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

A research team has introduced BCG-FM, a foundation model designed to monitor cardiac health passively through a bed-embedded piezoelectric sensor, according to a preprint posted to the arXiv repository [1]. The model was pre-trained on 2.75 million hours of nightly recordings from 145,985 individuals, making it the largest raw-waveform biosignal pretraining corpus reported to date [1][2]. The system captures ballistocardiography (BCG), a mechanical signal produced by the heart's ejection of blood, without requiring users to wear a device or visit a sleep laboratory [1]. The authors state that BCG-FM is the first foundation model for ambient mechanical biosignals, a departure from existing models that depend on wearables or clinical polysomnography [1][2]. The preprint was submitted on 5 June 2026 to arXiv, an open-access repository for electronic preprints that has hosted more than two million articles since its launch in 1991 and currently receives approximately 24,000 submissions per month [6]. Frozen embeddings from the model achieved a mean absolute error of 3.26 years on a biological-age estimation task, which the researchers describe as the lowest reported for any ambient, contactless modality [1][2]. The embeddings also provided clinically relevant discrimination across 15 self-reported health conditions and in three independent external cohorts [1][2]. In a data-efficiency experiment, representations derived from only 500 labeled participants outperformed a fully supervised baseline trained on 3,372 participants [1][2]. The paper further notes that representation quality scales log-linearly with contrastive batch size [1][2]. The work appears on arXiv through the standard preprint moderation process, which does not constitute peer review [6]. The abstract page includes integrations from arXivLabs, a framework launched in 2020 that allows community collaborators to build experimental tools on top of arXiv's infrastructure [5]. These tools, such as the Bibliographic Explorer and CORE Recommender, are designed to help readers discover related research and navigate citation networks [4][5]. arXiv has stated that third-party collaborators operating under the Labs framework receive only minimal, anonymized user data and are prohibited from using it for purposes beyond the functioning of the feature [5].