Physiological foundation modeling for subclinical disease assessment: a prospective pilot

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Clinical Trial Recruitment

Background:

• Clinical trial recruitment faces challenges due to undiagnosed patients and lack of specific labels.
• Efficient patient identification is crucial for the success and quality of clinical studies, particularly for conditions like steatotic liver disease.

Purpose of the Study:

• To develop and prospectively test a physiology-based patient representation, termed "Bioprofile," for improved patient identification in clinical studies.
• To evaluate if Bioprofiles can reduce screening numbers and enable more precise targeting of candidates for steatotic liver disease research.

Main Methods:

• Trained Bioprofile models using routinely collected health data from over 1 million subjects.
• Fine-tuned models against multiple endpoints and applied them to a cohort of 45,484 research subjects.
• Validated Bioprofile predictions against prospective study data and compared performance with existing foundation models and clinical risk scores.

Main Results:

• Bioprofile models achieved a Spearman coefficient of 0.65 against proton density fat fraction (PDFF), outperforming existing methods (ρ=0.361-0.542).
• Simulations indicated Bioprofiles could halve the number of subjects needing screening compared to current approaches.
• Prospective validation showed strong alignment between Bioprofile predictions and study data (ρ=0.740).

Conclusions:

• AI-based Bioprofiling can identify individuals with subclinical disease signatures, improving clinical trial quality.
• Widespread implementation could reduce screening failures and identify previously unknown subjects with specific conditions.
• Bioprofiles show potential for decision support in precision medicine and AI-augmented healthcare.