CardioFM: A Multimodal Foundation Model for Joint ECG and PPG Representation Learning

Electrocardiography (ECG) and photoplethysmography (PPG) arise from the same heartbeat and are routinely co-acquired at every monitored bedside, yet no foundation model jointly encodes both modalities. Existing approaches are either ECG-specific, PPG-specific, or domain-agnostic, and none captures the cross-modal physiological coupling between cardiac electrical activity and peripheral hemodynamics. We present CardioFM, a self-supervised multimodal foundation model that integrates ECG Lead-II and PPG through bidirectional cross-modal attention and adaptive residual vector quantization. CardioFM is pretrained on over 500,000 hours from approximately 63,000 patients across intensive care, surgical, ambulatory, and consumer-wearable settings, learning unified representations that transfer across contexts without retraining. CardioFM achieves an F1-score of 0.86 for cardiovascular disease classification on PTB-XL, estimates the QT interval with a mean error of 20.2 ms approaching expert inter-observer variability, and measures pulse arrival time with a mean error of 22.7 ms sufficient to support non-invasive hemodynamic trending. When used as a feature extractor, CardioFM embeddings provide superior discrimination for intensive care false alarm reduction compared with ECG-FM, PaPaGei, and TimesFM, despite requiring substantially smaller representations. In contrast, generic temporal pretraining fails to encode clinically relevant waveform morphology. Demographic inference from waveform embeddings (age MAE: 10.4 years; gender AUC: 0.97; BMI MAE: 0.66 kg/m ² ) confirms that the learned representations encode fundamental biological characteristics without requiring diagnostic labels. The model maintains zero-shot reconstruction fidelity across five independent datasets spanning heterogeneous sensor hardware, sampling rates, and patient populations, with the cross-modal attention mechanism providing robustness to single-modality signal degradation. The 17.11-million-parameter encoder is compatible with edge-deployment constraints, and the model uses only signal modalities already acquired by standard bedside monitors and consumer wearables, requiring no additional sensing hardware. These findings demonstrate that a single multimodal foundation model can consolidate the fragmented landscape of cardiac biosignal analysis, providing a unified representational framework across clinical monitoring systems and wearable health technologies that may extend to broader critical illness surveillance.