Healthcare - AlgoAnalytics

Healthcare Analytics
October 12, 2016
Outline
About AlgoAnalytics
Areas of Application for Healthcare
Our experience
Healthcare datasets
Page 2
About AlgoAnalytics
Analytics Consultancy
• Work at the intersection of mathematics and other domains
• Harness data to provide insight and solutions to our clients
Expertise in Mathematics and Computer Science
• Develop advanced mathematical models or solutions for a wide
range of industries:
• Retail, economics, healthcare, BFSI, telecom, …
Working with Domain Specialists
• Work closely with domain experts – either from the clients side or
our own – to effectively model the problem to be solved
Page 3
Healthcare: Areas of Analytics Application
Patient Aid
Resource Allocation
Revenue Optimization
Page 4
Patient Care Aid
• Analytics helps detect serious
disorders or diseases through
medico-genomic data as well
as through image analytics
• Classification technique
enables deciding whether a
particular pathology exists in a
particular patient or not
Clinical Decision
support
• Uses huge amount of past
patient data and genomics
data for diagnosing high risk
group through risk
stratification score.
• Offers better data driven care,
coordination and follow up in
prevention of lifestyle
diseases.
Chronic Disease
Management
Population health
management
• Generating risk score through
algorithm indicating high
chance of heart failure
amongst given patient
• Reducing readmission rates –
using scoring techniques to
locate high risk customers and
preventing readmissions
through higher monitoring
• Targeting cardiac patients
most in need of intensive
follow up
Adverse event
avoidance
Page
Page5 ‹#›
• Improved disease surveillance
• Identification of patients
needing more proactive
approach to manage their
conditions.
• Longitudinal patient analysis
• Improved patient compliance
Reducing
Readmissions
Resource Allocation
• Using historical data for staffing to
reduce costs
• Having the right clinician at right
time at right place
• Predict the number of days a
patient will spend in the hospital
over the next year
• Predict readmission for the same
illness
• Predict likelihood that an elderly
insurance policy holder will pass
away within 10 months in order to
trigger end-of-life counseling
• Risk of death in surgery based on
aspects of you and your condition
• Prevent bottlenecks in urgent care/
OR by analyzing patient flow during
peak times
Work Flow Optimization
Page 6
Efficient Use of Hospital
Resources
Death
Revenue Optimization
• Forecasting of cash flows
based on claims history
• Using historical data, it uses
reimbursement analysis and
potential denials to forecast
cash.
Cash Flow
Forecasting
• Identify opportunities to
collect missing income,
including claims that are
wrongfully rejected by payers
or overdue monies from
patients.
• Insurance coverage
Billing Errors
Patient Selection
• Patients expense of
treatment
• Pricing model based on past
data of previous illnesses,
costs, etc.
• Can quote min-max range,
with and without insurance
Expected Expense
of Treatment
Page
Page7 ‹#›
• Identifying patients those are
not likely to pay in full
IMAGE ANALYTICS
Page 8
MRI Image Analytics Overview
Goal
Value of Recommendation
Use image analytics on MRI scans to detect
whether a particular pathology exists
•
•
•
•
•
Automatic segmentation
Volumetric analysis
Classification
Feature selection
Analyzing voxel intensities – texture
features
Image Normalization and
Segmentation
Volumetric Analysis,
Image Processing – Feature
Extraction using voxel values
Classifier
Feature Extraction from
segmented part based on
pathology
Page 9
Probability
(Disease)
Data Analysis &
Results
Probability
(Normal)
MRI Analytics – Aim
 Image segmentation and registration
- Automatic registration to a default map
- Automatic segmentation of brain parts
 Image processing
- Volumetric Analysis of the images
- Extract intensity based features from
the image
- Texture features like entropy, energy,
correlation, etc.
