SensorFM:1兆分間の雑多なデータがいかに個別最適化ヘルスモデルを葬ったか

google research introduced sensorfm, a foundation model designed to unify the processing of messy wearable sensor data, trained on a dataset of over one trillion minutes collected from five million users across 100 countries using 20 different device models. the model utilizes self-supervised learning via adaptive and inherited masking to reconstruct masked and missing sensor segments from photoplethysmography, acceleration, and skin conductance inputs, demonstrating systematic performance improvements as model parameters and data volume increase. in evaluations against three separate clinical studies with nearly 14,000 participants, sensorfm outperformed monitored comparison models that used specially prepared features on 34 out of 35 tasks related to cardiovascular, metabolic, and mental health. additionally, health summaries augmented with sensorfm predictions were rated significantly higher than baselines, suggesting the model can improve user-facing health insights and reduce reliance on expensive labeled datasets.