BCG-FM: A Foundation Model for Ambient Cardiac Health Sensing

Researchers introduce BCG-FM, a foundation model trained on 2.75 million hours of ballistocardiography data from nearly 146,000 individuals, enabling non-invasive cardiac health monitoring through piezoelectric bed sensors. The model achieves state-of-the-art biological age estimation and demonstrates clinical relevance across multiple health conditions without requiring deliberate user action.

BCG-FM represents a significant advancement in passive health monitoring technology by leveraging ambient mechanical biosignals—specifically ballistocardiography recorded through embedded bed sensors—rather than wearable devices. This shift to zero-friction data collection removes adoption barriers that plague existing wearable-based health models, which demand consistent user engagement. The research demonstrates that foundation models pretrained on massive unlabeled datasets can achieve superior performance compared to traditionally supervised approaches, with 500 labeled participants outperforming 3,372 fully supervised training samples.

The scale of the pretraining corpus—2.75 million hours from 145,985 individuals—establishes a new benchmark for biosignal foundation models and reflects growing recognition that larger, more diverse datasets enable better transfer learning. The model's 3.26-year mean absolute error on biological age estimation surpasses previous contactless modality results, suggesting the approach has genuine clinical utility beyond research contexts.

This work bridges ambient sensing and AI-driven health analytics, opening commercial pathways for smart furniture manufacturers, healthcare providers, and remote patient monitoring platforms. Insurance companies and preventive health services could integrate such technology into homes to identify disease risks early. The log-linear scaling relationship with batch size indicates further improvements remain possible with additional compute resources.

The validation across three external cohorts strengthens claims of generalizability, though real-world deployment still requires regulatory approval and addressing privacy concerns around continuous in-home monitoring. Future development should focus on prospective validation—whether BCG-FM can predict adverse health events before clinical manifestation.

• →BCG-FM trained on 2.75 million hours of data from 145,985 participants achieves 3.26-year MAE on biological age estimation, outperforming previous contactless modalities
• →Ambient piezoelectric bed sensors enable passive health monitoring without user action or wearable device compliance burden
• →Foundation model approach requires 500 labeled samples to match performance of 3,372 fully supervised training samples, demonstrating transfer learning efficiency
• →Model validates across 15 self-reported health conditions and three independent external cohorts, suggesting broad clinical applicability
• →Technology enables new commercial opportunities in smart home health monitoring, preventive medicine, and remote patient surveillance