 Machine learning techniques
- Combine volumetric, textural features
and other possible features of the image
to classify
 Generalize the above process for multiple
pathologies
Page 10
Further Work:
 Consult domain expert for improving
the performance accuracy of automated
segmentation module
 Better feature selection using domain
knowledge of the pathology
 Addition of textural features to the
predictive model
 Combining unsupervised and supervised
methods
Number of features
used : 58
Diabetic Retinopathy Overview
Image
Processing
Goal
Use image analytics on fundus images to
detect whether diabetic retinopathy exists
Work done till date
• Image processing
• Feature extraction
• Blood-vessels
• Exudates
• Classification
Image-Texture
Based
Features
Blood Vessels
and Exudates
Input
Data
Ensemble Learning
Classify each image as retinopathy or not-retinopathy
retinopathy
using multiple classifying algorithms
Retinopathy grades 0 and 1 are considered
as class 0 (No-Retinopathy) and
grades 2 and 3 as class 1 (Retinopathy)
Page 11
Retinopathy – Future Work
Image processing & Feature computation
• Improvements in blood vessels extraction, exudates extraction, etc.
• Features based on hemorrhages, macula, optical disk, etc.
Feature selection
• Optimize features used for classification
Model selection & Ensemble Learning
• Model tuning
• Different techniques of ensemble learning
Further Analysis
• Trying the model on different dataset / Addition of new images as input
• Similar analysis for multiclass problem
• Change in threshold for categorizing as retinopathy
Page 12
REVENUE INTELLIGENCE
Page 13
Healthcare Revenue Intelligence – Cash Forecasting Model
Goal:
1.
Project N-day ahead cash using revenue and important KPIs
2.
Project target cash if all KPIs are at target level
3.
Find revenue leakage: (N-day Target cash – N-day Forecasted cash)
Read data
dates, cash, revenue,
KPI values, target values
Page 14
Lag the data
Lag the data by n days,
Forecast for n days then
Regression Model
Predict the cash values
for n days ahead using
data remaining after
taking lag
Three Level Optimization Problem
Team Level
Optimization
User Level
Optimization
• Determine teams best using how team performed in past n days
• Objective : Maximise Total Amount Resolved
• Optimized values : Average Amount Worked, Productivity, Effectiveness Ratio
• Determine teams best using how user and team performs in past n days
• Objective : Maximise User Level Amount Resolved
• Optimized values : Average Amount Worked, Productivity, Effectiveness Ratio
• Study the external actionable influencers affecting Productivity and ER combined
• Give actionable recommendation to achieve desirable Productivity and ER
Specific
Recommenda • Influencers divided in to two subtypes : Private and Government
tion
Each level can be used on it’s own as a separate recommendation system !!
Page 15
Daily Work Allocation
 Input : Available users and
number of Pvt. and Govt.
accounts
 Output : Accounts allocation
Total 200
accounts to
work on a
given day.
100 Private ,
100 Govt.
Divide Pvt.
and Govt.
accounts in
the proportion
determined
Find the work
allocation
percentage
from this
optimization
Page 16
Total 7 users
available for
that day to do
a job
Run 3rd level
optimization
for the
available
users
User
Govt.
