PhysioJEPA: Joint Embedding Representations of Physiological Signals for Real Time Risk Estimation in the Intensive Care Unit

Benjamin Fox, Dung T. Hoang, ANKIT PAREKH, Girish N Nadkarni, Ankit Sakhuja

Keywords: physiological signals, representation learning, foundation model, transformer, JEPA, critical care, risk estimation

TL;DR: PhysioJEPA learns representations of bedside monitoring physiological signal data via a JEPA learning procedure and frozen representations can be finetuned to estimate risk for hypotension and shock in critical care settings.

Abstract: Self-supervised learning of multichannel, high-frequency physiological signals is largely unexplored, despite their potential for critical care applications. We present PhysioJEPA, the first joint embedding predicting architecture (JEPA) designed for multichannel physiological signals from critical care bedside monitoring devices. PhysioJEPA learns representations from 30-minute segments of physiological signals from three channels: arterial blood pressure, electrocardiography lead II, and photoplethysmography, using the MIMIC-III Waveform Database. Trained on over 10.7 million minutes of arterial blood pressure, electrocardiography lead II, and photoplethysmography from 4,282 intensive care unit stays (N=2,631 patients), PhysioJEPA's learned representations can be finetuned to estimate 5-minute risk of hypotension (AUROC = 0.83 [Confidence Interval or CI 0.83-0.84]) and shock index (AUROC = 0.95 [0.95-0.96]) improving performance over a supervised convolutional baseline (AUROC = 0.78 [0.78-0.78] and 0.95 [0.95-0.95] for hypotension and shock index, respectively). Furthermore, it generalized to an external cohort from the Mount Sinai Health System (AUROC = 0.78 [0.78-0.78] and 0.92 [0.92-0.93]). These results suggest that self-supervised JEPA representation learning is a promising approach for critical care signal data.

Submission Number: 50