Quantifying surprise in clinical care: Detecting highly informative events in electronic health records with foundation models

TL;DR

This paper introduces a foundation model-based method to identify highly informative events in electronic health records, improving anomaly detection and interpretability for patient outcome prediction.

Contribution

It presents a novel approach leveraging foundation models to quantify event informativeness in EHRs, enhancing anomaly detection and model interpretability.

Findings

Model effectively flags significant anomalous events.

Informativeness helps improve downstream outcome predictions.

Dropping low-informative events maintains prediction accuracy.

Abstract

We present a foundation model-derived method to identify highly informative tokens and events in electronic health records. Our approach considers incoming data in the entire context of a patient's hospitalization and so can flag anomalous events that rule-based approaches would consider within a normal range. We demonstrate that the events our model flags are significant for predicting downstream patient outcomes and that a fraction of events identified as carrying little information can safely be dropped. Additionally, we show how informativeness can help interpret the predictions of prognostic models trained on foundation model-derived representations.