An AI tool for prioritizing candidate biomarkers from wearable sensor data

Google Research says the hard part of wearable medicine is no longer collecting the data. The hard part is turning endless streams from devices into biomarkers that are actually worth a clinician’s time.

Its answer is the Biomarker Discovery Framework, a multi-agent system built to run a research loop under human supervision. One agent breaks down the task, others pull literature, generate hypotheses, run statistical tests, train models, and then try to tear the candidates apart for leakage, overfitting, confounding, instability, and other ways time-series work can lie to you.

That matters because these systems usually chase predictive accuracy and stop there. On physiological data, that can mean spurious correlations dressed up as insight. Google’s setup tries to keep deterministic analysis separate from generative reasoning, with shared memory and a fact sheet so the workflow stays traceable and the numbers do not drift from the evidence.

The team ran it across three cohorts with 9,279 participant-observations, covering mental health and metabolic disease. It recovered known clinical signals, found converging biomarkers across independent datasets, and identified 41 candidate digital biomarkers for mental health plus 25 for metabolic outcomes. In one example, it linked sleep-duration variability to PHQ-8 severity in DWB with Spearman’s ρ of 0.252. In GLOBEM, sleep-onset variability showed up as a weaker exploratory signal for PHQ-4, with ρ of 0.126 and a CV AUC of 0.535.

The system also built new composite features instead of just selecting whatever was already in the dataset. One example was a cardiovascular fitness index made from steps divided by resting heart rate, which it tied to insulin resistance in the metabolic cohort. Google says the derived features improved prediction when added to demographic variables, with ΔR² of 0.040 for depression and 0.021 for insulin resistance.

There was also a human test. Fifteen experts reviewed blinded reports from this system and three other AI research tools. The Biomarker Discovery Framework got the highest mean scores across all seven quality dimensions, and it was the only one to earn any Accept or Minor Revision recommendations: 2 Accept, 8 Minor Revision, 8 Major Revision, and 3 Reject. The reviewers said they would keep 56.9% of its manuscript content on average. Not bad for a machine that still needs a grown-up in the room.