Here - Statistical Analysis of Neuronal Data

Transcription

SAND7 POSTER PRESENTATIONS
There are over 60 posters, and we have decided to split them into two sessions, one before
dinner and one after dinner. The poster abstracts are numbered below. The odd numbers
will be presented before dinner; the even numbers will be presented after dinner.
(1) Abbasi-Asl, Reza Explaining V4 Neuron’s’ Pattern Selectivity via Convolutional
Neural Network
(2) Adams, Terrence Development of a Big Data Framework for Connectomic Research
(3) Agarwal, Rahul Nonparametric Estimation of Band-limited Probability Density
Functions: Application to Rat Entorhinal Cortical Neuron
(4) Best, Matthew Using spatial patterns of primary motor corical activity to predict
behavioral state
(5) Brigham, Marco Non-stationary filtered shot noise processes and applications to
neuronal membranes
(6) Chase, Steve Recasting brain-machine interface design from a physical control system perspective
(7) Climer, Jason Examining rhythmicity in extracellular recordings
(8) Coffman, Brian Event-related potentials demonstrate deficits in auditory gestalt
formation in schizophrenia
(9) Constantino, Francisco Neural rhythm synchronizes with imagined acoustic rhythm
(10) David, Stephen More isn’t always better: The essential complexity of auditory
receptive fields
(11) Deng, Xinyi Clusterless decoding of postion from multiunity activity using a marked
point process filter
(12) Dimitrov, Alex Characterizing local invariances in the ascending ferret auditory
system
(13) DiTullo, Ron Hypothesis testing of grid cell parameters using a maximum likelihood
framework
(14) Dyer, Eva Quantifying mesoscale neuroanatomy with X-ray micotomography
(15) Effenberger, Felix Discovery of salient low-dimensional dynamical structure in neuronal population activity using Hopfield networks
(16) Ezennaya-Gomez, Salatiel Detecting statistically significant synchronous spiking
activity
(17) Gao, Yu-Rong Using the Thresholded in Radon Space (TiRS) algoirthm to reveal
mechanical restriction of intracortical vessel dilation during voluntary locomotion
(18) Gerhard, Felipe Generative models to discover structure in neural recordings of
human focal epilepsy
1
2
(19) Ghanbari, Abed Estimating short-term synaptic plasticity from paired spike recordings
(20) Glaser, Joshua Using generalized linear models to understand neural correlates of
saccade remapping and planning in natural scenes
(21) Green, Patrick Integrating source localization and spike sorting
(22) Gunnarsdottir, Kristin A look at the strength of micro and macro EEG analysis
for distinguishing insomnia whithin an HIV cohort
(23) Haigh, Sarah MMN to complex pattern deviants in schizophrenia
(24) Horkunenko, A.B. Mathematical modeling of EEG for their automated analysis
and forecasts
(25) Hoseini, Mahmood Characterization an dproposed mechanisms of intermittent oscillations in cerebral cortex
(26) Huo, Bing-Xing Linear models of the hemodynamic response and neurovascular
coupling in the behaving animal
(27) Kadakia, Nirag Parameter and State Estimation in HVC RA-Projecting neuron
(28) Kamal, Vineet Prediction of outcomes after severe and moderate head infury using simple clinical and laboratory variables by classificaiton and regression tree
technique
(29) Karimipanah, Yahya Scale-free cortical resting state activity in vivo at single-cell
resolution
(30) Kasi, Patrick Decoding of tactile afferents responsible for sensorimotor control
(31) Kerr, Matthew Event-related potentials in human attentional networks during
movement perturbations
(32) Koyama, Shinsuke On the spike train variability characterized by variance-to-mean
power relationship
(33) Leong, Josiah White-matter connecting anterior insula to nucleus accumbens is
associated with functional brain activity and risk-taking behavior
(34) Li, Yuanning Decoding the temporal dynamics of left mid-fusiform gyrus activity
during word reading
(35) Liang, Hualou Copula models of multivariate point process for the analysis of ensemble neural spiking activity
(36) Lopour, Beth Long-range functional connectivity in the epileptic human brain using
the spike-triggered impulse response
(37) Madahian, Behrouz A statistical approach for seizure risk forecasting
(38) Mahan, Margaret Utilizing time-varying graphps for discovering dynamic functional
connectivity
(39) Matano, Francesa Decoding velocity with kinematic models and direct regression
(40) Matzner, Ayala Quantifying spike train oscillations: biases, distortions
(41) Nielsen, Karen Regression spline mixed models for anlayzing EEG data and eventrelated potentials
(42) Onaga, Tomokatsu Spontaneous fluctuations in networks of spiking neurons
(43) Park, Yun Early detection of human epileptic seizures using MUA and LFPs from
intracortical microelectrode arrays
3
(44) Ramezan, Reza A flexible model with multivariate extensions for neural spike trains
(45) Sacre, Pierre On a reduced model of spinal cord simulation for chronic pain: selective relay of sensory neural activities in myelinated nerve fibers
(46) Singh, Arun Restoration of normal striatal dopamine responses with NMDA/AMPA
receptor blockade in parkinsonian monkeys
(47) Smith, Ryan Task-specific neuronal ensembles improve coding of grasp
(48) Stylios, Chrysostomos Sleep apnea detection using a reduced set of measurements
and symbolic time series analysis
(49) Subramanian, Sandya A novel method for seizure localization in medically refractory epilepsy patients
(50) Venkatesh, Praveen Some thought experiments on the applicability of Granger
causality and directed information in statsitically inferring the direction of information flows
(51) Walsten, Doran Orbitofrontal cortex and hippocampus role in bias under uncertainty
(52) Ma, Zhengyu Coordinated neocortical activity at cellular resolution during visual
processing
(53) White, Matthew Mixed-effects spline models for modeling corical rhythm dynamics
in the developing human brain
(54) Whitmire, Clarissa Information coding through adaptive control of synchronized
thalamic bursting
(55) Yaffe, Robert Reinstatement of distributed spatiotemporal patternsof oscillatory
power during associative memory recall
(56) Yaghouby, Farid A Probabilistic Model to Resolve Uncertainty in Clinical Sleep
Scoring
(57) Yang, Ying Exploring Spatio-temporal Neural Correlates of Face Learning
(58) Zheng, Charles Stimulus identification from fMRI scans: A statistical perspective
(59) Zhou, Pengcheng Establishing a Statistical Link Between Network Oscillations and
Neural Synchrony
(60) Zhang, Qiong Characterization of brain consistency via a data-driven brain parcellation
(61) Michalopoulos, P. Prefrontal neurons represent comparisons of motion directions
in the contralateral and the ipsilateral visual fields
(62) Stokes, Patrick Fundamental Problems in Granger Causality Analysis of Neuroscience Data
(63) Fiddyment, Grant Point process modeling of human seizures
4
Explaining V4 Neurons’ Pattern Selectivity via Convolutional Neural
Network
Abbasi-Asl, Reza
abba30.reza@gmail.com
In this poster, we present our recent model analysis for neurons in V4 area of visual cortex
using natural images as stimulus. Recorded activities of 55 neurons from area V4 of two
awake macaque monkey were used. We build a computational model based on a convolutional neural network trained on ImageNet dataset to predict the neuron responses and
we further examine and interpret their pattern selectivity. Convolutional neural networks
- as a successful tool to analyze big data problems - has been recently studied for a vast
variety of applications especially in machine learning. Here, it has been shown that they
are also successful to increase our understanding of visual cortex and especially V4 cells.
This is a joint work with Yuansi Chen, Adam Bloniarz, Jack Gallant and Bin Yu.
*******************************
Development of a Big Data Framework for Connectomic Research
Terrence Adams, U.S. Government
tma2131@columbia.edu
This poster outlines research and development of a new Hadoop-based architecture for distributed processing and analysis of electron microscopy of brains. We show development of
a new C++ library for implementation of 3D image analysis techniques, and deployment
in a distributed map/reduce framework. We demonstrate our new framework on a subset
of the Kasthuri11 dataset from the Open Connectome Project.
*******************************
Nonparametric Estimation of Band-limited Probability Density Functions:
Application to Rat Entorhinal Cortical Neuron
Rahul Agarwal
rahul.jhu@gmail.com
In this paper, a nonparametric maximum likelihood (ML) estimator for band-limited (BL)
probability density functions (pdfs) is proposed. The BLML estimator is consistent and
computationally efficient. To compute the BLML estimator, three approximate algorithms
are presented: a binary quadratic programming (BQP) algorithm for medium scale problems, a Trivial algorithm for large-scale problems that yields a consistent estimate if the
5
underlying pdf is strictly positive and BL, and a fast implementation of the Trivial algorithm that exploits the band-limited assumption and the Nyquist sampling theorem
(BLMLQuick). All three BLML estimators out-perform kernel density estimation (KDE)
algorithms (adaptive and higher order KDEs) with respect to the mean integrated squared
error for data generated from both BL and infinite-band pdfs. Further, the BLMLQuick
estimate is remarkably faster than the KD algorithms. Finally, the BLML method is applied to estimate the conditional intensity function of a neuronal spike train (point process)
recorded from a rats entorhinal cortex grid cell, for which it outperforms state-of-the-art
estimators used in neuroscience.
***********************************
Using spatial patterns of primary motor corical activity to predict behavioral
state
Matthew Best
mattbest@uchicago.edu
Recent work has shown that primary motor cortical (MI) activity traverses through a lowdimensional neural state space across time. These neural trajectories have been fruitfully
used to predict motor output, both in the form of movement kinematics and muscle activity.
And yet, these models have not incorporated information about the spatial interrelationships between recording sites despite the fact that MI is a highly spatially distributed
cortical area with heterogeneous response properties. We hypothesized that consideration
of spatial information in simultaneously recorded neural activity in MI will lead to better
predictions about the behavioral state of an animal. To this end, we recorded local field
potential activity from a 96-channel Utah array implanted in the MI of a rhesus macaque
while it performed an instructed-delay center-out reaching task. During the instruction
epoch, corresponding to motor preparation, the amplitude of beta band activity (18 Hz) is
high whereas during active movement, corresponding to motor execution, beta amplitude
is low. The transition between high and low amplitude beta, henceforth referred to as beta
attenuation, may be seen as a cortical correlate of the transition between motor preparation
and execution. Here, we show that beta attenuation does not happen simultaneously across
the entirety of motor cortex, but rather propagates across the MI surface as a linear wave.
We used the simultaneous beta amplitudes from each of our electrodes as input features to
a multinomial logistic regression model that predicted task epoch (i.e. instruction, reaction
time, or active movement). We found that a model that includes explicit information about
the spatial interrelationships of the electrodes outperforms spatially shuffled data.
Joint work with Kazutaka Takahashi and Nicholas G. Hasopoulos
***********************************
Non-stationary filtered shot noise processes and applications to neuronal
membranes
6
Marco Brigham
brigham@unic.cnrs-gif.fr
Intracellular recordings provide direct access to statistical properties of membrane potential (Vm) fluctuations. The recordings at the soma reflect biological characteristics of the
neuron, such as the number of synapses, synaptic time constant and synaptic strength for
excitatory and inhibitory synapses. The neuron samples the dynamics of afferent neural
populations through its synaptic input, which is likewise reflected in the statistics of Vm.
The raw data from intracellular recordings can be processed to extract its statistical characteristics. Such compact representation of the data can be used to infer biological properties
of the neuron and dynamics of afferent populations through a statistical inference model.
A key requirement is a robust statistical model of Vm fluctuations that yields Vm statistics given biological and synaptic input characteristics. Exact analytical descriptions and
several approximations have been developed in previous analytical work [1] for the joint
cumulants of Vm, in a linear passive model under conductance-based, non-stationary shot
noise input. Gaussian and higher order approximations have also been obtained for the
non-stationary distribution of Vm using an Edgeworth expansion. In the present work,
a statistical inference model is developed to leverage the increased accuracy in statistical
description of Vm and the availability of higher order statistics, such as the skewness and
the autocovariance.”
Joint work with Alain Destexhe.
***********************************
Recasting brain-machine interface design from a physical control system
perspective
Steven Chase
schase@cmu.edu
With the goal of improving the quality of life for people suffering from various motor control
disorders, brain-machine interfaces provide direct neural control of prosthetic devices by
translating neural signals into control signals. These systems act by reading motor intent
signals directly from the brain and using them to control, for example, the movement of a
cursor on a computer screen. Over the past two decades, much attention has been devoted
to the decoding problem: how should recorded neural activity be translated into the movement of the cursor? Most approaches have focused on this problem from an estimation
standpoint, i.e., decoders are designed to return the best estimate of motor intent possible,
under various sets of assumptions about how the recorded neural signals represent motor
intent. Here we recast the decoder design problem from a physical control system perspective, and investigate how various classes of decoders lead to different types of physical
7
systems for the subject to control. This framework leads to new interpretations of why
certain types of decoders have been shown to perform better than others. These results
have implications for understanding how motor neurons are recruited to perform various
tasks, and may lend insight into the brain’s ability to conceptualize artificial systems.
***********************************
Examining rhythmicity in extracellular recordings
Jason R. Climer
jrclimer@bu.edu
Many studies have attempted to examine the rhythmic modulation of the firing of individual neurons from extracellular recordings. In the rodent hippocampus, neurons are known
to have a strong relationship to theta rhythm (6-12 Hz) oscillations in the local field potential and to be intrinsically rhythmic in this frequency range. In contrast, recent recordings
of single units in the bat hippocampal formation have not yielded significant rhythmicity.
Theta rhythmicity is most often measured by spectral properties of the spike time autocorrelogram; however, this method is known to be biased by properties such as the firing rate
of the neuron. As such, an in depth study of the limits of existing techniques is warranted.
Here, we have examined properties which may affect the ability to observe rhythmicity and
bias traditional measures using large batteries of simulated data. Traditional methods are
biased by a number of features, including firing rate and dwell time in a cell s receptive
field. To combat this, we have used a maximum likelihood estimation approach as a less
biased and more sensitive way to examine rhythmicity. In this approach, each lag within
the autocorrelogram is treated as an observation. This allows statistical testing of changes
in individual rhythmicity features (e. g. frequency or amount of rhythmicity) in a single
cell across multiple manipulations. Additionally, because each spike is not binned into the
autocorrelogram, we can quantify the relationship between rhythmicity features and other
behavioral parameters such as running speed. This approach offers a marked improvement
