CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation
CardioState-JEPA learns one cardiac representation across ECG, PPG, and PCG, using delay alignment to compare signals at the same point in the cardiac cycle.
The paper says the model puts ECG, PPG, and PCG waveforms into a shared token space and trains a single Transformer encoder on masked latent cardiac states. Its delay aligner is meant to handle the offset between electrical, mechanical, and blood-flow signals. The authors pretrain first on unimodal data, then use paired recordings to align the modalities. As a frozen encoder, it reports gains across 25 downstream tasks, including PPG classification, PCG murmur detection, and ECG classification. HF Daily Papers' note
The paper says the model puts ECG, PPG, and PCG waveforms into a shared token space and trains a single Transformer encoder on masked latent cardiac states. Its delay aligner is meant to handle the offset between electrical, mechanical, and blood-flow signals. The authors pretrain first on unimodal data, then use paired recordings to align the modalities. As a frozen encoder, it reports gains across 25 downstream tasks, including PPG classification, PCG murmur detection, and ECG classification. HF Daily Papers' note
score 4