Event Fields: Learning Latent Event Structure for Waveform Foundation Models

Researchers introduce a novel waveform foundation model that represents physiological signals as latent event processes rather than sequential tokens, using self-supervised learning to capture clinically meaningful structure. The approach demonstrates improved performance on medical benchmarks including arrhythmia classification and hemodynamic prediction, suggesting event-centric representations may be more suitable for healthcare AI than traditional sequence-based methods.

This research presents a fundamental shift in how foundation models process physiological waveforms, moving away from sequence-based tokenization toward event-driven representations. The distinction matters because clinical signals contain temporally extended patterns with dependencies that don't map cleanly to local tokens—treating heartbeats or respiratory cycles as discrete events better reflects how clinicians interpret these signals. The self-supervised framework enforces consistency across stochastic segmentations and time-frequency projections, creating representations robust to signal perturbations while preserving event-level structure. This inductive bias proves particularly valuable in healthcare applications where label scarcity limits traditional supervised learning.

The work builds on growing recognition that general-purpose foundation models may not capture domain-specific structure efficiently. Healthcare time series differ fundamentally from natural language or vision—clinical events have variable duration, irregular timing, and complex interdependencies that sequence models handle poorly. By incorporating a segmentation-aware encoder and latent interaction operator, the model naturally accommodates multimodal physiological data through shared event representations, addressing a persistent challenge in clinical AI systems.

The empirical results across multiple benchmarks suggest measurable advantages in performance, robustness, and label efficiency compared to sequence baselines. For healthcare AI developers and institutions, this represents a viable alternative architecture for building more sample-efficient models—critical given the regulatory constraints and data limitations in medical applications. The approach could influence how foundation models are adapted for specialized domains beyond healthcare, wherever signals exhibit event-like structure.

• →Event-centric representations capture physiological dynamics more effectively than token-based sequence models by modeling clinically meaningful temporal structures.
• →Self-supervised learning across stochastic segmentations and time-frequency projections creates representations robust to signal perturbations while preserving event organization.
• →The model demonstrates improved performance on arrhythmia classification, hemodynamic prediction, and waveform retrieval with better label efficiency than baselines.
• →Natural multimodal extension through shared event representations addresses a significant challenge in clinical AI integration across multiple signal types.
• →This architectural shift suggests a complementary scaling path for healthcare foundation models distinct from general-purpose AI approaches.