over existing methods and can greatly aid in our ability to examine rhythmic properties of
extracellularly recorded neurons.
***********************************
Event-related potentials demonstrate deficits in auditory gestalt formation in
schizophrenia
Brian A. Coffman
coffmanb@upmc.edu
Grouping of auditory percepts is necessary for interpretation of patterns. Schizophrenia
patients have blunted responses to deviance from an established norm, such as reduced
8
mismatch negativity (MMN). Here we compared auditory event-related potential (ERP)
responses to complex patterns between schizophrenia patients (SZ; N=25) and matched
healthy controls (HC; N=23). ERPs were measured in an auditory pattern in which the
first 6 tones increased in pitch in 500 Hz steps, from 1.5 4 kHz, and the last 6 tones
decreased in pitch (4 1.5 kHz). In 8% of trials, the last 6 tones repeated the increasing
pitch pattern of the first 6 tones. Here we focused the analysis only upon the frequent tone
pattern (616 trials; 50 ms duration; SOA = 330 ms; ITI = 800 ms). Stimuli were presented
while participants watched a silent video. We observed a large sustained negativity (SN)
throughout the entire duration of each group that returned to baseline following completion
of the trial. Relationship between SN and ordinal stimulus position was compared between
SZ and HC.SN was sensitive to ordinal stimulus position (p¡0.01), with largest responses to
first and final tones. HC had greater SN than SZ across the entire trial, though differences
were greatest for first and final tones (p¡0.001). These results suggest stronger set formation in HC than SZ. Deficits in auditory pattern processing may be relevant to clinical
issues in SZ, such as conceptual disorganization. Future studies will examine relationships
between SN and clinical measures.
Funding Source: NIH MH094328 (PI: Dean Salisbury, PhD).
Joint work with Sarah M. Haigh, Tim K. Murphy, Kayla Ward, Christian Andraeggi, Dean
F. Salisbury
***********************************
Neural rhythm synchronizes with imagined acoustic rhythm
Francisco Cervantes Constantino
fcc@umd.edu
Perceptual filling-in is one mechanism to handle missing sensory information, possibly operating by interpolation from context cues. While driven by sensory data, filled-in epochs
are a direct outcome of endogenous neural processes. For example, listening to an acoustic rhythm locks in steady-state responses from the auditory system (aSSR), which are in
phase with respect to the input rhythm. aSSR oscillations are driven by real sensory input,
resulting from a combination of exogenous and endogenous neural processes. We created
conditions to observe steady neural oscillations driven by contextual but not real sensory
input, thus entirely reflecting endogenous neural processes.
Brief noise masker probes were pseudo-randomly added to a long, rhythmic, acoustic pulse
train (5 Hz rate). In half of the masker probes, the ongoing rhythmic pulse train was also
omitted for the duration of the masker probe. 35 listeners were asked to report, shortly
after each masker probe, whether it had been perceived-rhythmic (pR), or not (pN). To
9
make detection moderately difficult, the masker noise level was selected per subject.
Analysis of magnetoencephalography (MEG) neural responses at the 5 Hz rhythm rate
shows that incorrect pR trials showed higher evoked rhythmic power and higher trial-totrial rhythmic phase coherence, than did correct pN trials. This contrast alone accounted
for 30% of the variance in detection sensitivity.
In modulation rates relevant to human speech communication, as in the rate tested here, we
propose that the presence of neural dynamics synchronized to an actual rhythm (or sound
modulation) directs the subjective experience of a sound as rhythmic (or modulated), even
when such synchronized dynamics is not supported by sensory data - in analogy to some
illusions or hallucinations. This strategy underlies an internal model generated to extract
meaning from complex sound mixtures, as in the problem of active listening to multiple
speakers. It also raises the question of contextual interpolation as a common-principled
strategy found in other sensory modes.
Joint work with Jonathan Z. Sinmon
***********************************
More Isn’t Always Better: The Essential Complexity of Auditory Receptive
Fields
Stephen David
davids@ohsu.edu
Understanding how the brain solves sensory problems can provide useful new insight into
the development of automated systems such as speech recognizers and image classifiers.
Recent developments in nonlinear regression and machine learning have produced powerful
algorithms for characterizing the input-output relationship of complex systems. However,
the complexity of sensory neural systems, combined with practical limitations on available
data, make it difficult to apply arbitrarily complex analyses to neural data. In this study we
pushed analysis in the opposite direction, toward simpler models. We asked how simple a
model can be developed to capture the essential sensory properties of neurons in auditory
cortex. We found that a substantially simpler formulations of the widely-used spectrotemporal receptive field is able to perform as well as the best current models. Moreover,
these simpler formulations define new basis sets that can be incorporated into state-of-theart machine learning algorithms for a more exhaustive exploration of sensory processing.
***********************************
Clusterless decoding of position from multiunit activity using a marked point
process filter
10
Xinyi Deng
xinyi@math.bu.edu
Millisecond-timescale patterns of neural activity are the substrate for the computations
that underlie complex cognitive processes. Developing a causal understanding of the relationship between these patterns and the processes they support requires tools that allow
us to manipulate the patterns selectively. In the hippocampus, for example, sequences of
place cells are often replayed during sharp-wave ripple events that last 100-200 ms long. If
we are to understand how specific sequences drive information processing in downstream
regions, we need the tools to identify these sequences as they occur and manipulate targeted circuits based on sequence identity.
Previously, we have used point process theory to develop efficient decoding algorithms
based on spike train observations. However these algorithms assume the spike signals have
been accurately sorted into single units before the algorithms are applied. As the unsupervised spike sorting problem remains unsolved, we took an alternative approach that takes
advantages of recent insights about clusterless decoding (Kloosterman et al., 2014). Here
we present a new point process decoding algorithm that does not require multiunit signals
to be sorted. We use the theory of marked point processes to construct a function that
characterizes the relationship between a desired variable (in this case the animals location
in space) and features of the spike waveforms. Using Bayes’ rule, we compute the posterior
distribution of a signal to decode the spatial locations represented in hippocampal multiunit activity.
We illustrate our approach with a simulation study along with experimental data recorded
in the hippocampus of a rat performing a spatial memory task. Our decoding framework
is used to reconstruct the animal’s position from unsorted multiunit spiking activity. We
then compare the quality of our decoding framework to that of a traditional spike-sorting
and decoding framework.
Our analyses show that the proposed decoding algorithm performs as well as or better than
algorithms based on sorted single-unit activity. These results provide a path toward accurate real-time decoding of spiking patterns that could be used to carry out content-specific
manipulations of population activity in hippocampus or elsewhere in the brain.
Joint work with Daniel F. Liu, Kenneth Kay, Loren M. Frank, Uri T. Eden
*******************************
Characterizing local invariances in the ascending ferret auditory system
Alex Dimitrov
11
alex.dimitrov@wsu.edu
The sense of hearing requires a balance between competing processes of perceiving and
ignoring. Behavioral meaning depends on the combined values of some sound features but
remains invariant to others. The invariance of perception to physical transformations of
sound can be attributed in some cases to local, hard-wired circuits in peripheral brain
areas. However, at a higher level this process is dynamic and continuously adapting to new
contexts throughout life. Thus the rules defining invariant features can change.
In this project, we test the idea that high-level, coherent auditory processing is achieved
through hierarchical bottom-up combinations of neural elements that are only locally invariant. Local probabilistic invariances, defined by the distribution of transformations that
can be applied to a sensory stimulus without affecting the corresponding neural response,
are largely unstudied in auditory cortex. We assess these invariances at two stages of the
auditory hierarchy using single neuron recordings from the primary auditory cortex (A1)
and the secondary auditory cortex (PEG) of awake, passively listening ferrets.
Joint work with Jean Lienard and Stephen David.
***********************************
Hypothesis Testing of Grid Cell Parameters Using a Maximum Likelihood
Framework
Ron W. DiTullio
ron.w.ditullio@gmail.com
Since their discovery in 2004, grid cells have been a focal point of research for those investigating the neural basis of spatial navigation and memory as well as a unique statistical
challenge for those interested in the analysis of neuronal signals. Both the interest and
challenge of grid cells result from the unique, geometric pattern of the firing fields from
which these cells derive their name. Specifically: when grid cells are recorded from as an
animal explores an environment, their firing fields appear to fall on the vertices of equilateral triangles that tessellate in a grid like fashion throughout the explored space. Studies
investigating the properties of grid cells have focused on how the geometric properties of
the firing fields, such as the spacing between fields or the orientation of the fields, change
in response to various manipulations and have traditionally employed analyses based on
using a 2-d spatial autocorrelation, or autocorrelogram, of the data. Although intuitive,
autocorrelogram based analyses contain several biases and limitations that either entirely
precluded or at least make it very challenging to test certain hypotheses about grid cell firing properties. Here we develop an alternative method of analysis that utilizes parametric
modeling and particle swarm optimization in a maximum likelihood estimation framework
to allow for more accurate and more powerful testing. We demonstrate the accuracy of
12
this algorithm via the analysis of a battery of simulated cells and compare this accuracy
to current, autocorrelogram based techniques. We demonstrate the power of this method
in a hypothesis testing framework and discuss the advantages of using this algorithm in
common experimental designs for investigating grid cells. Finally, as a proof of concept
we re-analyze a previously published set of data regarding the response of grid cells to
environmental novelty.
Joint work with Jason R. Climer, Michael E. Hasslemo and Uri T. Eden.
***********************************
Quantifying mesoscale neuroanatomy with X-ray microtomography
Eva Dyer
edyer@ric.org
Neuroanatomy is essential for studying a number of neurological diseases as well as providing an atlas necessary to study brain function. Although relatively unused in neuroscience
to date, synchrotron-based X-ray microtomography (XRM) offers a new way of imaging
large brain volumes in order to quantify neuroanatomy; however, new computational methods are required to extract and analyze the underlying neural structures (cells, vessels, and
processes) in XRM data. To this end, we developed a host of methods for segmenting
and analyzing the spatial statistics of large brain volumes using XRM. To segment image
volumes, we extract multi-scale features from the 3D volume that characterize the shape
(for instance, whether the voxels can be well-approximated by a sphere or cylinder) and
intensity of voxels in a small cube of the data. Using these features, we trained a gradient
boosting classifier to distinguish between cell bodies, vessels, processes, and background
voxels. In addi tion, we used a non-parametric nearest-neighbor-based density estimation
technique to estimate a smooth continuous density function which describes the 3D distribution of cells in a volume of brain tissue. We applied this suite of tools to study and
compare the spatial distribution of cells in millimeter scale volumes from three different
animals: mouse, monkey, and human. Our results demonstrate that XRM provides a new
method for large-scale brain imaging that is complementary to optical and electron microscopy.
Joint work with Hugo L. Fernandes, Naryanan Kasthuri, Xianghui Xiao, Chris Jacobsen,
Konrad P. Kording.
***********************************
Discovery of salient low-dimensional dynamical structure in neuronal
population activity using Hopfield networks
13
Felix Effenberger
Felix.Effenberger@mis.mpg.de
We present here a novel method for finding and extracting salient low-dimensional representations of the dynamics of populations of spiking neurons. This is a classical problem
in data analysis of parallel spike trains, and quite a number of approaches to detect and
classify recurring spatiotemporal patterns (STP) of neural population activity were proposed [PWSN08, PMnBB+13, LdSRT13, GR10].
Yet, most published methods so far assume a noiseless scenario (apart from allowed jitter
in spike times for some methods) and either focus on (partial) synchrony detection and /
or seek to classify exactly recurring STP in neuronal activity. Yet, given the usually high
variability of population responses to stimuli, the re-occurrence of such exactly repeating
STP becomes more and more unlikely with increasing population size. Assuming that
despite this variability, network activity is not random per se (under the well-supported
hypothesis that the population has to code information about stimuli in some form of
STP), a much more plausible situation is that some underlying STP appears in several
corrupted variants differing in a few shifted, missing or excess spikes (characterized by a
low Hamming distance to some true, underlying STP).
The method proposed here uses Hopfield networks fitted to windowed, binned spiking activity of a population of cells using minimum probability flow (MPF) [SDBD11]. The method
is robust to the aforementioned variability in the signal and able to extract underlying
recurring patterns in an unsupervised way, even for seldom occurring STP and large population sizes. Modeling furthermore the sequence of occurring STP as a Markov process,
we are able to extract low-dimensional representations of neural population activity and
prominently occurring sequences of STP in the data. We demonstrate the approach on a
data set obtained in the rat barrel cortex [MCT+11] and show that it is able to extract a
remarkably low-dimensional yet accurate representation of the mean population response
to whisker stimulation that we computed using knowledge of the stimulus protocol. In
contrast, our method is able to extract this information without any knowledge of the
stimulus protocol. We thus prop ose the method as a novel tool in mining parallel spike
trains for possibly low-dimensional underlying network dynamics. An open source software
allowing for the wider application of the method is to be released soon.
References
[GR10] S. Grun and S. Rotter. Analysis of parallel spike trains. Springer, 2010.
[LdSRT13] V. Lopes-dos Santos, S. Ribeiro, and A. B. L. Tort. Detecting cell assemblies
in large neuronal populations. Journal of Neuroscience Methods, 220(2):14966, 2013.
14
[MCT+11] M. Minlebaev, M. Colonnese, T. Tsintsadze, A. Sirota, and R. Khazipov. Early
gamma oscillations synchronize developing thalamus and cortex. Science, 334(6053):226229,
2011.
[PMnBB+13] D. Picado-Muino, C. Borgelt, D. Berger, G. Gerstein, and S. Grun. Finding
neural assemblies with frequent item set mining. Frontiers in Neuroinformatics, 7(May):9,
2013.
[PWSN08] G. Pipa, D. W. Wheeler, W. Singer, and D. Nikolic. NeuroXidence: reliable and
