Event
Suchi Saria - PhD, Johns Hopkins University
Tuesday, February 7, 2017 15:30to16:30
Purvis Hall
Room 24, 1020 avenue des Pins Ouest, Montreal, QC, H3A 1A2, CA
Scalable Joint Models for Reliable Event Prediction: Application to Monitoring Adverse Events using Electronic Health Record Data.
Many life-threatening adverse events such as sepsis and cardiac arrest are treatable if detected early. Towards this, one can leverage the vast number of longitudinal signals---e.g., repeated heart rate, respiratory rate, blood cell counts, creatinine measurements---that are already recorded by clinicians to track an individual's health. Motivated by this problem, we propose a reliable event prediction framework comprising two key innovations. First, we extend existing state-of-the-art in joint-modeling to tackle settings with large-scale, (potentially) correlated, high-dimensional multivariate longitudinal data. For this, we propose a flexible Bayesian nonparametric joint model along with scalable stochastic variational inference techniques for estimation. Second, we use a decision-theoretic approach to derive an optimal detector that trades-off the cost of delaying correct adverse-event detections against making incorrect assessments. On a challenging clinical dataset on patients admitted to an Intensive Care Unit, we see significant gains in early event-detection performance over state-of-the-art techniques.