Foundation Models for Hemodynamic Time Series: A New Paradigm in Cardiovascular Data Modeling

Khuong Vo, Arash Shokouhmand, Arghavan Arafati, Cristhian Potes

Keywords: Foundation models, time series, blood pressure

Abstract: Hemodynamic waveforms encode rich physiological signals essential for cardiovascular assessment, but scalable interpretation has been limited by the need for labeled data and costly imaging. Leveraging approximately 34,000 hours of finger-cuff and arterial blood pressure waveforms from about 12,000 subjects, collected using Edwards Lifesciences ClearSight and FloTrac devices, we develop a transformer-based foundation model that learns robust representations of cardiovascular dynamics. Trained with self-supervised learning, the model achieves sample-efficient performance, matching state-of-the-art benchmarks while using only 30% of labeled data for detecting aortic stenosis and reduced left ventricular ejection fraction. To our knowledge, this is the first foundation model trained solely on blood pressure waveforms for cardiovascular disease screening.

Submission Number: 23