efficient analysis of an excess or deficiency of joint-spike events. Journal of Computational
Neuroscience, 25(1):6488, 2008.
[SDBD11] J. Sohl-Dickstein, P.B. Battaglino, and M.R. DeWeese. Physical Review Letters,
107(22):220601, 2011.
***********************************
Detecting Statistically Significant Synchronous Spiking Activity
Ezennaya-Gomez, Salatiel
salatiel.ezennaya@softcomputing.es
We consider the task of finding significant frequent synchronous events in parallel point
processes, and particularly in neural spike trains, where this task is connected to testing the
temporal coincidence coding hypothesis. Our approach transfers methods from frequent
item set mining (FIM) to a continuous time domain. It counts the number of synchronous
spiking events with a maximum independent set (MIS) approach, which ensures that no
spike contributes to more than one counted synchronous event. This leads to a natural and
efficiently computable support measure, which effectively handles the problem of temporal
imprecision (that is, it allows for imprecise spike synchrony) via a user-specified window
width.
We developed an efficient implementation of our algorithm, which we call CoCoNAD (for
Continuous-time Closed Neuron Assembly Detection). This basic algorithm for finding frequent patterns has been enhanced by so-called Pattern Spectrum Filtering (PSF), which
generates and analyzes surrogate data sets to identify statistically significant patterns, and
Pattern Set Reduction (PSR), which eliminates spurious induced patterns. The effectiveness of our approach has been demonstrated with a large number of artificially generated
data sets with injected synchronous activity, which can be identified almost perfectly provided it cannot be explained as a chance event resulting from background activity.
Recent extensions of our approach concern a graded notion of synchrony, which takes both
the number of synchronous events as well as the precision of synchrony into account, and a
method to handle selective participation, that is, means to detect synchronous activity of
15
a group of neurons, even if each individual synchronous spiking event contains only spikes
from a (randomly chosen) subset of the neurons.
Joint work with David Picado-Muiato, Christian Borgelt
***********************************
Using the Thresholded in Radon Space (TiRS) algorithm to reveal mechanical
restriction of intracortical vessel dilation during voluntary locomotion
Yu-Rong Gao
yzg102@psu.edu
To decipher the vascular basis of hemodynamic signals and neurovascular coupling, it is
essential to understand the spatial hemodynamics in the brain during natural behaviors. In
awake, head-fixed animals, locomotion drives large, rapid dilation in pial surface arteries,
and a smaller, slower surface venous distension. Penetrating arterioles enter into the brain
tissue and feed the surrounding neurons, serving as a bridge between the surface vasculature and the capillary bed. However, it is not known whether the intracortical arterioles
and venules behave similarly as those on the brain surface, and whether this hemodynamic
response is spatially localized during natural behavior. Because penetrating arterioles
and ascending venules are oriented to the imaging plane when visualized with two-photon
microscopy, and these vessels can change their shape during dilation or constriction, an
accurate method of quantifying the diameter changes of intracortical vessels in the brain
i s needed. We developed a novel algorithm Thresholding in Radon Space (TiRS). This
method transforms the vessel into Radon space, thresholds the transformed data and then
transforms it back into image space. The TiRS method makes use of the global structure
of the image, allowing us to determine the intracortical vessel cross-sectional area with
superior accuracy, and greater robustness to noise and vessel shape changes than previous
thresholding and full-width at half maximum (FWHM) methods.
We applied the TiRS algorithm to two-photon imaging data of vascular dynamics in the
somatosensory cortex of awake, head-fixed mice during voluntary locomotion. During
voluntary locomotion, intracortical arteriole dilation was correlated with nearby neural activity. Surprisingly, we found that surface arterioles and venules dilated significantly more
than the intracortical arterioles and venules. The smaller dilations of the intracortical arterioles were not due to saturation of dilation, as intracortical and surface arterioles dilated
to the same extent when the mouse was under isoflurane, which is a profound vasodilator.
Histology showed that, unlike surface vessels, intracortical vessels were tightly enclosed by
brain tissue. A mathematical model, using published values of vascular and brain tissue
stiffness, demonstrated that mechanical restriction by brain tissue could account for the
reduced amplitude of intracortical vessel dilation. Our results support the hypothesi s that
the mechanical properties of the brain may play an important role in sculpting the laminar
16
differences of hemodynamic responses.
References:
Gao Y-R, Drew PJ. Determination of vessel cross-sectional area by thresholding in Radon
space. J Cereb Blood Flow Metab 2014; 34: 11801187.
Gao Y-R, Greene SE, Drew PJ. Mechanical restriction of intracortical vessel dilation by
brain tissue sculpts the hemodynamic response, submitted.
Joint work with Stephanie E. Greene, Patrick J. Drew.
***********************************
Generative models to discover structure in neural recordings of human focal
epilepsy
Felipe Gerhard
felipe_gerhard@brown.edu
Electrophysiological recordings in humans with focal epilepsy have shown the emergence
of highly heterogeneous spiking patterns on the level of single neurons. Yet, across consecutive seizures these patterns seem to be consistently reactivated. Here, we present the
application of two complementary statistical models of multivariate point processes that
capture the dynamics of observed, recurring network patterns.
One model is based on a low-dimensional hidden linear dynamical system (LDS) that
drives the firing rates of individual single-neurons in the ensemble (Poisson-LDS). The
other model couples neurons’ firing rates to the ensemble activity through a dense network
of effective connections and a newly introduced mean-field coupling (ensemble-history Generalized Linear Model, GLM). We find that both models could predict single-neuron firing
equally well: Cross-validated Area-under-Curve (AUC) scores ranged from 0.7 to 0.9 for
two types of seizures with qualitatively different dynamics. Furthermore, predictions of
both models were partially correlated, indicating that models captured similar dynamical
features of the spiking patterns. We hypothesize that a combination of both models could
further improve prediction.
Both models provide a generative mechanism for the underlying seizure dynamics that
could not otherwise be derived from simpler, descriptive statistics. The ability to find
compact statistical descriptions of high-dimensional neural recordings is a major step towards better algorithms to detect and control seizures in humans.
17
Joint work with Sydney S. Cash, Wilson Truccolo
***********************************
Estimating short-term synaptic plasticity from paired spike recordings
Abed Ghanbari
abed.ghanbari@uconn.edu
Synaptic connections between neurons evolve over time, and these changes affect transmission and processing of information in neuronal circuits. Synaptic plasticity is traditionally
studied using intracellular recording techniques where the synaptic weight can be directly
estimated from postsynaptic potentials or currents. Since large-scale intracellular recordings are not possible in vivo, statistical methods that can estimate synaptic plasticity from
spike trains would be a valuable tool.
Recently, model-based methods were developed for estimating long-term synaptic changes
from spike trains. Estimating the type of timescale of short-term plasticity (STP), which
operates on timescales similar to the inter-spike intervals, represents an additional challenge. Here, we use a modified generalized-linear-model (GLM) to describe postsynaptic
dynamics and a time-varying coupling between a presynaptic neuron and a postsynaptic
neuron. We constrain the coupling term to vary according to the extended Tsodyks and
Markram (eTM) model a set of nonlinear differential equations that can accurately describe experimentally observed synaptic dynamics produced by facilitation and depression
using four parameters: the baseline release probability, the magnitude of facilitation, and
time constants for depression and facilitation.
We estimate model parameters using maximum likelihood with a coordinate ascent that
alternates between optimizing the GLM parameters and the eTM parameters. In order
to measure the accuracy of the plasticity parameter estimation in a realistic setting we
generated a postsynaptic current produced by spiking of 1024 model presynaptic neurons,
segregated in around 170 groups of different synaptic weights and 6 different sets of plasticity parameters. We recorded spike responses to injection of this artificial postsynaptic
current in layer 2/3 pyramidal neurons in slices from rat visual cortex in vitro. In both
this controlled experimental setting, as well as in simulation, we find that a model-based
approach (i) can recover short-term plasticity parameters from pairs of spike trains and
(ii) makes more accurate spike predictions than a model without plasticity.
Joint work with Vladimir Ilin, Maxim Volgushev, Ian H. Stevenson
***********************************
Using Generalized Linear Models to Understand Neural Correlates of
Saccade Remapping and Planning in Natural Scenes
18
Joshua Glaser
j-glaser@u.northwestern.edu
How is visual information transferred across eye movements (saccades) so that we can plan
future saccades and maintain a stable perception of the world? To answer this question, we
analyze recordings from the Frontal Eye Field, a region involved in saccade planning, while
monkeys searched natural scenes for an embedded target. We first used classical techniques
(comparing firing rates across conditions) to determine several factors that contributed to
the neurons firing rate, e.g. whether the search target was in the neurons receptive field
before and after the saccade. We then used generalized linear models to 1) fit more accurate receptive fields from the natural scenes data, and 2) disambiguate between the factors
that seem to affect neural activity. Lastly, we used neural activity to predict behavior,
specifically how long it would take the monkey to find the target. Our findings elucidate
a neural method for transfer of visual information across saccades, and more generally,
demonstrate techniques for analyzing neural data in complex naturalistic environments.
Joint work with Daniel K. Wood, Pavan Ramkumar, Patrick N. Lawlor, Mark A. Segraves,
Konrad P. Kording.
***********************************
Integrating source localization and spike sorting
Patrick Greene
pgreene@email.arizona.edu
In electrophysiology experiments, extracellular signals are recorded from multiple neurons
near the probe tip. As technology advances and probes become more sensitive, spike data
from increasingly large numbers of neurons can be simultaneously recorded. Interpreting
this data requires spike sorting grouping spikes according to the likely identity of the neuron that produced them. Current spike sorting methods often have difficulties with similar
spike waveforms, especially if they are low amplitude and have a significant component
consisting of other, more distant neurons that fired at approximately the same time (see
e.g. [Pedreira 2012 ]). Such ambiguous spikes are frequently discarded, possibly wasting
a significant amount of useful data. As the number of detectable neurons increases, this
problem becomes more severe both because spikes are more likely to overlap in time, and
because the likelihood of two neurons having similarly-shaped waveforms increases.
To increase the accuracy and yield of spike sorting algorithms, we study a novel method
that combines source localization, as introduced by Mechler and Victor [Mechler 2012],
with spike sorting. We investigate the extent to which positional information improves
19
spike sorting accuracy, and systematically quantify the uncertainty associated with putative sorts. The spatial relationships between neurons that we obtain by simultaneously
localizing and sorting neural signals may be useful for understanding local connectivity
within a brain region.
Citations
Mechler F, Victor J. Dipole characterization of single neurons from their extracellular action potentials. J. Computational Neuroscience. 32, 73-100, 2012.
Pedreira C, Martinez J, Ison M, Quiroga R. How many neurons can we see with current
spike sorting algorithms? J. Neuroscience Methods. 211(1), 58-65, 2012.
Joint work with Jean-Marc Fellows, Kevin K. Lin.
***********************************
A look at the strength of micro and macro EEG analysis for distinguishing
insomnia within an HIV cohort
Kristin M. Gunnasdottir
kgunnar1@jhu.edu
HIV patients are often plagued by sleep disorders and suffer from sleep deprivation. However, there remains a wide gap in our understanding of the relationship between HIV status,
poor sleep, overall function and future outcomes; particularly in the case of HIV patients
otherwise well controlled on cART (combined anti-retroviral therapy). In this study, we
compared two groups: 16 non-HIV subjects (seronegative controls) and 12 seropositive HIV
patients with undetectable viral loads and well managed cd4 counts. The two groups were
age-, PSQI-, and BMI-matched. We looked at sleep behavioral (macro-sleep architectural)
features and sleep spectral (micro-sleep architectural) features obtained from human-scored
overnight EEG recordings in order to observe if the annotated (i.e. scored) EEG data can
be used to distinguish between controls and HIV subjects in a more quantitative manner. Specifically, the behavioral features were defined by sleep stages and included sleep
transitions , percentage of time spent in each sleep stage, and duration of time spent in
each sleep stage. The sleep spectral features were obtained from the power spectrum of
the EEG signals by computing the total power across all channels and all frequencies, as
well as the average power in each sleep stage and across different frequency bands. While
the behavioral features do not distinguish between the two groups, there is a significant
difference and a high classification accuracy for the scoring-independent spectral features.
This suggests that the behavioral features, that are subjective and prone to human factors,
have limitations and do not appear to be useful for identifying sleep complications in HIV
20
patients. Furthermore, there are currently no biomarkers that predict the early development of cognitive decline in HIV patients, which have been shown to have a great impact
on morbidity and mortality. We take a special interest in this spectral separation of the
groups because evidence suggests a relationship between subjective sleep complaints and
cognitive dysfunction in HIV patients stable on cART. Thus, a micro-sleep architectural
approach could serve as a biomarker to identify HIV patients vulnerable to cognitive decline, providing an avenue to explore the utility of early intervention.
Joint work with Yu Min Kang, Matthew S. D. Kerr, Sridevi V. Sarma, Joshua Ewen,
Richard Allen, Charlene Gamado, Rachel M.E. Salas.
***********************************
MMN to Complex Pattern Deviants in Schizophrenia
Sarah Haigh
haighsm@upmc.edu
The neural mechanisms that generate mismatch negativity (MMN) are debated, yet MMN
is being assessed as a possible biomarker for schizophrenia (SZ). In SZ, MMN is smaller to
stimulus deviants that differ in simple physical characteristics such as pitch or intensity.
This suggests that primary auditory cortex is affected in SZ, but it is unclear whether it
reflects deficits in stimulus adaptation, novelty detection, or both. MMN is also elicited by
complex-pattern deviants, a finding that cannot be due to non-adapted cells. We measured
MMN to complex-pattern deviants to assess novelty detection MMN in SZ and healthy controls (HC). Eight tones differing in 0.5 kHz steps were used in a standard zig-zag ascending
pitch pattern (1, 2, 1.5, 2.5, 2, 3, 2.5, 3.5 kHz tones), with two final tone deviants: 2.5
kHz (repeat), or 4 kHz (jump). Subjects watched a silent video, and were presented with
