Tech and Risk Adjustment Using Data Analytics and Machine ...

47
Tech and Risk Adjustment: Using Data Analytics and Machine Learning to improve Risk Adjustment Management NPA Conference 10/25/2016

Transcript of Tech and Risk Adjustment Using Data Analytics and Machine ...

Tech and Risk Adjustment: Using Data Analytics and

Machine Learning to improve Risk Adjustment Management

NPA Conference 10/25/2016

On Lok Risk Adjustment Overview

▪Challenges▪Problems Attempting to Solve▪Tech Solutions

2

Agenda

•Sam – Technology and Reports•Kat – Reports Workflow•Travis – Machine Learning •Kat – Coder Workflows, Findings, and Provider Education

3

Tech Solutions Focus Areas

2 Main Focus Areas

•Free Text in Provider Documentation – Machine Learning and Chart Review Tool

•Missed or Incomplete RAPS Documentation in EHR - Operational Reports

Tech Solutions:Tech Solution

4

Machine Learning

Chart Review Tool

Operational Reports

Outcomes

Corrected Charts

Machine Learning

Chart Review

Operational Reports

Comprehensive HCC submissions

Improved Patient Documentation

Provider Education

Machine Learning – [RALE]

Risk Adjustment Learning Engine System•Machine Learning Engine•Provider Queues•Addendum Creation•Monitoring and Metrics•Generates Dx for RAPS

6

Chart Reviews – [Doc. Coder]

Tool to Assist with Chart Reviews•Advanced EHR Searching•RALE Integration•Provider Queues•Addendum Creation•Generates Dx for RAPS•Streamlines workflow

7

Operational HCC Reports

Missed Reporting

Opportunities< 2 HCC’s

Missing HCC after Cap

Incomplete Documentation

Missed Submission Opportunity

Dropped HCCProblems

recorded but not Assessed

Missed coding opportunity

Machine Learning

Chart Reviews/Doc

Coder

Monthly RAPS Processing Cycle

9

HCC Reports/Analysis

Extract/Submit EHR, Claims,

RALE, and CR

Chart updates

Machine Learning, Chart

Reviews

Next Up

•Kat – Reports Workflow•Travis – Machine Learning •Kat – Workflows, Findings, and Provider Education

10

Reports Work Flow

•Missing E&M Code: automatically distributed weekly to coder, who reviews SOAP note and determines E&M code•Missing Master IM: automatically distributed weekly to provider, who completes SOAP template and generates note•>2 HCCs, Dropped HCC, and Missing HCC: distributed periodically to provider, who determines whether or not patient needs to be brought in for assessment

Machine Learning

Missed Reporting

Opportunities0 or 1 HCC

Missing HCC after Cap

Incomplete Documentation

Missed Submission Opportunity

Dropped HCCProblems

recorded but not Assessed

Missed coding opportunity

Machine Learning

Chart Reviews/Doc

Coder

Machine Learning

Predicting Hierarchical Condition Categories

What is machine learning?

Making computers understand the meaning of observations

Categorized Observations Machine Learning

Algorithm Model

Where is Machine Learning?

● Your messaging app● The search bar● The Tesla● Your travel app● Your broker● And a billion other things

How does On Lok Apply Machine Learning?

● Observations = Chart Notes● Each chart note has a related diagnosis● Each diagnosis falls into an HCC● Collection of categorized chart notes = training set● Training set + machine learning = model● Model + new chart notes = prediction

On Lok example

HCC19 - Diabetes without complication

● Find chart notes w/diagnosis codes mapping to HCC19● This is the training set● The machine learning engine extracts the best features

from the documents. For HCC19 they could include: thirsty, hunger, losing weight, tired, etc

● The model is built and a subset of the training set is used to score the model’s accuracy

● We now have a model that we can run any new chart notes against

Making PredictionsWe have models, now what?

A new chart note is created

Run the note against the

model

Generate Predictions

Remove if HCC(s) already

assigned

Is this a high quality

prediction?

Add to the queue for the

provider

Provider queue review

Prediction is correct

Addendum added to chart

Results fed back into the

modelHCC submitted

Prediction Queue

Prediction Queue

Prediction Queue

Too good to be true?

Let’s look at the challenges with machine learning

Rocket Science

Data Dilemmas

We Still Need Humans

Models need attention

Rocket Science

There are a lot of moving parts, big concepts and math

Rocket Science

We use a collaboration of:

Data Scientists

Software Engineers

Database Administrators

Coders

Physicians

Rocket Science

Gather the data

Organize the data

Clean the data

Test the data

Build the models

Verify the models

Predict on new documents

Present the predictions

Create Addendums

Analyze the outcomes

We have applications to…

Data Dilemmas

QuantityToo few examples

QualityBad examples

QuantitySiamese Burmese

So what is this?

