2015 NJ/DV HIMSS Fall Conference - PowerPoint Presentation

Slide1

2015 NJ/DV HIMSS Fall Conference

Improving Predictive Models with Machine Learning & Big Data

Slide2

Predictive Modeling in Healthcare -

Why Predict?

Use Cases: Existing Predictive M

odeling

TechniquesReducing Preventable Readmissions Population Health ManagementImproving Healthcare Predictive Models with Machine Learning and Big DataIntegrating Machine Learning and Big Data into Predictive Models Use Cases: Enhanced Predictive ModelingReducing Preventable Readmissions Population Health Management

Learning Objectives

Slide3

Predictive Modeling: Why Predict?

Predict

Predict future risk for events of interest

Plan

Plan on how to act to intervene

Measure

Measure

effectiveness of prediction and intervention

PerformDeploy plan to intervene

Predictive Modeling helps institutions anticipate risks and better prepare, organize and align resources to tackle those risks.

Slide4

Predictive Models in Healthcare

- Ad-hoc querying and reporting

- Data mining techniques

- Structured data, typical sources

- Small to mid-size datasets- Optimizations predictive analytics Complex statistical analysis Machine learning- All types of data, and many sources- Very large datasets- More real-time Predictive models in healthcare extract useful insights from the industry’s rich and expanded data sources in real time.

Slide5

Readmissions in numbers

In 2011 there were

3.3 million hospital readmissions

They contributed

$41.3 billion in total hospital costsMedicare had the largest share of total readmissions (55.9%) and associated costs (58.2%) followed by Medicaid (20.6 %) Use Case 1: Existing Predictive Modeling Techniques to Reducing Preventable Readmissions

Slide6

Use Case 1: Existing Predictive Modeling Techniques to Reducing Preventable Readmissions

LACE How it works

L=Length of Stay (LOS)A= Acuity

C= Co-morbidities based on the “

Charlson Comorbidity Index” E= Previous emergency room visits ChallengesInfo from the LACE model is generally delivered too late to have an impact – when patient is already moving to dischargeLACE has moderate predictive powerLACE alone does not provide actionable insights

Slide7

ESRD (End-stage renal disease) in numbers

26M+ Americans have kidney disease: precursor to ESRD

1 in 3 American adults at risk to develop kidney disease

~450k Americans are on dialysis

Annual medical payments for a kidney disease patient increases from $15k in Stage 3 to $70k+ in Stage 5In 2012, Medicare expenditures for all stages of kidney disease was $87B+. ~$58B was spent caring for chronic kidney diseaseUse Case 2: Existing Predictive Modeling Techniques for Population Health Management (ESRD)

Slide8

ESRD

Age

Diabetes

CHF

RaceGenderHigh BP

Digital healthcare devices have made regular health monitoring possible which opens up a wealth of information to make better predictions to prevent or plan for events.

Current ESRD prediction models take into account various demographic and clinical factors.

Use Case 2: Existing Predictive Modeling Techniques for Population Health Management (ESRD)

Slide9

Use Case 2: Existing Predictive Modeling Techniques for Population Health Management (ESRD)

ChallengesModels rely on claims data

Models take into account limited risk factorsProblem of dealing with large numbers of potential predictors:

>90,000 ICD-10 diagnostic codes, >4,000 procedures, >7,000 medications

Managing the Utility-Privacy tradeoff: Inability to join healthcare data with other sources of critical information to ensure patient privacyModerate predictive powerLimited actionable insights

Slide10

Data Explosion in Healthcare

Progress and innovation are no longer hindered by the ability to collect data

Predictive Models

EHR Data

Social MediaMobile AppsClaims DataWearable DevicesDiagnostics

Slide11

Improving Healthcare Predictive Models with Machine Learning and Big Data

The next generation of predictive models incorporate real time data, text mining, machine learning and big dataUse real-time data to make results relevant and timelyThey make predictive models actionableImprove accuracy - prescriptive and customizable analytics based on the needs of the hospital

More generalizable

across patient sub populations

Easier to implement – machine learning makes them automated

Slide12

Understanding Big Data

Big Data: Data sets so large or complex that traditional data processing applications are inadequate.