80% standard patterns, 10% repeat- and 10% jump-deviants. HC (N=23) produced a late
MMN-lik e negativity (400-500 ms after stimulus-onset) that was significantly larger than
patients with chronic SZ (N=23) to both the repeat (p=.038) and jump-deviant (p=.014).
The topography and source of the activity was consistent with a typical MMN response.
The MMN from a complex deviant cannot be argued to be due to adaptation because
there was no repeated single tone to drive adaptation, and the MMN was too late to be
contaminated by a larger N1 response to novelty. Patients with schizophrenia did not produce a late-MMN to the repeat- or the jump-deviant suggesting deficits in novelty detection.
***********************************
Mathematical modeling of EEG for their automated analysis and forecasts
A.B. Horkunenko
21
horkunenko@gmail.com
Modeling, analysis, forecast cyclical EEG data are important tasks whose solution makes
it possible to predict and make decisions regarding the activities of the human brain.
Among the researchers of mathematical modeling and analysis of EEG there are such scientists as Bostem, Cooper, Arezzo etc. [1, 2]. Development of computer automated analysis,
prediction and simulation EEG requires creating a mathematical model of the EEG.
This report grounds using cyclic random process as a mathematical model EEG [3] that
occurs in most practical cases nature of EEG and stochasticity and variability of rhythmic
structures. Using this mathematical model makes the possibility of spreading of developed
methods of statistical evaluation of stochastic characteristics of random cyclic processes
for EEG study [3].
References:
1. L. Patomaki, J. Kapio, and P. Karjalainen, Traking of nonstationary EEG with the
roots of ARMA models, IEEE Conf. EMBC-95, - 1995.
2. Wojciech Zaremba Modeling the variability of EEG/MEG data through statistical machine learning, cole Polytechnique, M.Sc. - 2012.
3. Lupenko S.A. Determined and casual cyclic function as a model for oscillatory phenomena and signals : definition and classification Electronic simulation. Institute of modeling
problems in power them. GE Pukhov . Volume 28 , ?4, 2006. - P. 29-45 ).
***********************************
Characterization and proposed mechanisms of intermittent oscillations in
cerebral cortex
M. Hoseini
sayedmahmood@go.wustl.edu
Rhythmic oscillations are ubiquitous in cerebral cortex and their potential functional roles
continue to excite the imagination of neuroscientists. These oscillations are (i) intermittent, are (ii) of variable durations and frequencies, and (iii) typically are accompanied by
sparse and irregular single neuron spiking. To our knowledge, no one spiking model has
succeeded in capturing these three characteristics of cortical oscillations. What combination of neuronal and network properties mediates the characteristic features of observed
cortical oscillations? To address this question, we recorded spontaneous and evoked neuronal oscillations in the visual cortex of turtle. This preparation was chosen because the
22
local field potential (LFP) oscillations generated in this cortex are sufficiently large to allow
single-trial analysis of characteristic features without the need for averaging. We determined the frequency profiles for LFP oscillations, as well as the variability in the amplitude
and the frequency of oscillations across trials, recording sites, and visual stimuli. Importantly, we designed a network of spiking model neurons with the objective to investigate
model parameter values such that the network reproduces the observed features of cortical
oscillations. The primary results of this study are that (a) visually-evoked activity often
exhibits very large power increases with peaks in multiple narrow frequency bands, (b)
from trial to trial, these peaks in relative power occur among different sets of frequencies within the 0.7 - 100 Hz range, and (c) for individual trials, spectral peaks are often
shared among groups of electrodes across the electrode array, but these electrode groups
may vary across trials and by frequency. The intermittent oscillations of variable duration and frequencies, accompanied by sparse and irregular spiking were reproduced with
a model network consisting of excitatory neurons, and fast and slow inhibitory neurons
. Model neurons with spike-rate adaptation were connected randomly to form a sparse
network. Our model results indicate that fast interneurons help to keep the balance between excitation and inhibition, while slow interneurons play a critical role in turning off
oscillations and causing intermittency. Adding dendritic non-linearity to the model allows
for intermittency over a broader range of network parameters and makes the system more
robust to noise. Our investigation demonstrates the possibility of generating intermittent
and variable gamma-band oscillations in a network with realistic parameters and irregular
and sparse single-neuron spiking.
***********************************
Linear models of the hemodynamic response and neurovascular coupling in
the behaving animal
Bing-Xing Huo
bih5103@psu.edu
Cerebral hemodynamic responses to sensory stimuli are widely used to infer neural activity.
However, whether the hemodynamic response accurately reflects neural activity, and if the
cerebral hemodynamic signals are affected by cardiovascular changes is unclear. To better
understand neurovascular coupling during normal behavior, we measured neural and vascular response in in the frontal and parietal cortices of head-fixed mice during voluntary
locomotion. We measured the cerebral blood volume (CBV) responses to voluntary locomotion using intrinsic optical signal (IOS) imaging, cerebral blood flow (CBF) using laser
Doppler flowmetry (LDF), and neural activity using stereotrodes. We found that locomotion drove CBV and CBF increases together with an increase in the local field potential
(LFP) and multi-unit activity (MUA) in the parietal cortex. The neural activity increase in
the frontal cortex was not accompanied by a significant hemodynamic response. This result
showed that neurovascular coupling is brain region specific. We then developed a simple
linear model, based on 2-photon microscope measurements of individual vessel dynamics,
23
to quantify the spatial extent of cortical CBV increases seen during voluntary locomotion.
This model allowed us to linearly decompose the cortical hemodynamic response to locomotion into a spatially localized arterial component, and a more diffuse venous component.
We then tested if the hemodynamic response within the parietal cortex was affected by
the cardiovascular changes that accompany locomotion. We occluded locomotion-induced
heart rate increases with glycopyrrolate, or reduced heart rate increase with atenolol. Using
this model of the hemodynamic response, we found that the arterial responses and CBF
were not detectably affected by cardiovascular perturbations, while the venous responses
were significantly attenuated by atenolol. Our results show that cortical hemodynamic signals can be decomposed int o arterial and venous components, with distinct spatial profiles
and sensitivities to cardiovascular perturbations.
Huo, B.-X., Gao, Y.-R., Drew, P.J., 2015. Quantitative separation of arterial and venous
cerebral blood volume increases during voluntary locomotion. Neuroimage 105, 369379.
Huo, B.-X., Smith, J.B., Drew, P.J., 2014. Neurovascular Coupling and Decoupling in the
Cortex during Voluntary Locomotion. J. Neurosci. 34, 1097581. doi:10.1523/JNEUROSCI.136914.2014
Joint work with Yu-Rong Gao, Jared Smith, Stephanie Greene, Patrick Drew
***********************************
Parameter and State Estimation in HVC RA-Projecting Neurons
Nirag Kadakia
nkadakia@physics.ucsd.edu
The brief, stereotypical songs produced by zebra finch songbirds have been studied and
extensively characterized in terms of auditory output and neural behavior. Neural firing
patterns and connectivity have been studied in various regions of the songbird brain known
as the HVC and RA, which are together responsible for learning, shaping, and producing
the birds vocal output. While several archetypal features of neural firing and the auditory
output have been measured, it is as yet unclear exactly how the peculiar firing patterns can
be explained by appropriate combinations of cellular and network properties. In particular,
neurons in the HVC that excite neurons in the RA (HVCRA neurons) have been found
to have extremely sparse, short bursts during the song. On the other hand, neurons in
the HVC that inhibit these HVCRA neurons (so-called HVC interneurons, or HVCI) burst
densely and broadly throughout the song, as do the RA neurons which receive excitation
from the HVC and in turn connect to the vocal box itself. Recent data has suggested that
the inhibitory effect of the interneurons may play a role in the shaping of individual syllables in the song, but the exact cellular mechanisms and network connectivity are largely
conjectural.
24
This work seeks to incorporate experimental data through a refined data assimilation technique to give insight into cellular properties. In this technique, a proposed model with
unknown parameters and dynamical state variables is combined with measured data (presumably sparse) to determine the parameters of the system. The data assimilation method
is defined as a path integral representation of transition probabilities, together defining the
model trajectory, conditioned on the measurements. We have evaluated this path integral
in several methods, using combinations of numerical procedures and variational approximations. It has been successful in many toy models, and here we extend the applications
to neural data. We show that it can determine a host of linear and nonlinear parameters
and unmeasured state variables in the neural model of the zebra finch HVC to excellent
accuracy.
***********************************
Prediction of outcomes after severe and moderate head injury using simple
clinical and laboratory variables by classification and regression tree technique
Vineet Kumar Kamal
vineetstats@gmail.com
Traumatic brain injury is the leading cause of disability and death all over the Globe.
Our aim is to develop and validate a prognostic model, which is simple and easy to use
for In-hospital mortality and unfavourable outcome at 6-months in patients with moderate and severe head injury involving rapidly and easily available variables in daily routine
practice. For this, a classification and regression tree (CART) technique was employed in
the analysis by using trauma database (n=1466 patients) of consecutive patients. A total
of 24 prognostic indicators were examined to predict In-hospital mortality and outcome
at 6 months after head injury. For In-hospital mortality, there were 7 terminal nodes and
the area under curve was 0.83 and 0.82 for learning and test data sample respectively.
The overall classification predictive accuracy was 82% for learning data sample and 79%
for test data sample, with a relative cost 0.37 for learning data sample. For 6-months
outcome, there were 4 terminal nodes and the area under curve was 0.82 and 0.79 for
learning and test data sample respectively. The overall classification predictive accuracy
was 79% for learning data sample and 76% for test data sample, with a relative cost 0.40
for learning data sample. Methodologically, CART is quite different from other commonly
used statistical methods with the primary benefit of illustrating the important prognostic
variables as related to outcome. This seems less expensive, less time consuming, and less
specialized measurements and may prove useful in developing new therapeutic strategies
and approaches.
Joint work with RM Pandey
***********************************
25
Scale-free Cortical Resting State Activity in vivo at Single-cell Resolution
Yahya Karimipanah
yahya@wustl.edu
Mounting evidence from fMRI, EEG, and LFP recordings of resting state activity in vivo
reveals a high level of coordination among the neuronal populations at the recording sites
and specifically indicates a lack of a characteristic scale in the spatiotemporal patterns
of activities. This scale-free nature of cortical activity suggests the attractive hypothesis that the cortex operates near a critical state between order (large-scale activity) and
disorder (small-scale activity), which, on theoretical grounds, has long been suggested to
be optimized for computation. The coarse spatial resolution (¿100 m) of the fMRI, EEG,
and LFP recording methods, raises the question whether the scale-free nature of cortical
activity extends to a small cortical volume consisting of some 40 neurons.
To address this question, we labeled layer 2/3 cells in the primary visual cortex of urethaneanaesthetized adult mouse by bolus injection of the calcium indicator dye Oregon Green
488 BAPTA-1 AM, used two-photon calcium imaging to monitor ensemble activity, and
inferred spikes as described previously (Kwan, Dan 2012). We thus obtained the inferred
spike trains of several minutes duration from up to 40 simultaneously recorded neurons in
primary visual cortex from 42 mice.
Recordings of ongoing cortical L2/3 activity at single-cell resolution revealed pronounced
coordinated activity among the population of some 40 closely-spaced neurons. First, temporal correlations for both single neuron and network activity were exposed using the
Detrended Fluctuation Analysis, which showed a linear trend for a long range of time
windows, indicating the existence of long-term memory. Second, the cross-correlation coefficients among the spike trains of pairs of neurons were generally small with a skewed
non-Gaussian distribution dominated by a long tail. Third, correlations in time and among
neurons were further revealed using the neuronal avalanche concept. The avalanche size
and duration distributions were best fit by a power law function (both truncated and
with exponential cutoff), compared to other commonly tested functions (e.g., exponential, lognormal, etc). Fourth, consistent with the properties of a dynamical critical state,
the avalanche sizes scaled with avalanche duration. Fifth, a critical model network with
synaptic depression qualitatively reproduced the four observed hall marks of coordinated
activity. Taken together, the data and model investigations support the hypothesis that
the mouse primary visual cortex operates near a critical state including at the cortical
microcircuit level.
***********************************
Decoding of Tactile Afferents Responsible for Sensorimotor Control
26
Patrick Kasi
pkasi@yahoo.com
Humans manipulate objects, in daily activities, with great precision. Experimental studies
have demonstrated that tactile signals encoded by mechanoreceptors are key to the precise
object manipulation in humans, however, little is known about the underlying mechanisms.
Current models range from complexthey account for skin tissue propertiesto simple regression fit. These models do not describe the dynamics of neural data well. We propose analyzing these data within the point process framework because it allows for characterization
of neural dynamics. The knowledge acquired may provide insight into some fundamental
sensory mechanisms that are responsible for coordinating force components during object
manipulation. We envisage that the knowledge may guide the design of sensory-controlled
biomedical devices and robotic manipulators.
***********************************
Event-Related Potentials in Human Attentional Networks During Movement
Perturbations
Matthew Kerr
matthew.sd.kerr@gmail.com
While both the neural substrates of attention and motor control have been extensively
studied in recent decades, minimal research has been done on the role of associative cortices including the orbitofrontal cortex, precuneus, and hippocampus during motor tasks
with unexpected perturbations. In a center-out reaching task with unexpected force perturbations performed in human subjects with Stereo-tactic EEG implants, evoked potential
responses were observed in the hippocampus, precuneus, and orbitofrontal cortex timelocked to the perturbation. Based on existing non-motor literature, these may correspond
to recognition of violated expectations, a shift in spatial attention, and the inhibition of
the prior movement plan respectively.
Joint work with Kevin Kahn, Hyun-Joo Park, Mathew Johnson, James Lee, Susan Thopson, Juan Bulacio, Jorge Gonzalez-Martinez, Sridevi V. Sarma, John T. Gale
