Can Smartwatches Predict Diseases? Let's Find Out!

Data from our wearable devices, like smartwatches and fitness trackers, are reshaping how doctors predict diseases. Traditionally, hospitals rely on electronic health records (EHR), which only capture events that happen in the clinic: visits, tests, diagnoses. But what about everyday life, with its ups and downs, that stays out of these records?\n\nA research team merged EHR information with continuous activity, heart‑rate, and sleep data collected by wearables, using the massive "All of Us" database. They tested ten clinical outcomes, from diabetes to depression, comparing different encoding methods: from traditional feature engineering to more advanced foundation models.\n\nThe result was striking: adding wearable data boosted model predictive power by about 8.5% on average. For depression, AUROC rose by 6.8%; for hypertension by 9.7%; and for diabetes by a remarkable 12.6%.\n\nThese numbers show that signals gathered outside the doctor's office – like steps taken or sleep quality – can fill gaps in clinical information, making predictions more personal and timely. The next step will be to fine‑tune AI models to fully exploit wearable data, paving the way for more preventive, customized health care for everyone.

Integrating data from wearable devices with electronic health records (EHR) poses a multidisciplinary challenge involving medical informatics, bioinformatics, and advanced statistics. Methodologically, researchers compared two encoding approaches: (i) traditional feature engineering, which transforms raw signals (e.g., step count, heart‑rate variability) into aggregated variables (daily mean, standard deviation, percentiles); (ii) embeddings based on foundation models pre‑trained on large biometric time‑series, capable of capturing non‑linear patterns and long‑term dependencies.\n\nPerformance was assessed using AUROC (Area Under the Receiver Operating Characteristic curve), a robust metric for binary classification, especially when classes are imbalanced. The average AUROC increase of 8.5% indicates that wearable signals provide complementary information absent from EHR alone, likely linked to daily behaviors (physical activity, sleep quality) that directly affect risk factors.\n\nA critical issue is handling missing data, common in wearables due to removal or malfunction. The authors employed imputation techniques based on sequential learning models (e.g., LSTM) to reconstruct time series. Additionally, data normalization reduced bias stemming from device or population differences.\n\nClinically, the most pronounced gain for diabetes (+12.6%) may reflect the tight correlation between activity, circadian rhythm, and glycemic control. For depression, sleep patterns and heart‑rate variability are known emotional state biomarkers, explaining the +6.8% improvement.\n\nFuture directions include adopting more sophisticated multimodal models, such as hybrid transformers that combine wearable time series with textual encodings of clinical notes. Personalizing models for demographic subgroups (age, gender, ethnicity) could mitigate performance disparities. Finally, data privacy remains a challenge: federated learning techniques would allow training on distributed data without transferring sensitive information to a central server.