An AI tool for prioritizing candidate biomarkers from wearable sensor data

Wearable devices continuously capture physiological signals like heart rate and sleep patterns that could reveal early health changes, but converting these data streams into reliable biomarkers has been difficult. Researchers introduced the Biomarker Discovery Framework, a multi-agent AI system that automates biomarker discovery through structured hypothesis generation, statistical analysis, and literature review while maintaining human oversight and preventing common errors like spurious correlations and data leakage. The system was tested across three large datasets totaling over 9,000 participant-observations in mental health and metabolic disease domains, where it identified candidate biomarkers including sleep variability linked to depression severity and a cardiovascular fitness index associated with insulin resistance. The framework combines computational rigor for numerical analysis with AI reasoning for hypothesis formation, using specialized agents that check each candidate through an 11-point validation process before presenting results for expert review.