***********************************
On the spike train variability characterized by variance-to-mean power
relationship
Shinsuke Koyama
27
skoyama@ism.ac.jp
We propose a statistical framework for modeling the non-Poisson variability of spike trains
observed in a wide range of brain regions. Central to our approach is the assumption that
the variance and the mean of ISIs are related by a power function characterized by two
parameters: the scale factor and exponent. This single assumption allows the variability
of spike trains to have an arbitrary scale and various dependencies on the firing rate in the
spike count statistics, as well as in the interval statistics, depending on the two parameters
of the power function.
On the basis of this statistical assumption, we show that the power function relationship
between the mean and variance of ISIs with various exponents emerges in a stochastic leaky
integrate-and-fire model under certain conditions. We also discuss based on this result that
the conventional assumption of proportional relationship between the spike count mean and
variance could lead to the wrong conclusion regarding the variability of neural responses.
Finally, we propose a statistical model for spike trains that exhibits the variance-to-mean
power relationship, and a maximum likelihood method is developed for inferring the parameters from rate-modulated spike trains.
***********************************
White-matter connecting anterior insula to nucleus accumbens is associated
with functional brain activity and risk-taking behavior
Josiah K. Leong
josiah@stanford.edu
Introduction
Neuroimaging studies utilizing FMRI have implicated activity in the nucleus accumbens
(NAcc) and anterior insular cortex in anticipation of uncertain rewards (1). Their whitematter connections, however, have not been mapped in humans (2). These connections, as
well as modulatory projections from ventral tegmental area (VTA) dopamine and medial
prefrontal cortex (MPFC) resist mapping with atlas-based tractography approaches (3).
To map white-matter paths from the anterior insula to the NAcc for the first time in humans, and determine the association of their structure with FMRI recordings during risky
choices, we combined diffusion weighted imaging and probabilistic fiber tractography. We
additionally sought to replicate previously reported MPFC-NAcc and VTA-NAcc pathways (4). We validated each of these pathways using a novel method involving a Virtual
Lesion (5). Thus, we present a novel approach for tracking, validating, and quantifying the
structural characteristics of white-matter pathways in circuits associated with motivated
behavior. We further relate the circuit’s structure with functional brain activation during
28
risky choice.
Method
In a community sample of 32 healthy adults (14 F, age range = 21-85), we acquired HARDI,
FMRI during a gambling task, and a T1-weighted scan for alignment and region of interest
(ROI) identification. To define anatomical ROIs as seed areas for tracking, we processed
subjects T1 scans with FreeSurfer (6). NAcc ROIs were identified from subcortical tissue
classification and anterior insula ROIs were derived from cortical parcellation (7).
Fiber tracking between anterior insula and NAcc ROIs was performed using constrained
spherical deconvolution-based probabilistic tracking (8). Fiber pathways were generated
by randomly seeding a voxel in a starting ROI and tracking until the fiber reached the
end-pair ROI. Fibers leaving the white matter volume were discarded. Fiber tracking
for MPFC-NAcc and VTA-NAcc pathways was performed using the method reported by
Samanez-Larkin (2012). A probabilistic tractography algorithm (ConTrack) was used to
generate a set of 50,000 candidate fibers connecting the ROI pairs within each hemisphere
(9). Candidate fibers were scored using the ConTrack scoring algorithm and the top-scoring
1% of fibers were retained. We tested the statistical validity of each pathway using a novel
method called Linear Fascicle Evaluation (LiFE) with Virtual Lesions. Indices of tract
coherence (e.g., fractional anisotropy or FA) from validated tracts were correlated with
FMRI and behavior across subject s.
Results
Each white-matter pathway of interest was successfully tracked in all subjects. Virtual Lesion analysis validated our three main pathways. Regression analyses revealed individual
differences in right hemisphere anterior insula-NAcc tract coherence were associated with
acceptance of postively-skewed gambles (β = -0.40, p = 0.02). Critically, this association
was statistically mediated by NAcc activation during risky choice. Tract coherence was
associated with decreased NAcc activation during the decision period for positively-skewed
gambles (?? = -0.35, p = 0.03), and NAcc activation was associated with choosing to
gamble (β = 0.46, p = 0.003). This indirect effect reduced the direct association between
tract coherence and gambling to nonsignificance (β 0 = -0.24, p = 0.14; ?? = -0.40, p =
0.02), consistent with full statistical mediation.
Conclusion
We tracked and validated for the first time a white-matter pathway connecting the anterior
insula to the NAcc in humans. In addition, we validated previously observed projections
from the VTA and MPFC. Structural characteristics of this circuit related to functional
brain activations and behavioral risk-taking. Building from tracts observed in comparative
29
research, these results raise the possibility of linking white-matter properties to neural activity in dopaminergic reward circuits.
References
1) Knutson, B. (2008), Anticipatory affect: neural correlates and consequences for choice,
Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences,
vol. 363, no. 1511, pp. 3771-86.
2) Chikama, M. (1997), Insular cortical projections to functional regions of the striatum
correlate with cortical cytoarchitectonic organization in the primate, Journal of Neuroscience, vol. 17, no. 24, pp. 9686-705.
3) Haber, S.N. (2010), The reward circuit: linking primate anatomy and human imaging,
Neuropsychopharmacology, vol. 35, no. 1, pp. 4-26.
4) Samanez-Larkin, G.R. (2012), Frontostriatal White Matter Integrity Mediates Adult
Age Differences in Probabilistic Reward Learning, Journal of Neuroscience, vol. 32, no.
15, pp. 5333-5337.
5) Pestilli, F. (2014), ’Evaluation and statistical inference for human connectomes’, Nature
Methods, vol. 11, no. 10, pp. 1058-63.
6) Fischl, B. (2004b), Automatically parcellating the human cerebral cortex. Cerebral Cortex, vol. 14, pp. 11-22.
7) Destrieux, C. (2010), Automatic parcellation of human cortical gyri and sulci using
standard anatomical nomenclature. NeuroImage, vol. 53, no. 1, pp. 1-15.
8) Tournier, J.D. (2007), Robust determination of the fibre orientation distribution in diffusion MRI: non-negativity constrained super-resolved spherical deconvolution, NeuroImage,
vol. 35, no. 4, pp. 1459-72.
9) Sherbondy, A.J. (2008), ConTrack?: Finding the most likely pathways between brain
regions using diffusion tractography., Journal of Vision, vol. 8, no. 9, pp. 116.
10) Wu, C.C. (2011). ’The affective impact of financial skewness on neural activity and
choice’. PloS One, 6(2), e16838.
Joint work with Franco Pestilli, Gregory R. Samanez-Larkin, Brian Knutson
***********************************
30
Decoding the temporal dynamics of left mid-fusiform gyrus activity during
word reading
Yuanning li
yuanninl@andrew.cmu.edu>
The nature of the visual representation for words has been fiercely debated for over 150
years. We used direct brain stimulation, pre- and post-surgical behavioral measures, and
intracranial electroencephalography to provide support for, and elaborate upon, the visual word form hypothesis. This hypothesis states that activity in the left mid-fusiform
gyrus (lmFG) reflects visually organized information about words and word-parts. We
applied classification methods to analyze the event-related potentials (ERPs) from the
electrophysiological data. We found that information contained in early lmFG activity
was consistent with an orthographic similarity space. Furthermore, disrupting lmFG activity through stimulation or surgical resection led to impaired perception of whole words
and word-parts. Finally, classification for individual visual word stimulus based on timewindowed ERP signals demonstrated that early lmFG response to words reflected a coarse
visual representation organized by orthographic similarity, while later activity reflected a
finer representation, capable of individuation. These results provide strong support for the
visual word form hypothesis and demonstrate lmFGs role in a dynamic coarse-to-fine shift
in word representations.
Joint work with Elizabeth A. Hirshorn, Michael Ward, Ellyana Kessler, Breana Gallagher,
R. Mark Richardson, Julie A. Fiez, and Avniel Singh Ghuman
***********************************
Copula Models of Multivariate Point Process for the Analysis of Ensemble
Neural Spiking Activity
Hualou Liang
hualou.liang@drexel.edu
We present a new statistical technique for analyzing neural dependence of simultaneously
recorded multiple spike trains. The method is based on the copula models that can account for both the marginal distribution over spiking activity of individual neurons and
the joint distribution over ensemble activity of multiple neurons. Considering the popular
generalized linear models (GLMs) as marginal models, we develop a general and flexible
likelihood framework that uses the copula to integrate separate GLMs into a joint regression analysis. The resulting joint analysis essentially leads to a multivariate analogue of the
marginal GLM theory and hence an efficiency gain in the model estimation. In addition,
we show that Granger causality between neural spike trains can be readily assessed via
the likelihood ratio statistic. The performance of the estimation procedure is validated
by extensive simulations, and compared favorably to the widely used GLMs. Finally the
31
method is applied to spiking activity of simultaneously recorded frontal eye field (FEF)
and inferotemporal (IT) neurons of a monkey performing an object-based working memory
task. We observe significant Granger causality influence from FEF to IT, but not in the
opposite direction, indicating the role of the FEF in the selection and retention of visual
information in working memory. The results of the real neural data analysis suggest that
spatial selection in FEF precedes object identification in IT during memory task and that
our approach has the potential to provide unique neurophysiological insights about network
properties of the brain.
Joint work with Meng Hu, Kelsey L. Clark, Xiajing Gong, Behrad Noudoost, Mingyao Li,
Tirin Noore
***********************************
Long-range functional connectivity in the epileptic human brain using the
spike-triggered impulse response
Beth A. Lopour
beth.lopour@uci.edu
Functional connectivity analysis has revealed important characteristics of the networks that
contribute to epileptic seizures. For example, it has been shown via fMRI that epilepsy
causes changes to both local (near the site of seizure onset) and long-range connectivity.
Here we take advantage of a unique situation in which we can bilaterally record multi-unit
activity (MUA) and local field potential (LFP) from the brains of patients with intractable
seizures who are surgical candidates. We assess MUA/LFP functional connectivity between pairs of electrodes in distinct regions of the brain, and we find that the spatial
characteristics are consistent with general notions of anatomical connectivity, e.g. selfconnections are most common, followed by ipsilateral connections within the same lobe of
the brain. Further, the timing of the impulse response appears to be related to the type of
connection, e.g. contralateral responses are delayed relative to the timing of the multi-unit
spike. However, while the spatial characteristics of the impulse response are consistent
with anatomical connectivity when measured across all subjects, we find distinct, localized
networks within each subject. We hypothesize that these connections are related to the
unique pathological epileptogenic network(s) in each subject, and therefore this work may
have implications for the diagnosis and surgical treatment of epilepsy.
Joint work with Richard J. Staba, John M. Stern, Itzhak Fried, Dario L. Ringach
***********************************
A Statistical Approach for Seizure Risk Forecasting
32
Behrouz Madahian
bmdahian@memphis.edu
About 30% of all patients with epilepsy experience seizures that are unresponsive to medication or resective surgery. Although seizure frequency in these patients could be moderate,
the constant threat of an impending seizure prevents them from doing several daily routine
activities. Hence, an effective seizure forecasting system that identifies periods associated
with elevated seizure risk would improve the quality of life in patients with intractable
seizures. Early seizure warning would help patients avoid potentially risky activities (e.g.
driving or swimming), and enable individually tailored closed-loop anti-seizure therapies.
Research over the past decade has shown that seizures are not quite random events and
that statistical models can be applied to predict seizures to some extent. The goal of a
seizure prediction algorithm is typically to differentiate interictal (baseline) and preictal
(pre-seizure) periods. In this study, a statistical algorithm for anticipating s eizures based
on a random forest classifier is proposed and tested on prolonged Intracranial EEG recordings in dogs. The possibility of classifying preictal and interictal states are explored and
results from out-of-sample testing showed perfect sensitivity and a very low false positive
rate for the proposed algorithm.
***********************************
Utilizing time-varying graphs for discovering dynamic functional connectivity
Margaret Mahan
mahan027@umn.edu
Functional connectivity analyses commonly take advantage of graph theoretical properties
to assess the brain as a network. However, these analyses capture static graph measures
without incorporating the inherently temporal aspect of brain function. Therefore, to
examine the brain as the dynamic network it is, functional connectivity analyses need to
include the temporal dimension as part of graph construction. Time-varying graphs are
a valuable tool for such purposes. These graphs are characterized by incorporating the
temporal dimension into the graph components (i.e. edges, nodes). For example, an edge
between node A and node B is only present during certain time points (say, 1-3, 5, & 8).
Here, we present two methods, lagged-based and window-based, to construct time-varying
graphs from magnetoencephalography (MEG) data and apply these methods to assessing
dynamic functional connectivity across the lifespan.
MEG recordings were collected from 140 women (32-97 years old; age-grouped into 12
groups: ¡ 40, 40:5:90, ¿ 90 years old) for two sessions. MEG time series were prewhitened
using ARIMA(50,1,1) to yield practically white noise innovations, and nodes were defined
to be the individual sensors (n = 248). To construct the lagged-based time-varying graph,
each subjects crosscorrelations (CCs) were computed for all sensor pairs (n = 30,628) for
33
k lags and significant CCs were retained for further analysis. Then, a combined age-group
correlation coefficient was calculated for each sensor pair and lag combination. Ultimately,
four lagged-based time-varying graphs were constructed for each of the twelve age-groups,
from combinations of unweighted/weighted and undirected/directed graph-types. To construct the window-based time-varying graph, zero-lag crosscorrelations were computed for
non-overlapping time windows. Undirected graphs of both unweighted and weighted were
const ructed for multiple time windows. Graph metrics for all constructed time-varying
graphs were calculated for each age-group and session. To determine reliability, an intraclass correlation between sessions was calculated for each metric. Discussion focuses on
evaluating the two methods for constructing time-varying graphs, exploring the reliability
of metrics across the method and parameter choice, and the patterns of dynamic functional connectivity across the age-groups. Finally, aims towards constructing a model of
how brain communication patterns change with age, in such a way that brain function
remains healthy, are explored.
Joint work with Apostolos P. Georgopoulos.
