Beyond Patient Invariance: Learning Cardiac Dynamics via Action-Conditioned JEPAs

Abstract: Self-supervised learning in healthcare has largely relied on invariance-based objectives, which maximize similarity between different views of the same patient. While effective for static anatomy, this paradigm is fundamentally misaligned with clinical diagnosis, as it mathematically compels the model to suppress the transient pathological changes it is intended to detect. We propose a shift towards Action-Conditioned World Models that learn to simulate the dynamics of disease progression, or Event-Conditioned. Adapting the LeJEPA framework to physiological time-series, we define pathology not as a static label, but as a transition vector acting on a patient's latent state. By predicting the future electrophysiological state of the heart given a disease onset, our model explicitly disentangles stable anatomical features from dynamic pathological forces. Evaluated on the MIMIC-IV-ECG dataset, our approach outperforms fully supervised baselines on the critical triage task. Crucially, we demonstrate superior sample efficiency: in low-resource regimes, our world model outperforms supervised learning by over 0.05 AUROC. These results suggest that modeling biological dynamics provides a dense supervision signal that is far more robust than static classification. Source code is available at https://github.com/cljosegfer/lesaude-dynamics

• The paper introduces an action-conditioned JEPA framework that models cardiac disease onset as latent transition vectors.
• It employs an xResNet1d50 backbone with isotropic Gaussian regularization to create robust and separable latent representations.
• Experimental results on the MIMIC-IV ECG dataset show improved AUROC and sample efficiency over traditional SSL and supervised methods.

Action-Conditioned Latent Dynamics for Cardiac Monitoring: Overview

"Beyond Patient Invariance: Learning Cardiac Dynamics via Action-Conditioned JEPAs" (2604.22618) addresses a central misalignment in self-supervised learning (SSL) for medical time series, specifically cardiac monitoring with 12-lead ECGs. Whereas most prevailing SSL methods enforce invariance to patient identity—potentially collapsing transient pathological changes essential for diagnosis—this work proposes an Action-Conditioned World Model paradigm. The framework explicitly simulates disease evolution by modeling pathology as a transition vector operating on patient-specific latent representations, harnessing the Joint-Embedding Predictive Architecture (JEPA) with isotropic Gaussian regularization.

Motivation: Limits of Patient Invariance in Clinical Time Series SSL

SSL in medical AI, especially for physiological signals, has traditionally relied on contrastive and invariance-based objectives (e.g., SimCLR, PCLR). These methods are effective for capturing static anatomical features or patient re-identification but are fundamentally ill-suited to tasks where the detection of temporal change—such as disease onset—is critical.

In the clinical domain, mapping both healthy and pathological states of the same patient to proximate locations in latent space eradicates critical diagnostic signals. Furthermore, canonical data augmentations (cropping, jittering, scaling) commonly used in contrastive pipelines are often destructive for 1D signals, either removing diagnostic segments or mimicking pathologies. This creates models invariant to clinical changes and robust to artifacts but poor at capturing meaningful longitudinal dynamics.

Methodology: Action-Conditioned JEPA for Cardiac State Dynamics

The proposed model formalizes the learning of disease dynamics as a predictive latent process rather than static classification. The central innovations are:

• Action-Conditioned World Model: Disease onset is encoded as a sparse transition vector ("action") between sequential cardiac states: $a_t = y_{t+1} - y_t \in \{-1,0,1\}^C$. This action operates in latent space to predict the future patient state, enabling explicit modeling of pathological events as movements on the latent manifold.
• Latent Dynamics Architecture:
  • An xResNet1d50 backbone encodes raw ECGs to a latent embedding $h_t$.
  • The action vector is projected with an MLP projector to produce an action embedding.
  • A residual MLP predictor estimates the next latent state $h_{t+1}$ conditioned on $(h_t, \text{action embedding})$.
  • The joint embedding is regularized by Sketched Isotropic Gaussian Regularization (SIGReg), ensuring maximally entropic, rotationally symmetric, and linearly separable latent geometry.
• Disentanglement Objective: The framework forces separation between patient-invariant features (anatomical baseline) and dynamic pathological transitions (disease onset), implemented as transitions in the latent space rather than as static class labels.
• Avoidance of Generative Pitfalls: Unlike generative pixel-reconstruction world models, the Action-Conditioned JEPA predicts in abstract latent space, circumventing high computational costs and the risk of physiologically implausible generation.

Experimental Evaluation

Data and Setup: Experiments use the longitudinal MIMIC-IV-ECG dataset (~800k records, ~160k patients, rich diagnostic variety). Key design choices include patient-stratified train/test splits, action computation from truncated ICD-10 codes, and a focus on both acute and chronic cardiac state transitions.

Performance Benchmarks:

• Triage Task (First ECG of new patient): The action-conditioned dynamics model achieves superior AUROC (0.742) over both fully supervised (0.735) and naive patient-invariant SSL (0.666–0.679).
• Monitoring Task (All ECGs, temporal redundancy): Parity with supervised baselines is observed, but strong gains persist in sample efficiency.
• Low-Resource Learning: In the 10% data regime, the dynamics model outperforms supervised by over 0.05 AUROC on critical tasks. At 1% data, both approaches degrade to random, but the latent space from pretraining still supports linearly separable clinical features.

Ablation:

• SIGReg outperforms VICReg regularization, providing more stable and robust latent spaces for clinical discriminability.

Failure Analysis:

• In extremely low data, learned latent representations remain meaningful, but conventional finetuning objectives overwrite them, indicating the need for improved adaptation protocols.

Theoretical and Practical Implications

The shift from invariance to dynamics represents a significant advance for time-series SSL in healthcare. The explicit modeling of disease as a latent action enables:

• Richer supervision, capturing the causal, temporal, and entropic character of disease processes.
• Enhanced sample efficiency, critical for real-world clinical deployment where annotated data is sparse.
• Robust disentanglement, supporting both generalization to new patients and precision triage.

Limitations:

• The binary action encoding cannot distinguish between acute and chronic progression, nor model irreversible conditions’ temporal asymmetry.
• ICD code labeling is subject to delays and ambiguity relative to waveform acquisition, introducing unavoidable label noise.

Future Directions:

• Transition to continuous action embeddings representing clinical trajectories, temporal asymmetry, and morphologic similarity.
• Time-aware modeling (e.g., elapsed time between ECGs) aligning with Neural ODE frameworks for continuous disease progression simulation.
• Extensions to broader ICD ontologies, rare pathology detection, and counterfactual patient simulation.

Conclusion

This work demonstrates that action-conditioned latent dynamic models, implemented via JEPA with isotropic Gaussian regularization, provide a clinically meaningful, sample-efficient, and robust alternative to patient-invariant SSL for cardiac time-series. The framework aligns the learning objective with the true nature of clinical monitoring: detecting and anticipating change, rather than suppressing it. These innovations set a foundation for future, more powerful simulation-based clinical AI, capable of generalizing across pathologies, data regimes, and time scales (2604.22618).