CLaI: Collaborative Learning and Inference for Low-Resolution Physiological Signals: Validation in Clinical Event Detection and Prediction

While machine learning (ML) techniques have been applied to detection and prediction tasks in clinical data, most methods rely on high-resolution data, which is not routinely available in most Intensive Care Units (ICUs), and perform poorly when faced with class imbalance. Here, we introduce and validate Collaborative Learning and Inference (CLaI) for detection and prediction of events from learned latent representations of multivariate physiological time series, leveraging similarities across patients. Our method offers a new way to detect and predict events using low-resolution physiological time series. We evaluate its performance on predicting intracranial hypertension and sepsis using the KidsBrainIT (minute-by-minute resolution) and MIMIC-IV (hourly resolution) datasets, respectively, comparing our approach with classification-based and sequence-to-sequence benchmarks from existing studies. Additional experiments on sepsis detection, robustness to class imbalance, and generalizability-demonstrated via seizure detection using the CHB-MIT scalp electroencephalogram dataset-confirm that CLaI effectively handles class imbalance, consistently achieving competitive performance and the highest F1 score. Overall, our approach introduces a novel method for analyzing routinely collected ICU physiological time series by leveraging patient similarity thus enabling ML interpretability through case-based reasoning.