***********************************
Decoding velocity with kinematic models and direct regression
Francesca Matano
fmatano@andrew.cmu.edu
We compare two approaches to decoding velocity, and other kinematic variables, from neural activity in primary motor cortex (MI): a conventional state-space (Kalman filter) model
based on improved kinematic models, and a direct or forward regression approach which
reverses the relationship between the stimulus and the response. In the first, Bayesian
approach, we sought to improve decoding by developing better state-space models for the
evolution of the kinematic variables over time. The resulting models are much better fits
than the usual random-walk-in-velocity model to kinematic data in a hand-reaching experiment. When used in Kalman or particle filters to estimate velocity from neural data,
however, the results are no better than a random-walk model; we argue that this is a
general problem and not a specific defect of our model. Our forward model, by contrast,
directly regresses current velocity on past kinematic variables and current neural activity.
We stabilized the regression using both the ridge penalty, and a variant which separately
penalized neural and kinematic coefficients. Either way, we selected neural tuning curve
models to minimize the error of predicting trajectories. This forward method is fast, simple, and out-performs Kalman filters.
Joint work with Steven Chase, Cosma Shalizi, Valerie Ventura
***********************************
34
Quantifying spike train oscillations: biases, distortions
Ayala Matzner
ayalamatzner@gmail.com
Estimation of the power spectrum is a common method for identifying oscillatory changes
in neuronal activity. However, the stochastic nature of neuronal activity leads to severe
biases in the estimation of these oscillations in single unit spike trains. Different biological and experimental factors cause the spike train to differentially reflect its underlying
oscillatory rate function. We analyzed the effect of factors such as the mean firing rate
and the recording duration on the detectability of oscillations and their significance, and
tested these theoretical results on experimental data recorded in Parkinsonian non-human
primates. The effect of these factors is dramatic, such that in some conditions, the detection of existing oscillations is impossible. Moreover, these biases impede the comparison
of oscillations across brain regions, neuronal types, behavioral states and separate recordings with different underlying parameters, and lead inevitably to a gross misinterpre tation
of experimental results. We introduce a novel objective measure, the ”modulation index”,
which overcomes these biases, and enables reliable detection of oscillations from spike trains
and a direct estimation of the oscillation magnitude. The modulation index detects a high
percentage of oscillations over a wide range of parameters, compared to classical spectral
analysis methods, and enables an unbiased comparison between spike trains recorded from
different neurons and using different experimental protocols.
Joint work with Izhar Bar-Gad
***********************************
Regression Spline Mixed Models for Analyzing EEG Data and Event-Related
Potentials
Karen Nielsen
karenen@umich.edu
Analysis of EEG data tends to be a nuanced, subjective process. For example, filtering
is common, primarily to reduce noise, but a wide variety of filters are available with only
heuristic (not theoretical) recommendations for use. This work focuses on Event-Related
Potentials (ERP), which generally involve waveforms with only one or a few oscillations.
Since EEG readings consist of highly-correlated multi-channel readings, an ideal modeling approach should make use of this structure. Here, we will show how Regression Spline
Mixed Models (RSMM) can combine the features of splines with a hierarchical framework to
explore EEG data at any of the many levels that are collected and of interest to researchers.
35
Joint work with Rich Gonzalez
***********************************
Spontaneous fluctuations in networks of spiking neurons
Tomokatsu Onaga
onaga@scphys.kyoto-u.ac.jp
Spontaneous fluctuation in neuronal activity is widely observed in the cortical neural network not only in vivo and also in vitro. In recent study, it was proposed that a rich variety
of temporal dynamics in neuronal firing can be utilized for working memory and motor
control in the brain. When considering an isolated network of neurons, the firing rates
remains constant if the interactions among neurons are weak. However, if the interactions
are strong, the network may exhibit non-stationary fluctuation in the firing rates even in
the absence of external inputs. Recently we have revealed that the self-exciting process
may exhibit a transition above which the rate of event occurrence fluctuates spontaneously.
The condition of the transition does not depend on the time course of interection, but is determined solely by the strength of interaction. In this contribution, we apply this analysis
to a network of spiking neurons to explore the condition for the stationary-nonstationary
transition.
Joint work with Shigeru Shinomoto
***********************************
Early detection of human epileptic seizures using MUA and LFPs from
intracortical microelectrode arrays
Yun Park
yun_sang_park@brown.edu
Reliable early seizure detection could significantly improve the therapeutic alternatives
for people with pharmacologically resistant focal epilepsy. Most current approaches rely
on scalp or intracranial EEG, with potential for improvement in false positive rates and
detection latencies. Here, we examined early seizure detection based on intracortical neural signals, recently made available by microelectrode array (MEA) recordings in people
with epilepsy. In particular, we studied the use local field potentials (LFPs) and multiunit
activity (MUA) recorded from 96-channel MEAs. We used a patient-specific framework
for the detection that consisted of (1) extraction of LFP and MUA; (2) feature extraction from LFP and MUA signals; (3) nonlinear cost-sensitive SVM classification of ictal
and interictal states based on features extracted from LFP, MUA, or their combination;
and (4) postprocessing. LFP features included statistical summaries of power spectrum
36
in seven frequency band s and measures related to spatial coherence. MUA was defined
as the count of threshold crossing events in 0.1 s time bins. MUA features consisted of
statistical summaries of the counts and coherence measures. We assessed the frameworks
performance on data including 17 seizures and 38.2-hour interictal recordings from six
patients: six gamma-band type seizures (i.e. seizures characterized by 40 60 Hz LFP
oscillations) from one patient, and 11 spike-wave complex type seizures from five patients.
Seizure onsets were determined based on ECoG recordings. Under cross-validation, detection based only on LFP features produced 100% sensitivity, 0.10 false alarms per hour, and
an average latency of 3.7 s. (median: 3.0 s). Detection based on MUA features achieved
100% sensitivity, 0.13 false alarms per hour, and an average latency of 4.5 s (median: 4.0
s). Furthermore, detection based on the combination of LFP and MUA features resulted
in shorter latencies: 100% sensitivity, enhanced latency (average: -5.4 s; median: 3.0 s),
and six false alarms (0.16 per hour). Importantly, three of these false alarms were related
to epileptiform activity, two to subclinical seizure events, and one to artifact. Our findings
indicate that the combination of MUA and LFP signals recorded from MEAs may lead to
reliable human epileptic seizure detection by improving latency and reducing the number
of false alarms.
***********************************
A Flexible Model with Multivariate Extensions for Neural Spike Trains
Reza Ramezan
rramezan@fullerton.edu
We present Skellam Process with Resetting (SPR), a new model for the analysis of neural
spike trains. SPR is the difference between two Poisson processes with an adjustment for
the neural refractory period. We show that modeling spike trains as realizations of the
records of SPR is efficient, powerful, and informative. One interesting property of SPR is
that it allows for flexible behavior of the inter-spike interval distribution, including a wide
range from exponential to Inverse Gaussian.
A challenging problem at the juncture of statistics and neuroscience is the simultaneous
analysis of multiple neural spike trains within a multivariate point process framework–
particularly modeling negative correlation. We show that SPR has easy-to-implement
multivariate extensions, which allow for both positive and negative correlations. SPR also
generalizes the traditional inhomogeneous Poisson process, and the inhomogeneous Inverse
Gaussian process in modeling ISI distribution. Simulations, and real data analyses based
on computationally efficient algorithms show promising results of this new flexible model
for neural spike train data.
Joint work with Paul Marriott, Shojaeddin Chenouri
37
***********************************
On a reduced model of spinal cord stimulation for chronic pain: selective
relay of sensory neural activities in myelinated nerve fibers
Pierre Sacre
p.sacre@jhu.edu
Chronic pain affects about 100 million adults in the US. Despite their great need, neuropharmacology and neurostimulation therapies for chronic pain have been associated with
suboptimal efficacy and limited long-term success as their mechanisms of action are unclear. Over the past decades, detailed computational models have been used to understand
the effects of electrical neurostimulation on dorsal column fibers, the first target of neurostimulation in the complex pain system. Although these models reproduce some observed
behaviors, none of these models—to our knowledge—include the fundamental underlying
sensory activity (either normal or pathological) traveling in these nerve fibers. We developed a (simple) simulation testbed of electrical neurostimulation of myelinated nerve fibers
with underlying sensory activity and we reduced it to allow for tractable mathematical
analysis. This poster reports our findings so far. Interactions between stimulation-evoked
and underlyi ng activities are mainly due to collisions of action potentials and losses of
excitability due to the refractory period following an action potential. In addition, intuitively, the reliability of sensory activity decreases as the stimulation frequency increases.
This first step opens the door to a better understanding of pain transmission and its modulation by neurostimulation therapies.
Joint work with Sridevi V. Sarma, Yun Guan, William S. Anderson
***********************************
Restoration of normal striatal dopamine responses with NMDA/AMPA
receptor blockade in parkinsonian monkeys
Arun Singh
arun.singh@emory.edu
In non-human primate models of advanced parkinsonism, medium spiny neurons (MSNs)
are markedly hyperactive and often exhibit reversal of levodopa-induced firing rate changes
(inversion of dopamine responses) in correlation with levodopa-induced dyskinesias (Liang
et al., 2008). Hyperfunction of striatal glutamate signaling is thought to play a primary
role in the mechanisms of dyskinesias. However, the impact of glutamatergic transmission
on abnormal MSN responses to dopamine has not been studied. The electrophysiological
effects of striatal NMDA or AMPA receptor antagonism were studied in four awake, behaving, parkinsonian rhesus monkeys. The competitive NMDA antagonist LY235959 or AMPA
38
antagonist NBQX was delivered by microinjection at the site of extracellular recordings
in the striatum of monkeys followed by systemic levodopa administration (s.c.) during
the recording session. The doses of antagonist were determined on the basis of in vitro
tests for selectivity of receptor binding and in vivo tests of magnitude of firing frequency
reduction. Behavioral effects of the antagonists were also evaluated with systemic administration. We found that the reduction of MSN baseline activity via local microinjection
of LY235959 or NBQX completely abolished the abnormal inversions of firing rate changes
induced by dopamine inputs. Comparisons with the vehicle alone as control confirmed the
specific effect of the local drug microinjection. These NMDA/AMPA antagonists also reduced dyskinesias following systemic injections, demonstrating correlated behavioral effects
in the same animals that exhibited physiological effects. These results indicate that the
ionotropic glutamate transmission primarily controls the MSN activity in the parkinsonian
state. This has profound implications for the striatal pathology developed in advanced PD
that is associated with abnormal responses to dopamine.
Support contributed by NS045962
***********************************
Task-specific Neuronal Ensembles Improve Coding of Grasp
Ryan J. Smith
ryansmith@jhu.edu
Reaching and grasping motions require the activation and coordination of functional networks of neurons. Models of motor-related neuronal activity have commonly focused on the
encoding of behavioral signals by individual neurons independently within a population.
Interactions among the population may provide additional insights into the encoding of behavior by individual neurons as well as encoding of behavior by the population as a whole.
As recording technologies improve and the number of simultaneously observable neurons
increases, models of neuronal activity must also expand to better incorporate information
contained within the ensemble structure.
Spiking activity of individual neurons is often modeled as covarying with relevant motor
variables but independent of the activity of the remaining observed population. Recent
studies have demonstrated that accounting for effective connectivity among simultaneously
observed neural signals can result in dramatic improvements to encoding performance at
the single neuron level. In this study, we extend these models to enable effective connectivity to vary with motor behavior. This model structure then allows for behavior-related
activity to be encoded both within the firing rate of individual units as well as in the
effective structure of the ensemble.
39
We recorded spiking activity from multiple microelectrode arrays in primary motor cortex
(M1) and premotor cortex (PM) of two rhesus monkeys performing a center-out reachand-grasp task. Using generalized linear models (GLMs), we constructed point process
encoding models of firing activity that account for task-specific baseline firing activity as
well as task-specific effective connectivity. Models were evaluated in terms of their encoding capabilities as well as their ability to properly classify the grasp being performed.
Incorporating these task-specific ensemble effects significantly improved decoding performance over alternative models. Additionally, task-specific changes to effective connectivity
appear to reflect only small deviations from a common underlying connectivity structure.
Joint work with Adam G. Rouse, Alcimar B. Soares, Marc H. Schieber, Nitish V. Thakor
***********************************
Sleep apnea detection using a reduced set of measurements and symbolic
time series analysis
Chrysostomos D. Stylios
stylios@teiep.gr
The most prevalent sleeping disorder, affecting 2-4% of the adult population, is obstructive
sleep apnea (OSA) [1]. In spite of its frequent appearance, especially among men, it is
surprisingly passes undetected to about 90% of the cases and thus untreated. The main
reason for that is the fact that breathe stoppages do not cause a full awakening of the patient. Another reason is that widely diagnosis means an overnight OSA test, which usually
requires for the patient to sleep for at least two consecutive nights at a sleeping lab for the
acquisition of polysomnographic (PSG) signals that constitute the gold standard for OSA
detection. Since OSA is related to other more serious health problems as well as excessive
daytime sleepiness and fatigue which has reported as cause of traffic accidents.
PSG analysis is very efficient but quite uncomforting for patients. Therefore a more practical way is needed to detect OSA in the general population without the need for are
overnight PSG. With the advances in sensors and mobile technology this is close to becoming a reality [2]. Here we present a new approach to OSA detection, which combines a
single measurement of the ECG acquired and stored with the help of a smartphone along
