Integrating real-world wearable data into breast cancer risk assessment: Evidence from the All of Us Research Program

Lifestyle and genetic factors are known contributors to breast cancer risk, yet their integration with clinical data into breast cancer risk assessment remains limited. Traditional, self-reported lifestyle measures are subject to recall bias, whereas wearable devices provide objective, continuous measurements of physical activity and sleep behaviors. Using data from the National Institutes of Health All of Us Research Program (n=633,540 participants), we conducted a retrospective matched case-control study to evaluate the association between objectively captured wearable data and breast cancer risk, and to establish a scalable analytical framework for causal and machine learning modeling. Females diagnosed with breast cancer at age ≥50 years with at least five valid weeks of Fitbit data (two or more days per week) within the five years preceding diagnosis (n=154) were each matched to up to 20 cancer-free controls by date of birth (±1 year) and availability of wearable data within the same time temporal window. Numerical variables were analyzed using Wilcoxon signed-rank tests, and categorical variables via chi-square analysis. Cases exhibited lower average daily steps (6766 ± 3040) compared to controls (7248 ± 3266; p=0.011), as well as fewer daily light active and very active minutes (179.8 ± 69.0 and 13.6 ± 13.7 vs. 190.2 ± 69.3 and 16.0 ± 16.9; p = 0.043 and p < 0.001, respectively). Sleep metrics were not significantly different between groups, while family history of breast cancer was more common among cases (p < 0.001). Building on these findings, we propose a multimodal integrative framework that merges wearable, survey, and electronic health record data, with future incorporation of genomic features and causal inference techniques (e.g., propensity score matching and causal forests) to refine individualized risk estimation. Explainable machine learning approaches, including ensemble and time-series models, will enable interpretable and dynamically updated risk predictions. This study demonstrates the feasibility of using real-world wearable data within the All of Us infrastructure and underscores the translational potential of multimodal, causal, and interpretable modeling for precision breast cancer screening and prevention at a population scale.