Physiological foundation modeling for subclinical disease assessment: a prospective pilot

Clinical studies struggle to locate the right patients, in part because many remain undiagnosed or lack relevant labels. We develop and prospectively test a physiology-based patient representation ("Bioprofile") to evaluate whether it can reduce the number of individuals who must be screened and enable more precise targeting of candidates for steatotic liver disease studies.

We trained Bioprofile patient representations from routinely collected health data and fine-tuned against multiple endpoints. Bioprofiles were trained using 1 million subjects from the UK Biobank and other cohorts. The trained model was applied to 45 484 research subjects at Vanderbilt University Medical Center. Based on Bioprofile nominations, 31 subjects were recruited for a prospective validation study. The primary outcome measure was proton density fat fraction (PDFF), an imaging-based metric of steatosis severity.

Bioprofile models achieved a Spearman coefficient of 0.65 against PDFF, outperforming existing foundation models (ρ = 0.361) and clinical risk scores (ρ = 0.522-0.542) in retrospective validation. Simulations found that Bioprofiles required half as many subjects needed to screen compared to existing methods depending on task. Bioprofiles were further validated against 5 global cohort studies. Bioprofile predictions aligned strongly with prospective study data (ρ = 0.740).

Artificial intelligence (AI)-based profiling of patient physiology can reveal individuals who have subclinical signatures of disease. If implemented widely, this approach can identify unknown human subjects, reduce screening failures, and improve trial quality. Bioprofiles may have additional utility in decision support applications in precision medicine and AI-augmented healthcare.