Physiological foundation modeling for subclinical disease assessment: a prospective pilot

Clinical studies across disease areas struggle to locate the right patients, in part because many remain undiagnosed or lack relevant subclinical disease labels. We created an AI framework that converts routine health-record data into physiology-based “Bioprofiles,” enabling faster, more precise identification of candidates for prospective research. We aimed to evaluate whether Bioprofile-guided recruitment could reduce the number of individuals who must be screened and enable more precise, physiology-informed targeting of candidates for steatotic liver disease studies. We trained Bioprofile patient representations from routinely collected health data including demographics, medical history, lab results, and lifestyle factors, and fine-tuned against multiple endpoints including MASLD severity, evaluated using Proton density fat fraction (PDFF), an imaging-based metric of steatosis severity. We trained Bioprofiles using 1 million subjects from the UK Biobank and other cohorts and recruited 31 subjects from Vanderbilt University Medical Center for study visits based on Bioprofile nominations. Bioprofile models achieved a Spearman coefficient of 0.646 against PDFF, strongly outperforming existing foundation models (r = 0.361) and clinical risk scores (r = 0.516-0.545). Simulations of the Bioprofile patient nomination pipeline suggest it requires half as many subjects needed to screen compared to existing methods depending on task. These findings were validated in a prospective pilot at VUMC, where 31 Bioprofile-nominated subjects were recruited to undergo liver imaging. Bioprofile predictions aligned strongly with study data (r = 0.740). Bioprofiles were further validated against 5 global cohort studies for non-PDFF MASLD endpoints. AI-based profiling of patient physiology can surface individuals who harbor subclinical or otherwise unrecognized disease signatures. This approach can facilitate subject nomination for human studies, reduce screening failures, and improve trial quality. Bioprofiles may have additional utility in decision support applications in precision medicine and AI-augmented healthcare.