Google Debuts Biomarker Discovery Framework for Wearables

Google Research, in collaboration with MIT, has developed the Biomarker Discovery Framework. This multi-agent AI system automates the pipeline of translating continuous physiological data from consumer wearables into statistically sound clinical biomarkers. Led by intern Yubin Kim alongside Google researchers Hamid Palangi and Daniel McDuff, the framework coordinates specialized agents—such as Scout, Critic, Defender, and Mechanism—under an Orchestrator agent to generate hypotheses, run statistical tests, and write reports under human supervision.

To test the system, researchers deployed it across three cohorts totaling 9,279 participant-observations spanning metabolic disease and mental health. The framework successfully flagged 41 candidate biomarkers for mental health and 25 for metabolic outcomes. For instance, in the DWB depression dataset, it linked sleep-duration variability to PHQ-8 severity (ρ = 0.252, p < 0.001). In the GLOBEM cohort, it associated sleep-onset variability with PHQ-4 severity (ρ = 0.126, p < 0.001; CV AUC = 0.535). It also engineered a cardiovascular fitness index—steps divided by resting heart rate—to correlate with insulin resistance. Adding these AI-derived features to demographic data boosted predictive performance, yielding a ΔR² of 0.040 for depression and 0.021 for insulin resistance.

To ensure statistical rigor, the framework subjects candidates to an 11-test adversarial filtering battery to prevent target leakage and overfitting. In a blinded evaluation by 15 medical and machine learning experts, the system outperformed Google DeepMind’s AI co-scientist, Biomni, and Google ADK’s Data Science Agent. It was the sole system to receive Accept or Minor Revision editorial decisions, earning two accepts, eight minor revisions, eight major revisions, and three rejections. Experts estimated they would retain 56.9% of the framework's generated report content, far exceeding the 18.8% to 30.4% retention rates of baseline models.

For clinical researchers, this framework shifts AI from a black-box predictor to a transparent assistant. By combining deterministic code execution for statistical math with generative reasoning for literature-grounded hypotheses, the tool helps practitioners safely accelerate the discovery of new digital biomarkers while maintaining strict scientific oversight.