with a light data mining algorithm for the detection and quantification of OSA.
The approach is based on a well-known algorithm from the field of the symbolic time series
analysis, the Symbolic Aggregate approXimation (SAX) algorithm [3] and an invariant
bag-of-patterns representation [4], inspired from from the field of information retrieval, for
the extraction of OSA-sensitive features. The very low computational requirement of the
algorithm makes it ideal for smartphone applications. Therefore the smartphone can not
only acts as a recording devices coupled with an off-the-self ECG sensor but also as a
40
diagnosis tool alerting for further investigation and treatment. Figure 1 summarizes the
feature extraction process with the feature extraction stage also visualized using intelligent
icons [5].
The proposed approach was tested on a set of single channel ECGs with very promising results compared to other computationally more demanding algorithms which employ
specifically tailored features and state of the art classification algorithms [2].?
Joint work with George Georgoulas, Petros Karvelis
***********************************
A Novel Method for Seizure Localization in Medically Refractory Epilepsy
Patients
Sandya Subramanian
ssubra15@jhu.edu
Epilepsy is a neurological disorder characterized by abnormal electrical activity in the brain,
called seizures. The region of the brain that causes the seizures is called the epileptogenic
zone (EZ), and may differ for each patient. Epilepsy affects 60 million people worldwide, of
whom over 30% of cases do not respond to medication or have medically refractory epilepsy
(MRE). There are currently two treatments for patients with focal MRE: surgical resection,
in which the EZ is removed in hopes of stopping seizures, or neurostimulation, in which
the EZ is electrically stimulated to suppress seizures. Both treatments depend on accurately localizing the EZ, and when successful, both treatments are life-changing. EZTrack
generates a simple-to-read heat map overlaid over the patient’s brain scan that displays to
clinicians which regions of the brain are highly likely to be in the EZ. EZTrack was tested
in a small retrospective study that included 19 patients who had resective surgeries. To
test its efficacy, we compared EZTrack’s “red-hot” regions (ROI) to resected regions using
electrocorticographic data from only 2 seizure events per patient. If the complete ROI
was resected, then we predicted a successful surgery; else we predicted a failure. For 19
patients, EZtrack achieved a prediction accuracy of 95%. It also correctly predicted all 8
failed surgeries, which is especially important to indicate to clinicians whether to resample
different areas of the brain before deciding to resect.
***********************************
Some thought experiments on the applicability of Granger causality and
Directed Information in statistically inferring the direction of information
flows
Praveen Venkatesh
41
vpraveen@cmu.edu
Not without controversy, Granger causality and, more recently, Directed Information, have
emerged as measures of the “causal influence” of one stochastic process on another. Taking a step further, many recent works interpret obtained direction of causal influence as
the direction of “information flow” in the neural circuit. To test the interpretation on
information-flow directions, this paper constructs two simple theoretical examples to test
whether these causal-influence measures predict the correct direction of information flow.
To better define and distinguish these terms, it is useful to think of them in the context of
the question of how the brain computes. We might, for instance, seek to describe the brain
as a block diagram of discrete computational units. In such a picture, each computational
unit receives information, processes it and then passes it on to another unit. In order to
arrive at such an understanding of the brain, a natural method is to probe it and apply
measures of directed causal influence (such as Granger causality) to the time series data
obtained from the probes. The question we ask is: must the message flow in the direction
of greater causal influence?
Our counterexamples are based on a simple feedback system where a transmitter communicates to a receiver using a well-known strategy pioneered by Schalkwijk and Kailath in
1966. Here, the “ground truth” for the direction of information flow is known by construction: from the transmitter to the receiver. We show that for reasonable values of
model parameters, even for this two node problem, the direction of information flow can
be opposite to the direction indicated by Granger causality and Directed Information. We
conclude that while it might be reasonable to infer direction of causal influence using these
techniques, one needs to exercise care in interpreting the direction of causal influence as
the direction of information flow.
Joint work with Pulkit Grover
***********************************
Orbitofrontal Cortex and Hippocampus Role in Bias Under Uncertainty
Doran Walsten
dwalste1@jhu.edu
Being able to make decisions under uncertainty is an important aspect of our lives. Both
the orbitofrontal cortex for its role in decision making and hippocampus for its role in short
term memory play an important role for examining how history biases our decisions in these
situations. This study examines oscillations in orbitofrontal cortex and hippocampus as
measured by stereotactic electroencephalography in human subjects playing a gambling
card game. For the task, patients must decide how much to bet for their card being higher
42
than a hidden computer’s card. Trials with 50% chance to win are split by high and low
bets. Significant oscillation differences are clustered across time and frequency and are
assessed using a permutation test. Before the subject even sees their card on these trials,
the power in orbitofrontal cortex (30-50Hz, 0.8s - 1s before) and hippocampus (10-30Hz,
0.7 - 1s before) correlate with the subject’s future bet. Specifically, higher power in both
areas correlate with a higher chance to bet high. This relationship indicates that the orbitofrontal cortex and hippocampus activity encodes a bias on our future decisions when
uncertain choices are given.
***********************************
Coordinated neocortical activity at cellular resolution during visual processing
Zhengyu Ma
zhengyuma@wustl.edu>
The highly interconnected nature of cerebral cortex supports the hypothesis that cortical
function emerges from coordinated neural activity across scales of space and time. Testing
this tantalizing coordination hypothesis ultimately requires recording neural activity at
spatial scales from synapses, dendrites, neurons, microcircuits to brain regions and during
sensory processing.
Here we performed two-photon population calcium imaging of layer 2/3 neurons in primary
visual cortex of awake and behaving mice during three conditions of visual stimulation:
black screen, drifting grating, or natural movie. For each mouse and stimulus condition we
obtained the inferred spike trains from some 100 closely-spaced neurons for several minutes.
We analyzed the population of spike trains for each mouse and condition with respect to
(i) the statistical properties of individual spike trains, (ii) the pairwise correlation of spike
trains, and (iii) the coordination across neurons and time for all spike trains of a given
data set.
The analysis of the population of spike trains revealed three important features. First,
spike trains showed a broad distribution of mean rates and highly irregular spiking. The
latter resulted in a broad distribution of the coefficients of variation of the inter spike intervals with a population mean larger than one. Second, pairs of spike trains were weakly
correlated resulting in a distribution of small values of cross correlation coefficients for all
pairs. Third, neuronal avalanches, which are cascades of contiguous spikes within the population of imaged neurons, had power law size and duration distributions. Furthermore,
avalanche sizes and durations followed a scaling relation, which is an important fingerprint
of a dynamical critical system. In addition, the statistical properties of the population
spike trains were largely independent of the stimulus condition, thus indicating a dominant
contribution from intracortical dynamics. Taken together, this collection of quantitative
43
observations of cortical population activity during visual stimulation provides valuable constraints for future models of cortical dynamics and sensory processing. We have started to
design such models.
Joint work with Zhengyu Ma, Yahya Karimipanah, Jae-eun Miller, Raphael Yuste.
***********************************
Mixed-effects spline models for modeling cortical rhythm dynamics in the
developing human brain
Matthew White
matthew.white@childrens.harvard.edu
Human electrophysiological data (EEG) acquired longitudinally during early development
provide a unique opportunity to characterize the dynamically evolving characteristics of
brain activity as a result of neural maturation. To date, fundamental aspects of neural
activity in the typically developing brain, such as cortical oscillations (rhythms), their
individual maturation raters and their inter-infant variability, remain poorly understood.
This ongoing study aims to characterize the maturation of fundamental cortical rhythms
in the developing human brain during the critical period of the first 3 years of life, using
a relatively large EEG dataset collected longitudinally at multiple time points from 6 to
36 months of life. During this period the neuroarchitecture of the human brain undergoes
profound changes, including significant reorganization of neural networks as a result of
the acquisition of increasingly complex cognitive skills. Consequently, the trajectories of
cort ical rhythm parameters may vary non-linearly with age. In addition, substantially
inter-infant variability of rhythm trajectories is expected, given a wide range of unique
early experiences that may influence neural maturation. Statistical models that capture
potential non-linearities and inter-subject variability of neural trajectories are, therefore,
desirable.
Mixed-effects spline regression models represent a promising framework for describing the
non-linearity of cortical rhythm trajectories (via the spline representation) while accounting
for the variability of individual infant trajectories (via the inclusion of subject-specific random effects). We developed spline-based mixed effects models to describe cortical rhythm
frequency, amplitude and corresponding rhythm-specific network connectivity as a function of age. These models were estimated from a preliminary dataset of high-density,
non-task related EEGs from typically developing infants collected at 6, 9, 12, 18, 24 and 36
months of age. Subject-independent parameter trajectories were modeled by splines and
subject-dependent contributions were modeled by random effects (random intercept and
slope). The Akaike Information Criterion (AIC) was used to select an optimal combination
of model parameters, including the spline (piecewise polynomial) order, number of knots
44
(connecti on points of polynomial pieces), and the number of random effects (random intercept versus random intercept and random slope for age).
The AIC-based optimization resulted in parsimonious, distinct models for individual cortical rhythms. For example, frequency trajectories in the range of the gamma (30-80 Hz)
and beta (13-30 Hz) oscillations were best described by linear spline models with an internal knot at 24 months that also included a random intercept (gamma and beta) and
slope (gamma). These models show a relatively small increase in rhythm frequencies up
to 24 months of age followed by a rapid increase in these frequencies after 24 months. In
contrast, the trajectory of the delta oscillation (¡1-4 Hz) was best described by an interceptonly model with a random intercept (i.e., the delta oscillation is constant with respect to
age). These models reflect potentially distinct rhythm maturation rates and inter-infant
variability. The delta oscillation, predominantly associated with sleep, may already be
robust at birth and may not vary significantly with age and across infants. Consequently,
an intercep t-only model may adequately describe the dynamics of this oscillation. In contrast, the gamma and beta oscillations, which may change significantly as a function of age
as a result of the development of cognitive function and may vary substantially between
infants, are best described by linear mixed-effects models that include an internal knot at
24 months. Therefore, mixed-effects spline regression models provide a promising statistical framework for describing the dynamics of the electrophysiological correlates of neural
maturation during the first 3 years of life.
Joint work with Charles A. Nelson, Catherine Stamoulis
***********************************
Information coding through adaptive control of synchronized thalamic
bursting
Clarissa J. Whitmire
clarissa.whitmire@gatech.edu
Beyond acting as a simple relay from the periphery to cortex, the thalamus acts as a gate
for the peripheral signals, controlling what does and does not get transmitted to cortex.
Furthermore, this gating is dynamic, and can be influenced through both bottom-up sensory influence, and top-down mechanisms related to wakefulness and attention. In this
work, we explored the bottom-up effect of stimulus adaptation on the encoding of features
in the whisker thalamocortical circuit of the fentanyl-cocktail anesthetized rat using a classic signal-in-noise paradigm. Previous work has demonstrated that adaptation can lead to
enhanced discriminability paired with reduced detectability, but the underlying mechanism
is unknown. In the context of the signal-in-noise paradigm, we investigate the role of the
level of adaptation due to the background sensory noise on thalamic spiking, burst spiking,
and synchronous firing. Increasing levels of adaptation reduce the amplitude of the evoked
45
response and effectively shift thalamic neurons from burst to tonic firing when conveying
information related to the embedded signal. Furthermore, increasing adaptation leads to
reduced levels of synchrony across pairs of simultaneously recorded neurons. These experimental results demonstrate that thalamic cells fire more burst spikes in response to signals
presented in isolation than in noise and that this leads to a higher detectability, but a lower
discriminability, as assessed using an ideal observer analysis of the thalamic unit spiking activity. Direct depolarization of the thalamic neurons using channelrhodopsin can also shift
the encoding of sensory features from burst to tonic spikes. We developed an integrate and
fire neuron with an incorporated burst mechanism to investigate the role of depolarization
on thalamic encoding. Consistent with the experimental findings, the model suggests that
the sensory noise is depolarizing the membrane potential of the simulated cell and that
this is sufficient to explain the shift in bursting. Taken together, these results suggest that
the level of sensory adaptation may have a sustained depolarization effect that dynamically gates information flow through modulations to the sensory evoked response, the burst
spiking activity, and the synchrony across neurons. Furthermore, these results could have
implications for a more comprehensive coding strategy whereby the continuity of sensation
dynamically alters the state of the thalamus based on the statistics of the encoded sensory
information to transition between processing states (i.e. detection/discrimination states).
Joint work with Christian Waiblinger, Cornelius Schwarz, Garrett B. Stanley
***********************************
Reinstatement of distributed spatiotemporal patterns of oscillatory power
during associative memory recall
Robert Yaffe
robertbyaffe@gmail.com
Reinstatement of neural activity is hypothesized to underlie our ability to mentally travel
back in time to recover the context of a previous experience. We used intracranial recordings to directly examine the precise spatiotemporal extent of neural reinstatement as 32
participants with electrodes placed for seizure monitoring performed a paired-associates
episodic verbal memory task. By cueing recall, we were able to compare reinstatement
during correct and incorrect trials, and found that successful retrieval occurs with reinstatement of a gradually changing neural signal present during encoding. We examined
