An AI Co-Data-Scientist for Prioritizing Candidate Biomarkers from Wearable Sensor Data
Yubin Kim, Salman Rahman, Samuel Schmidgall, Chunjong Park, A. Ali Heydari, Ahmed A. Metwally, Hong Yu, Xin Liu, Xuhai Xu, Yuzhe Yang, Hyeonhoon Lee, Hyewon Jeong, Kyungho Lim, MingYu Lu, Dongjae Lee, Theodora Pappa, Hanseul Cho, Maxwell A. Xu, Zhihan Zhang, Cynthia Breazeal, Tim Althoff, Petar Sirkovic, Ivor Rendulic, Annalisa Pawlosky, Nicolas Stroppa, Juraj Gottweis, Elahe Vedadi, Alan Karthikesalingam, Pushmeet Kohli, Mark Malhotra, Shwetak Patel, Samir Tulebaev, Hae Won Park, Vivek Natarajan, Hamid Palangi, Daniel McDuff
cs.AI
2026-04-16
CoDaS ranked wearable biomarker candidates across 9,279 observations; sleep-duration variability (ρ=0.252) and steps/RHR (ρ=−0.374) added ΔR²=0.040 and 0.021 over demographics.
Consumer wearables dump continuous heart rate, steps, sleep, and phone-use logs. Turning those streams into biomarker hypotheses a clinician can actually review still depends on handcrafted features, usually locked to one disease. CoDaS (AI Co-Data-Scientist), from Google Research with DeepMind, MIT and clinical collaborators, is a supervised multi-agent pipeline that generates, stress-tests, and ranks candidate digital biomarkers from wearable time series.
The hard part is leakage and spurious correlation in high-dimensional physiology. Squared glucose and a TG/HDL ratio already sitting on a lipid panel will both look like strong hits. The system is built to produce an auditable shortlist that can be rejected, not a black-box score.
Gemini-3.1 Pro Preview handles literature-heavy reasoning and code generation; Gemini-3 Flash Preview handles the high-frequency steps. An Orchestrator runs six phases: data profiling, hypothesis generation, parallel stats/ML search, adversarial validation, mechanistic and novelty assessment, and report drafting. A Scout agent maps the schema and endpoint. Deterministic runners do univariate tests and participant-level 5-fold cross-validation. Critic and Defender agents argue over leakage, construct overlap, confounding, and physiological implausibility.
Leakage controls are procedural:
A Fact Sheet of deterministically computed numbers is attached to every writing agent so sample sizes and effect sizes are copied, not invented. Humans can intervene on mechanistic interpretation and search stagnation, but they do not see raw labels or fold-level predictions during feedback.
Three cohorts, 9,279 participant-observations (9,072 people): DWB (n=7,497, PHQ-8), GLOBEM (704 waves, PHQ-4), WEAR-ME (n=1,078, HOMA-IR).
| Cohort | Lead candidate | Spearman ρ | Versus baseline |
| DWB | Main sleep-duration variability | 0.252 (95% CI 0.23–0.27) | Demographics Ridge CV R² 0.188 → 0.228 with top-5 features, ΔR²=0.040 |
| GLOBEM | Sleep-onset variability | 0.126 | All-screened-feature CV AUC=0.535, near chance; Data Science Agent 0.523 |
| WEAR-ME | Steps / resting heart rate | −0.374 (95% CI −0.42 to −0.32) | Wearable-only ΔR²=0.021 over demographics; full model with labs R²=0.389 |
DWB also ranked nocturnal social-app use (ρ=0.246), night-to-day social-media ratio (ρ=0.222), and polyphasic-sleep percentage (ρ=0.193). The two depression cohorts do not share an identical feature or instrument, so the overlap is construct-level circadian instability, not direct replication. On WEAR-ME, TG/HDL (ρ=0.542) was kept only as a positive control: statistically loud, clinically redundant. Disable construct-overlap gating and a glucosesq leak inflates WEAR-ME CV R² to 0.963; in the full pipeline the Critic rejected that feature even though it passed 10 of 11 checks. Dropping the Scout agent cut DWB R² from 0.228 to 0.120, the largest ablation drop. Adversarial debate barely moves the cross-validated point estimate; it changes what gets reported. DWB started with 145 statistical metrics and kept 34 screened.
Twelve clinicians spent about 25 hours on 16 candidates plus the TG/HDL control. Mean validity was 3.69 / 5 and tracked CoDaS confidence tiers (ρ=0.67, p=0.005). Added clinical value and confidence to act sat at 2.37 and 2.34. Evening incoming-call duration, already flagged unstable by the pipeline, scored the lowest validity (1.75). In blinded report review, CoDaS scored 3.74 overall against 1.63–2.12 for baselines, with 18/21 non-reject decisions; baselines were almost all rejected. On DWB the full pipeline ran about 8.3 wall-clock hours at an estimated $3.91 in API cost.
This is a hypothesis-ranking tool, not a diagnostic. The transferable piece is the engineering: keep computation deterministic, use an adversarial step to kill tautologies, pin numbers to a Fact Sheet, and let clinicians score validity separately from actionability. Sleep variability and a steps/RHR index are not new biology. What CoDaS shows is that a guarded agent loop can recover those signals at population scale, assign confidence tiers, and still report that incremental R² is small.
Follow the pipeline. Do not ship the ρ values as clinical biomarkers.
The whole analysis is exploratory and unregistered. GLOBEM's PHQ-4 endpoint was chosen by the Scout agent for coverage. No identical feature replicated in an independent cohort with a matched instrument. Effect sizes are modest (ρ about 0.15–0.25), and PHQ-8 itself contains a sleep item, so sleep-variability hits may partly be criterion overlap. GLOBEM has 54.6% feature missingness and near-chance AUC. Cohorts skew young to middle-aged; steps and resting heart rate change meaning under mobility limits or beta-blockers. Mechanistic paragraphs are post-hoc literature narratives. Benchmark comparisons pit a specialized multi-agent stack against single models and generic agents, so those margins are system-level, not model-to-model. The 11-check battery is not FDA BEST analytical, clinical, or interventional validation.