Quantity

Chart Notes

● On Lok has a small number of patients which generates a relatively small number of documents

● Machine learning is learn by example, a shallow training set will yield shallow learning

● Patient data is protected and therefore hard to come by● We are always looking for ways to find more categorized

data.

QualitySiamese Burmese

So what is this?

Quality

Our data wasn’t gathered with machine learning in mind

● Chart notes sometimes contain nothing that indicates the diagnosis attached to them

● When reviewing the predictions the predicted HCC category may match the participant but not the document

We Still Need Humans

● Because predictions are never 100% accurate, we need a human to verify the results (the queue mentioned earlier)

● If the predictions are < 50% accuracy it means there are more misses than hits. That can get very frustrating for a busy physician

● Frustrated physicians lead to lower participation and less accurate results

Models Need Attention

● Unless you have millions of really high quality documents, you will have to frequently rebuild and tune your models.

● Every time you get a pile of high quality results you will want to add them back into your model

● Models can take a long time to build● We frequently Check our models against actual results

What’s Next?

● Integration wherever documents are used○ EHR integration○ etc.

● Faster on-demand predictions○ Composite Model○ GPU processing

● Accuracy Improvements○ Composite Model○ Expanded Datasets○ Integration of Google’s NLP tools

RALE Workflow

RALE predictions run

Providers validate

predictions

Dx selected and addendum

created

RAPS file created

Dx and addendum added

to HER

Matches and No Matches fed back

into RALE

Missed Reporting

Opportunities0 or 1 HCC

Missing HCC after Cap

Incomplete Documentation

Missed Submission Opportunity

Dropped HCCProblems

recorded but not Assessed

Missed coding opportunity

Machine Learning

Chart Reviews/Doc

Coder

Doc Coder

Document Coder Workflow - K

Recommendations entered into Doc

Coder.Providers validate

suggestionsSelect dx and

addendum created.

RAPS file created.Dx and addendum added to EHR.

Matches and No Matches are fed back

into RALE.

Doc Coder – Results - K

38

Findings - Unique HCC’s Submitted for 2014 DOS

•As of 1/29/2016•Reports: 229 (4.9%)•Machine Learning: 336 (7.4%)•Doc Coder: 220 (4.8%)•Improvement in HCC risk scores 2014/2015 Dates of Service:• Base: +21% (provider education)• Tools: +17% (Rpts, Mach Lrn, Doc Coder)

39

Top Three Machine Learning Findings

1. Vascular Disease (21%)•Specifically, atherosclerosis of the aorta

2. Major Depressive, Bipolar, and Paranoid Disorders (18%)

•Specifically, major depression3. Dementia without Complication

(13%)

40

Top Three Doc Coder Findings

1. Major Depressive, Bipolar, and Paranoid Disorders (18%)

•Specifically, Major Depression2. Polyneuropathy (14%)3. Vascular Disease (12%)

41

Education of Providers(Kat)

•Diagnostic sources• Review all labs, specialist referrals, etc. on re-eval

• Example: atherosclerosis of aorta on chest x-rays

•Diagnostic guidelines• Review medical journals for diagnostic requirements

• Example: depression vs. major depressive disorder

•Coding Guidelines• Review coding resources for coding requirements

• Example: dementia with vs. without complications

42

RAPS Processing Overview

44

Run Tools and Reports

Validate and update Charts

Extract Claims, EHR, Machine Learning, Doc

Coder

Submit to RAPS

Process Errors

Missed documentationMissed face to face encounters

Add E&M code and Dx

Extract face to face encounters

RAPS return filesstored in EDW

Run monthly and during sweeps

Tools/Technology (Sam)

•Analytics•Reports•Machine Learning•Productivity Tools

•Doc Coder

45

Miss HCC

0 or 1 HCC

Missing HCC after Cap

Incomplete Documentation

Missed Submission Opportunity

Dropped HCCProblems

recorded but not Assessed

Missed coding opportunity

RALE

Chart Reviews/Doc

Coder

HCC’s

HCC Count103 2011108 3791111 4452122 67134 462136 1704138 5539139 4623169 8218 956228 35548 59451 411752 979255 24758 330575 60578 175579 81484 823388 18519 45996 5377 Total 69,797

46

About Doc Coder – PLACE Holder

•Assists with Manual Chart Reviews•Advanced Search Functionality•Knowledge of RALE Findings •Builds PCP Review Queue •Automated Generation of Addenda•Generates DX Code Files for RAPS

47