ABSTRACT Physical activity is central to cancer survivorship care, yet the prognostic value of its multiple dimensions remains poorly characterized. We aimed to define multidimensional, wearable‐derived digital phenotypes of physical activity; determine their independent associations with all‐cause mortality; identify dose–response thresholds for survival benefit; and evaluate their utility in machine learning mortality prediction among cancer survivors. This observational cohort study analyzed 9379 National Health and Nutrition Examination Survey (2011–2014) participants with wrist accelerometry and mortality follow‐up, including 864 cancer survivors. Three Monitor‐Independent Movement Summary (MIMS) phenotypes were derived: Volume, Intensity, and Variability. In fully adjusted Cox models, vigorous versus light patterns were associated with lower mortality (hazard ratios 0.33, 0.22, and 0.29 for Volume, Intensity, and Variability, respectively). Dose–response thresholds for survival benefit were identified (Intensity > 24, Volume > 4945, and Variability > 2.4). Protective associations were stronger in survivors with low functional limitation. Machine learning models discriminated well (Random Survival Forest AUC = 0.799), with external validation confirming generalizability (AUC = 0.773). Multidimensional digital phenotypes, particularly for peak intensity and variability, are independent mortality predictors, supporting their integration into precision cancer survivorship care. Interpretation is limited by the absence of cancer staging and treatment data.
Li et al. (Wed,) studied this question.