using rwd in the form of ehr data to predict adverse ...• 1.04% of study population had a bleed...

18
Using RWD in the Form of EHR Data to Predict Adverse Events and Outcomes PHUSE US Connect 2020 Monday March 9 th , 2020

Upload: others

Post on 01-May-2021

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Using RWD in the Form of EHR Data to Predict Adverse Events and OutcomesPHUSE US Connect 2020Monday March 9th, 2020

Page 2: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Adverse Drug Events: Huge costs to Patients and Providers Alike

1 in 6senior hospital admissions is due to an ADE

Annually,

1 million patient admissions nationally with an estimated cost of more than

$14 billion

More than 90% of 30-day readmissions due to ADEs are preventable.1

2

Research Need

1. Dalleur, O., et al (2017). 30-Day Potentially Avoidable Readmissions Due to Adverse Drug Events. Journal of Patient Safety, [Online, published ahead of print]. Available from: https://journals.lww.com/journalpatientsafety/

Page 3: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

• Patient-level risk scoring can improve patient outcomes and prevent costs to the health system

• Limited scale of prior predictive modeling of ADE risk for anticoagulant users

• Edj’s machine learning modeling excels at decision support and resource allocation, making it a natural fit for this problem

Early Risk Detection Can HelpResearch Impact

3

Page 4: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Objective

Dataset

Adverse Event PredictionObjectives and Data

4

• Increase patient safety by predicting serious bleeding event risk and risk factors for individual patients undergoing anticoagulant therapy

• Evaluate risk on a 90-day horizon • Drive reduction in readmission rate with risk factors identified by machine

learning

• 39,058 patients on anticoagulants from a health system touching about two million lives

– Time span: 2017-2018• 1.04% of study population had a bleed event in 3-month window• Received first-party EHR data and integrated proprietary SDoH factors

Page 5: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Results

Key Findings

Adverse Event PredictionResults and Key Findings Snapshot

5

• Edj queried and compiled diagnoses for all encounters (inpatient, ER, outpatient, ambulatory) for serious bleeding events.

• Edj built a highly predictive model well-suited for resource allocation– Model’s highest ADE risk group has an event 10x more often than the total

population

• Edj’s model produces risk and anti-risk factors at both a population level and an individual level

• Edj generated significant predictive lift (8%) by combining proprietary SDoH data with health system EHR data.

Page 6: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

• Describes real events in a health system

• First-hand data from service providers

• Modeling fit to the local population

• No third-party data costs

Why Model with EHR Data?Benefits of EHR Data as RWD

6

• Event-level data source with clear time component

– Can be updated near real-time

• Patient-level data allows for predictions by individual

• Easier to integrate back into the EHR for provider decision support

In-House Data Granularity Integration

Page 7: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

To model with EHR data, we:• Perform extensive ETL/Data Processing

– Hundreds of person-hours of data exploration and cleaning to prepare for modeling stage

• Creatively engineer predictive features– Created more than 1000 features for ADE model– Example: Calculating change in a patient’s vitals

over various time periods• Added a temporal component to static model types

Preparing EHR Data for InsightsData Cleaning and Feature Engineering

7

Page 8: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Modeling ResultsUsing RWD in the Form of EHR Data to Predict Adverse Events and Outcomes

Page 9: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Instant Risk Detection and ContextModeling Results

9

Highly Accurate and InterpretableRisk Scores for Each Patient

• Predicts risk of a bleed event within a three-month window based on EHR and non-clinical factors

• Promotes optimizing care continuum resource allocation by risk score

Population-level and Patient-level Risk and Anti-Risk Factors

• Details “the why” for a risk score, making it truly a decision support tool

• Individual-level risk factors identified, driving personalized risk-prevention plans

Page 10: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Features

10

• CPT and ICD codes– Indicator vectors for high occurring individual codes– Rolled up versions of individual codes– Bag of words representations of code descriptions

• Medications– Indicator vectors of generic name and pharmacy class

• Labs and vitals– Summary stats of labs and vitals over past two years, as well as rate of change over time.

• Demographics– Indicator vectors for age, sex, smoking status, etc.

• Various encounter information– Frequency of visits, service locations, discharge disposition, insurance type

Modeling Results

Page 11: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

• AUC (Area Under the Curve) is a score for how accurate a model’s predictions are in aggregate.

- It ranges from 0 to 1- Also referred to as a C-

Statistic• The random forest model, using

EHR data coupled with SDoH data, provides strong predictive accuracy within the three-month prediction window.

- Area under curve metric (AUC = .8) indicates that the model distinguishes effectively between patients with bleed event and no bleed event.

11

Evaluating Predictive PowerModeling Results

Page 12: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

• There is a low target event rate with an overall bleed rate of .0104.

• As your risk score increases, the actual bleed rate for patients scored in that strata is increasing.

• The scores from the model output are conducive to allocating resources to higher risk patients.

Model PerformanceModeling Results

12

Page 13: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

• At a high level, feature importance tells how much of a given prediction is coming from that feature.

– A feature being important doesn’t imply a positive or negative linear relationship.

– These features are important in a series of rules rather than being important in isolation.

Key Features of ImportanceModeling Results

13

• Key features of importance in determining an individual’s risk score include:– Procedure descriptions, diagnosis descriptions– Medications– Vitals and general health– Engagement with the health system

Page 14: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Sample Patient-Level Risk FactorsModeling Results

14

Page 15: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Why Do Social Determinants Matter?

• A 2016 study determined that only 16% of a health outcome could be explained by clinical care.1

• The remainder of the health outcome could be explained by social determinants of health (SDoH)

– Physical environment – Healthy behaviors – Socioeconomic conditions

• To complete the picture of the patient, Edj integrates proprietary SDoH data into clinical modeling

Modeling Results

151. Hood, C. M., Gennuso, K. P., Swain, G. R., & Catlin, B. B. (2016). County Health Rankings:

Relationships Between Determinant Factors and Health Outcomes. American Journal of Preventive Medicine, 50(2), 129–135. doi: https://doi.org/10.1016/j.amepre.2015.08.024

Page 16: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Key SDoH factors for this model included:• Percent of never delinquent transactions (i.e. bills paid)

– Possibly an indicator for drug adherence• Life insurance score

– Credit-based measure of mortality risk• Credit score

SDoH Features Drive Better PredictionsModeling Results

16

8 percent increase in predictive power

Page 17: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Q&AUsing RWD in the Form of EHR Data to Predict Adverse Events and Outcomes

Page 18: Using RWD in the Form of EHR Data to Predict Adverse ...• 1.04% of study population had a bleed event in 3-month window • Received first-party EHR data and integrated proprietary

Contact Us!Susan Olson, [email protected]+1 574-210-6592

Corey YoungSr. Data [email protected]