reinstatement in individual frequency bands and individual electrodes and found that neural reinstatement was largely mediated by temporal lobe theta and high-gamma frequencies. Leveraging the high temporal precision afforded by intracranial recordings, our data
demonstrate that high-gamma activity associated with reinstatement preceded theta activity during encoding, b ut during retrieval this difference in timing between frequency
bands was absent. Our results build upon previous studies to provide direct evidence that
46
successful retrieval involves the reinstatement of a temporal context, and that such reinstatement occurs with precise spatiotemporal dynamics.
Joint work with Matthew S. D. Kerr, Srikanth Damera, Sridevi V. Sarma, Sara K. Inati,
Kareem A. Zaghloul
***********************************
A Probabilistic Model to Resolve Uncertainty in Clinical Sleep Scoring
Farid Yaghouby
f.yaghouby@uky.edu
Scoring sleep in polysomnographic recordings is a tedious and subjective task. Uncertainty
and variability between assessments of expert raters are the major obstacles. Hence, algorithms for automated sleep segmentation are in great demand. These algorithms either
use inherent patterns in the data to differentiate between vigilance states (unsupervised
classification) or mimic a human raters behavior by modeling labeled samples to predict
vigilance state in unlabeled data (supervised classification). Here we propose a novel technique to address three problems related to human sleep scoring: 1. The rater is confident
of scoring only some of the states; 2. The rater scores all states but is uncertain of some
epochs; and 3. Two raters score all states and epochs but with some disagreement. To
address these problems EEG, EMG, and EOG features were extracted in 30s epochs from
human-scored polysomnograms from 42 healthy human subjects in an anonymized database. A framework for quasi-supervised classification was devised in which unsupervised
probabilistic models (viz. hidden Markov models) are estimated from unlabeled training
data, but the training samples are tagged with variables whose values depend on available
scores. Variations on this theme are used to address each of the scoring scenarios and
classifier performance assessed using Cohen’s kappa statistic. The quasi-supervised classifier performed significantly better than an unsupervised model and sometimes as well as a
completely supervised model despite limited access to scores. This addresses the need for
classifiers that mimic human scoring patterns while compensating for human uncertainty.
Acknowledgement: We acknowledge support from National Institutes of Health (USA)
grant NS083218 during the writing of this manuscript.
Joint work with Sridhar Sunderam
***********************************
Exploring Spatio-temporal Neural Correlates of Face Learning
Ying Yan
47
yingyan1@andrew.cmu.edu
Faces are among the most important visual stimuli in our everyday life, and we are all
experts in learning new faces. Understanding the neural mechanisms of such expertise is
one of the fundamental goals of cognitive science. With recent functional neuroimaging,
researchers have discovered some temporal signatures of face-processing, and a spatial network of face-sensitive regions in the brain, distributed in the ventral visual cortex, superior
temporal cortex and frontal cortex. However, we still lack a joint spatio-temporal characterization in the process of learning novel faces.
In this work, we analyzed magnetoencephalography (MEG) recordings when human participants learned to distinguish two categories of faces in an on-line way. To examine whether
the MEG signals were correlated with behavioral learning, we regressed the MEG sensor
recordings across trials against the behavioral accuracy, which increased monotonically
with the trial number. In addition, we developed a structured-sparsity-inducing regression
model to facilitate inference in the face-sensitive regions in the brain space.
We found that the MEG sensor data were significantly correlated with the learning curve,
at 170-600 ms after the face stimulus onset, and peaked at around 250 ms, which may
correspond to the N250 EEG component that indexes familarity of faces. This correlation
effect was predominant in face-sensitive regions in the ventral visual cortex, whereas regions outside the ventral visual cortex did not show as strong effects. Our results revealed
the spatio-temporal dynamics in the face-sensitive areas during the online face-learning, on
a finer-grained level than previous literature, and suggested different roles of the regions
in and outside the ventral visual cortex during learning.
***********************************
Stimulus identification from fMRI scans: a statistical perspective
Charles Zheng
snarles@stanford.edu
Functional MRI studies frequently employ statistical or machine learning models to describe the relationship between stimuli and a subject’s multivoxel response.
Such encoding models can be used to predict the subject’s response to a new stimulus:
the accuracy of this prediction quantifies how well the model describes the encoding of
stimulus features to neurological responses. Furthermore, these same models can be used
to recover the stimulus presented from the brain response by solving an inverse problem
from a candidate set. In contrast to many classification strategies, this formulation allows
decoding of stimuli that were not seen in the training stage. We focus on an identification
task: the rate in which the observed stimulus can be chosen from large but finite library of
48
candidates. Performance on this identification task is often used to quantify the sensitivity
of the model to changes in the stimulus, as in Kay et al 2008.
In this work, we develop a theoretical framework for studying the problem of identification.
We observe that even under linear or approximately linear models, the model-estimates for
optimal identification differ from those that would give optimal encoding. We consequently
develop heuristics for how to improve performance for the identification task. We further
analyze how issues such as sample size, the size and distributional properties of the image library, the dimensionality of the feature space, signal-to-noise ratio and the presence
of nonlinearities affect the feasibility of identification and the interpretability of the results.
Joint work with Yuval Benjaini
***********************************
Establishing a Statistical Link Between Network Oscillations and Neural
Synchrony
Pengcheng Zhou
zhoupc1988@gmail.com
Pairs of active neurons frequently fire action potentials or ”spikes” nearly synchronous (i.e.,
within 5 ms of each other). This spike synchrony may occur by chance, based solely on
the neurons’ fluctuating firing patterns, or it may occur too frequently to be explicable
by chance alone. When spike synchrony above chances levels is present, it may subserve
computation for a specific cognitive process, or it could be an irrelevant byproduct of such
computation . Either way, spike synchrony is a feature of neural data that should be explained. A point process regression framework has been developed for this purpose, using
generalized linear models (GLMs). In this framework, the observed number of synchronous spikes is compared to the number predicted by chance under varying assumptions
about the factors that affect each of the individual neuron’s firing-rate functions. An important possible source of spike synchrony is network-wide oscillations, which may provide
an essential mechanism of network information flow. To establish the statistical link between spike synchrony and network- wide oscillations, we have integrated oscillatory field
potentials into our point process regression framework. We first extended the spike-field
association models of Lepage et al. and showed that we could recover phase relationships
between oscillatory field potentials and firing rates. We then used this new framework
to demonstrate the statistical relationship between oscillatory field potentials and spike
synchrony in: 1) simulated neurons, 2) in vitro recordings of hippocampal CA1 pyramidal
cells, and 3) in vivo recordings of neocortical V4 neurons. Our results provide a rigorous
method for establishing a statistical link between network oscillations and neural synchrony.
49
Joint work with Rob Kass
***********************************
Characterization of brain consistency via a data-driven brain parcellation
Qiong Zhang
qiongz@andrew.cmu.edu
It is of interest to examine the degree fMRI brain activations under one condition are similar to brain activations under the other, which is traditionally compared over a series of
pre-defined brain regions. We propose the use of a data-driven brain parcellation via spectral clustering to characterize brain consistency. The functional homogeneity of the brain
voxels during the clustering procedure is defined by a sequence of brain states identified in
a hidden semi-Markov model as a way to normalize trials with different number of scans.
We demonstrate the effectiveness of this method in identifying a neural level indicator of
behavior performance in a mathematical problem-solving task. We observe that subjects
who showed consistent brain patterns performed better.
Joint work with John R. Anderson, Rob E. Kass
***********************************
Prefrontal neurons represent comparisons of motion directions in the
contralateral and the ipsilateral visual fields
K. Michalopoulos
Prefrontal neurons represent comparisons of motion directions in the contralateral and the
ipsilateral visual fields. K. Michalopoulos, P. Spinelli, T. Pasternak Neurons in the lateral
prefrontal cortex (LPFC) are active when monkeys decide whether two stimuli, S1 and S2,
separated by a delay, move in the same or in different directions. Their responses show
direction selectivity reminiscent of activity in motion processing area MT, and during S2,
their responses are modulated by the remembered direction. A similar modulation, termed
comparison effect (CE), has also been observed in area MT. These parallels between the
two areas are consistent with their connectivity, although the nature of this connectivity suggests a possibility of asymmetries in the way contralateral and ipsilateral motion is
represented in the LPFC during the motion tasks. Specifically, while signals about the contralateral motion reach LPFC directly from MT of the same hemisphere, ipsilateral motion
processed by MT in the other hemisphere can only reach the LPFC indirectly via callosal
connections from the opposite LPFC. We explored the role of direct and indirect motion
signals during this task by examining responses of LPFC to contralateral and ipsilateral
stimuli during S1 and S2. During S1, responses to the contralateral motion were stronger
and preceded ipsilateral responses by 40ms, an indication of the apparent dominance of
50
direct inputs from the ipsilateral MT. The asymmetry between contralateral and ipsilateral
responses during S1 was not reflected in their DS activity, since it was equally robust for
both stimulus locations. During S2, responses to ipsilateral but not contralateral motion
were enhanced, eliminating the dominance of the contralateral stimuli observed during S1.
The CE was measured by comparing response to identical S2 stimuli on trials when S2 was
preceded by S1 moving in the same direction(S-trials) with trials when it was preceded by
S1 moving in a different direction (D-trials). ROC analysis revealed two distinct groups
of neurons preferring either S-trials or D-trials. CE effects were equally likely to occur
for ipsilateral and contralateral stimuli. These results demonstrate that the comparison
between the current and the remembered stimulus can be carried out in the LPFC even in
the absence of direct inputs from area MT.
Joint work with P. Spinelli and T. Pasternak
***********************************
Fundamental Problems in Granger Causality Analysis
of Neuroscience Data
Patrick A. Stokes
Granger causality methods analyze the flow of information between time series. The
Geweke measure of Granger causality (GG-causality) has been widely applied in neuroscience because its frequency-domain and conditional forms appear well-suited to highlymultivariate oscillatory data. Here, we analyze the statistical and structural properties of
GG-causality in the context of neuroscience data analysis.
We analyzed simulated examples and derived analytical expressions to demonstrate how
computational problems arise in current methods of estimating conditional GG-causality.
We found that the use of separate full and reduced models leads to either large biases
or large uncertainties in the causality estimates, and high sensitivity to uncertainties in
model parameter estimates, producing spurious peaks, valleys, and even negative values
in the frequency domain. We also analyzed how the generative systems properties and
frequency structure relate to the structure of GG-causality estimates. We used simulated
examples and derived analytical expressions to show that GG-causality is independent of
the receiver dynamics, i.e., the dynamics of the effect node that receives the input of the
putatively causal node. In particular, the magnitude of the receiver response is ignored by
GG-causality. This would mislead analysts in situations where the magnitude of the response is a central feature of the underlying physical or physiological process. In addition,
we found that GG-causality combines transmitter and channel dynamics in a way that cannot be disentangled without evaluating the component dynamics of the full model estimate.
The separate-model fit computation in GG-causality leads to either large bias or large uncertainties that make the interpretation of frequency-domain structure highly problematic.
51
Even if these computational issues are overcome, correct interpretation of GG-causality
values is challenging, since GG-causality ignores receiver dynamics, and is not informative of the system dynamics without consideration of the full model estimate. Our work
suggests that GG-causality analyses could be easily misinterpreted without careful consideration of these factors. Through this work we hope to provide conceptual clarification of
GG-causality and place it in the broader framework of modeling and system analysis, which
may enable investigators to better assess the utility and interpretation of such methods.
Joint work with Patrick L. Purdon
***********************************
Point process modeling of human seizures
Grant Fiddyment
Epilepsy is a serious brain disease afflicting 1% of the population. Ictal discharges (IDs)
– transient, large-amplitude changes in brain voltage – are a hallmark of the disease and
are thought to promote both seizures and epilepsy. However the mechanisms and networks
underpinning IDs are not well understood. Likewise, how IDs evolve in space and time
during human seizures also remains unclear.
Here we examine in vivo microelectrode array (MEA) recordings from eleven human
seizures. After identifying IDs with an automated algorithm, we apply an established
tool for spike train analysis: the point process generalized linear model (PP GLM). PP
GLM estimates are similar to traditional descriptive measures (e.g. correlation, coherence)
but can flexibly deal with confounded variables (e.g. spike rate) and are more physically
interpretable, Specifically, following Truccolo et al. (2005), we build a model with selfhistory-dependent (“intrinsic”) effects and ensemble-history-dependent (“spatial”) effects.
We show that both types of effect are necessary for a complete characterization of seizure
dynamics. Moreover model estimates of the IDs show two general patterns as seizures
terminate: (1) a discontinuous change in the dominant rhythm from 3Hz to 1Hz; and (2)
changes in the magnitude – but not the direction – of spatial influence.
Joint work with Uri Eden, Sydney Cash, Mark Kramer

Here - Statistical Analysis of Neuronal Data

Transcription

Similar documents

The Journal of Neuroscience Journal Club Articles

CSCC85 Lecture 4: Control Systems

Association for the Study of Neurons and Diseases REGISTRATION

TOC - The Journal of Neuroscience

Neuroembryology as a Process of PaYern FormaVon Outline The

CS534 — Written Assignment 4 — Due Nov 18th

Neuroanatomy - UCSD Cognitive Science

What is a neural network

Table of Contents - The Journal of Neuroscience

The Journal of Neuroscience Journal Club