Fitbit-derived activity and heart-rate patterns for predicting 12-month hospitalization in an NIH All of Us cancer cohort.

Abstract

1629 Background: Hospitalizations in oncology are common and diverse, and are often limited in their predictive accuracy with routine variables. Wearable devices capture functional and physiological patterns that may indicate vulnerability. We examined whether Fitbit-derived activity and heart rate features enhance the prediction of all-cause hospitalization beyond demographic and clinical factors like frailty. Methods: Using data from the NIH All of Us Research Program, we conducted a retrospective analysis of participants aged ≥50 years with a history of cancer, linked electronic health records, and short-term Fitbit data (with imputed missing days). Baseline predictors included demographic variables, clinical factors such as the All of Us frailty index, and indicators of cancer subtype. The outcome was any hospitalization within 12 months of the index date, relative to the last Fitbit data day. Features included step count, step variability, activity distribution, rhythmicity, heart rate, and heart rate variability. Missing data were imputed using predictive mean matching. Models were optimized via nested stratified cross-validation and evaluated with out-of-fold predictions. Discrimination was measured by AUC and average precision with bootstrap confidence intervals. Predictive value for the top 10% at-risk group was also assessed. Results: The study cohort consisted of 278 participants (average age 64.6). Twelve percent (35) experienced hospitalization within a year. Time to hospitalization ranged up to 342 days. Hospitalization diagnoses were heterogeneous. Models using demographic and clinical data alone had limited discrimination (AUC 0.41–0.50). Incorporating Fitbit activity features improved logistic performance: average steps yielded an AUC of 0.522, step-pattern quartiles yielded 0.559, and distributional rhythmic features yielded 0.628 (CI 0.520–0.733; average precision 0.283). Heart rate and variability did not materially enhance logistic prediction. Gradient-boosted trees combining activity and heart rate data achieved the highest discrimination (AUC 0.704; CI 0.621–0.786; average precision 0.286). Using the best-performing model, the top 10% predicted-risk group had a hospitalization rate of 28.6% (8/28), compared with the overall cohort hospitalization rate of 12.6%, capturing 8 of 35 events. Conclusions: Fitbit-derived activity patterns and heart rate patterns, when modeled using machine learning, improved near-term hospitalization prediction in an NIH cancer cohort beyond demographic and clinical models, supporting their potential for practical risk assessment. Ongoing analyses will address missingness, calibration, and subgroup robustness to inform prospective validation.