Use of parsimonious Fitbit step metrics to improve frailty classification in adults ≥50 with cancer in NIH All of Us.

1657 Background: Frailty predicts treatment intolerance and adverse outcomes in older adults with cancer, yet scalable screening remains limited. Scalable frailty screening is particularly relevant for treatment selection and risk stratification in routine oncology care. We tested whether routine Fitbit step metrics improve frailty classification beyond basic variables and whether high-dimensional step-pattern features add benefit. Methods: Retrospective cross-sectional study in the NIH All of Us Research Program including adults ≥50 years with cancer history and Fitbit step data (N=378). Frailty was defined by a 33-item All of Us deficit frailty index (fit 0.25); primary endpoint was frail vs not frail (fit+vulnerable). Valid wear day required ≥10 hours with activity or heart-rate evidence. Missing days were imputed for daily totals and minute-level profiles reconstructed using within-person donor days or weekday/weekend templates. Prespecified feature tiers were: basic (9), basic+average steps (10), and high-dimensional step-pattern features (108; minute-level intensity, variability, and frequency-domain summaries). Class-weighted elastic-net logistic regression, random forest, and gradient-boosted trees were tuned and evaluated with nested 5×5 cross-validation. Discrimination was pooled out-of-fold area under the receiver operating characteristic curve (AUC) with stratified bootstrap 95% confidence intervals (CI); incremental value used paired bootstrap ΔAUC. Results: Participants were 64.8±8.6 years; 67.7% were female. Frailty categories were fit 43.7%, vulnerable 40.7%, and frail 15.6% (59/378). Mean daily steps were lower in frail vs not frail participants (5505 vs 7648 steps/day). Elastic-net logistic regression with step features achieved the highest AUC (0.713, 95% CI 0.636–0.783) vs the basic model (0.644, 95% CI 0.567–0.720; ΔAUC 0.068, paired bootstrap 95% CI 0.002–0.133). At ≥90% specificity, sensitivity increased from 25.4% (15/59) to 39.0% (23/59). In secondary analyses predicting not-fit status (vulnerable/frail vs fit), adding step features improved AUC from 0.551 (95% CI 0.493–0.609) to 0.635 (95% CI 0.580–0.694). High-dimensional step features did not improve AUC. Conclusions: In adults ≥50 with cancer, a parsimonious step-based metric improved frailty classification and increased sensitivity at high specificity, while complex feature sets and models did not add benefit. These findings support low-burden step metrics as scalable adjuncts to frailty screening in oncology; external validation is needed.