Accounts
Private
Accounts
Total
Accounts
21842
7
10
17
21843
0
37
37
21844
9
17
26
21845
8
7
15
21852
35
9
44
21853
39
11
50
21867
2
9
11
Total
100
100
200
1. Available Online Datasets
2. Potential Sources of Data
HEALTHCARE DATA
Page 17
Available Healthcare Data Sets Online
Acute Inflammations
- Diagnose two diseases of the
urinary system
Arrhythmia
- Distinguish/classify presence or
absence of cardiac arrhythmia
Breast Cancer Wisconsin
- Classify cancer as benign or
malignant based on FNA
samples
Cardiotocography
- Fetal cardiotocograms
processed and classified
Dermatology
- Determine type of ErythematoSquamous Disease
Page 18
Liver Disorders
- Data sensitive to liver disorders
which may arise from excessive
alcohol consumption
Lung Cancer
- Classify three types of
pathological lung cancers
Thoracic Surgery Data
- Determine post-operative life
expectancy of lung cancer
patients
Vertebral Column assessment
- Classify orthopaedic patients
into 3 classes: normal, disk
hernia or spondylolisthesis
Mammographic Mass
- Determine benign and
malignant mammographic
masses based on BI-RADS
attributes and patient’s age
Diabetes
- Analyze factors related to
readmission and other
outcomes
Parkinson’s
E. Coli Genes
- Predict clinician’s Parkinson’s
disease symptom score on the
UPDRS scale
- Data gives characteristics of
each ORF in the E. coli genome
Post-Operative Patient
Echocardiogram
- Determine where patients in a
postoperative recovery areas
should be sent to next
- Classify if patients will survive a
year after a heart attack
Example Data sets are available online and have been studied
Healthcare Data Assets
Page 19
The Analytics Process
Once a client requirement comes in:
Define and outline the
problem statement
Business Requirement
Understand data
Data preparation
Data Situation
Value-adding Results
Explainable results
Operational outcomes
Page 20
Develop models
Evaluate performance
Analytics Solution
Seamless Integration
Work with clients to
integrate solution
Machine Learning Techniques
Page 21
Decision Trees
Kernel Learning
• Flow-chart like structure
• Maps observations of an item to
conclusion on item’s target value
• SVM extension – different kernel
functions for feature subsets
• Effective when data comes from
a variety of sources
Random Forests
Deep Learning
• Extension of classification trees
• High accuracy and efficient on
large databases
• Model high level abstractions
• Architecture composed of
multiple non-linear
transformations
Artificial Neural Networks
Clustering
• Idea analogous to biological
neural networks
• Used to discover complex
patterns in data
• Unsupervised learning to group
data into 2+ classes
• Clustering based on similarity or
dissimilarity between data points
Logistic Regression
Optimization
• Probabilistic statistical
classification model
• Binary predictor
• Modifying a system to make
some aspect of it work more
efficiently or use fewer resources
Technology:
Page 22
Analytics Projects Overview
News / Social Media
Analytics
Risk Score Modeling
Algorithmic Trading
Network Modeling of Risk
Quantitative strategies
in the Indian markets
Event-based or real-time score
that is updated in real time
Strategy Development
Brokerage Analytics
Recommender Systems
Multilanguage Sentiment
Dormancy prediction for
the upcoming quarter
Machine learning methods to
recommend items to users
Calculating sentiments of news
articles in 7 different languages
Improvements to pre-existing
algorithmic strategies
Event identification based
on topic detection
Country risk scores on
pre-decided risk factors
Cash Flow Forecasting
Work Flow Optimization
Contracts Management
Performance Manager
Cash forecasting for 90
days in future
Work Allocation for team
and team members
Automate existing system to
analyze legal contracts’ text
Forecast various KPIs concerning
operational performance
Predictive Maintenance
IoT Analytics
Assisted Living
Retinopathy Analytics
Fault detection in ovens
for energy optimization
Analytics on signals obtained
from a cloud network
Outlier detection based on
sensor input
Automated detection of
diabetic retinopathy
Network Failure Model
Analytics Engine
Clickstream Analytics
MRI Analytics
Monitor communication
of a network of routers
Connect to an infrastructure
and provide required analytics
Compute inherent features of
users based on website logs
Evidence-based decision
support system
Page 23
AlgoAnalytics
 Areas of expertise
•
Predictive analytics consultancy
•
Algorithmic and quantitative trading in financial markets
 About us
•
Founded by Dr. Aniruddha Pant in 2009
•
Firm focused on the intersection of mathematics, computer science and other domains
 Team
Dr. Aniruddha Pant
CEO
Ph.D. (Berkeley, USA)
Control theory, Machine
Learning, Analytics,
Statistics
+ 30 Quantitative Analysts
with a solid foundation in mathematics and engineering
Page 24