Event Fields: Learning Latent Event Structure for Waveform Foundation Models

Researchers have introduced a novel class of waveform foundation models designed to improve the analysis of physiological time series data. Unlike conventional sequence-based approaches that treat signals as collections of local tokens, this new method models data as realizations of latent event processes. The framework assumes that clinically significant structures emerge from temporally extended, interacting events with unobserved boundaries. To capture this, the team developed a self-supervised learning system that ensures consistency across stochastic segmentations and time-frequency projections. This approach creates representations invariant to signal-level perturbations while preserving event-level organization. The model features a segmentation-aware encoder and a latent interaction operator to map dependencies among inferred events, allowing for natural extension to multimodal settings. Benchmark tests on arrhythmia classification, hemodynamic prediction, and waveform retrieval demonstrate superior performance, robustness, and label efficiency compared to existing sequence-based baselines. These findings suggest that shifting from signal-centric to event-centric representations offers a more effective inductive bias for modeling physiological dynamics, providing a promising pathway for scaling foundation models in healthcare applications.