|Friday, February 21|
|CS11 Risk Prediction & Modeling||
Fri, Feb 21, 1:30 PM - 3:00 PM
Predicting Acute Cardiac Events using Electronic Health Records (302709)*Benjamin A. Goldstein, Stanford University - Quantative Sciences Unit
Keywords: Electronic Health Records, Risk Prediction, Machine Learning, Risk Assessment
Individuals undergoing hemodialysis (HD) are at increased risk of cardiac events such as myocardial infarction, stroke and sudden cardiac arrest. Electronic health records (EHRs) capture detailed and changing information about patients' health status. An HD session represents an ideal environment to utilize EHRs - patients typically receive treatments 3 times a week where hemodynamics, medications and blood laboratory values are collected. Using this detailed data in a large cohort of patients undergoing HD (100,000 people over 22 million sessions) I will illustrate a diverse array of methods for assessing the risk of a cardiac events: (1) functional spline regression will be used to identify biomarkers for cardiac events; (2) Machine learning methods (e.g. Random Forests, Lasso) to derive a predictor and assess how far out an event can be forecast; (3) Ensemble methods to combine predictors. Different evaluation metrics (c-statistics, IDI, PPV) are applied based on the particular question of interest. The talk highlights both the potential and challenges for predicting a rare event with a dense set of data.