Key Themes of Big Data

Data explosion

Need to access data and store itStructured-unstructured dataNeed for big data architecture to harness it“Data Lakes” and other big data conceptsBig data tools – Hadoop and MapReduce

Slide13

Four Dimensions of Big Data

Big data is characterized by 4Vs that set it apart from traditional data

Slide14

Big Data Technical Architecture

Enables faster, scalable storage and retrieval by harnessing processing and storage capabilities of multiple nodes

Enables execution of complex algorithms across nodes and then aggregates - where predictive modeling and machine algorithms are executed

Determines what questions need data based answers and how the outputs need to be presented

StorageAnalyticsInsightsThe Emerging Big Data Stack

Slide15

Machine Learning to the Rescue!

Increased computing power and advancement in computer science have resulted in the development of sophisticated machine learning algorithms that enable intelligent mining of the big data.Machine Learning uses Big Data to Improve the Performance of Predictive Models

Slide16

Why Machine Learning?

Relationships and correlations hidden within large amounts of data can be discovered using Machine LearningAmount of knowledge available about certain tasks might be too large for explicit encoding by humans

New knowledge about tasks is constantly being discovered. It is inefficient and difficult to continuously re-design systems “by hand”Environments change over time

Machine Learning enables us to generate meaningful insights from Big Data to drive business more effectively

Slide17

Machine learning explores the study and construction of algorithms that can learn from and make predictions on data – Wikipedia

What is Machine Learning=

Slide18

Machine Learning Structure

Source: http://xiaochongzhang.me/blog/wp-content/uploads/2013/05/MapReduce_Work_Structure.png

Slide19

Use Case 1A:

Enhanced Predictive Modeling to Identify Patients at High Risk for Readmissions

Slide20

Predicting

Readmissions

L

ength Of Stay

Co-morbiditiesAcute vs. Emergent

E

D Visits in past 6 months

Predictors in LACE

Additional Predictors

in our modelOther Potential Predictors Additional variables are important predictors of readmission Sophisticated machine learning techniques like SVM enhanced predictive accuracy of the algorithmPrimary ConditionDischarged to,Admit SourcePatient Age

Financial Class# of VisitsOther Behavioral factorsOther Socio Economic factorsMedication detailsPredictive model algorithms with sophisticated ML techniques and a wide variety of predictors will exhibit better accuracy

Slide21

Predictive Model: Readmission Risk Prediction

Slide22

Use Case 2a:

Enhanced Predictive Modeling to Predict ESRD in Elderly Patients

Slide23

Current ESRD Predictive Models

ESRD

Age

Diabetes

CHF RaceGenderHigh BP

CMS hierarchical condition categories (CMS-HCC) model

Slide24

Scope

Challenges

Solution

Benefits

Identify patients at high risk of developing chronic kidney diseaseIdentify patients at high risk of transitioning from chronic kidney disease to end stage renal disease

Extensive data collection,

storage and processing

Developing trend analysis based on incomplete data

Incorporating unstructured data into predictive models that utilize cutting edge machine learning techniques

Enable wider data collection by tapping into unconventional sources like labs, radiology and personal dataUse big data architecture to store data in data lakes and use on need basisUtilize machine learning techniques to do best possible predictions with available data on case to case basis Enhanced predictive power (Higher True Positives, Lower False Positives)Key indicators increasing the patient risk for patient specific intervention strategies How we Improved the Existing Models

Slide25

Patient Demographics

Behavioral Factors

Medical History

Machine Learning Tech.

Genetic FactorsStatistical Modeling

Patient Data Analysis

Modeling is performed

using patient level data

Population Health Risk Assessment Modeling

Analysis of data using various statistical and machine learning techniques

helped

identify patients at high

risk of ESRD progression

Low

Medium

High

Very high

Need of intervention

Slide26

Questions?

Slide27

Presenter Contact Information

Raj Lakhanpal, MDCEO, SpectraMedix609-336-7733, x301 (office)609-865-3244 (cell)

Raj.Lakhanpal@SpectraMedixcom

Indranil

GangulyVice President & CIO JFK Health732-321-7702IGanguly@JFKHealth.org