Smartphone step-count patterns and later QMCOO-defined frailty-related risk

Abstract

Smartphone steps offer low-burden signals. Their incremental value for later frailty-related risk defined by the Questionnaire for Medical Checkup of Old-Old (QMCOO) remains uncertain. We analyzed users of Osaka’s Asmile app. Step-count features from November–December 2024 were related to a 12-item QMCOO score ≥ 4 in November–December 2025. Five prespecified Light Gradient Boosting Machine (LightGBM) feature sets underwent internal–external validation across 34 municipality groups. We assessed discrimination, calibration, Brier score, risk enrichment, and exploratory net benefit. Among 22,658 participants (mean age, 63.8 years; 62.5% female), 3,250 (14.3%) met the criterion. Primary versus age-and-sex mean fold areas under the receiver operating characteristic curve (AUC) were 0.592 (95% confidence interval [CI], 0.576–0.609) and 0.547 (0.536–0.559); areas under the precision–recall curve (PR-AUC) were 0.217 (0.197–0.238) and 0.166 (0.153–0.178). The Brier score decreased by 0.0014. Pooled calibration-in-the-large was 0.015 and slope 1.139. Risk increased from 9.6% in the lowest to 23.6% in the highest predicted-risk decile. When both models were redeveloped exclusively among 12,399 participants aged ≥ 65 years, the primary mean fold AUC was 0.592 and risk increased from 8.7% to 18.0%. Step-count patterns modestly enriched identification of later QMCOO-defined frailty-related risk. Discrimination and estimated net benefit remained limited. Without baseline frailty status, these results do not establish incident frailty. These findings provide a rationale for prospective evaluation of targeted in-app prompting for QMCOO completion at the population level; the model is not a standalone screen or clinical decision tool.