Abstract
<title>Abstract</title> <p>Background Later survey waves may include participants who contributed to model development, compromising the independence of cross-wave evaluation. We evaluated screening for possible sarcopenia in unseen CHARLS participants, compared a parsimonious model with an anthropometry-augmented model, and examined threshold trade-offs. A supportive analysis explored whether changes after participant isolation could be attributed to overlap alone. Methods We analysed adults aged 60 years or older, using CHARLS Wave 1 for development and Wave 3 for evaluation. Possible sarcopenia was defined using an AWGS 2019-aligned operational definition based on handgrip strength or five-chair-stand criteria. The primary evaluation excluded all Wave 1 IDs from Wave 3. We compared a logistic regression model using core demographic, socioeconomic, lifestyle, and chronic-disease predictors with an XGBoost model that additionally included height, weight, body mass index, and waist circumference. Both fitted models were applied without retraining. Model differences were assessed on a common unseen-participant subset. A supportive post hoc analysis compared overlap-only with unseen-only participants using fixed models, bootstrap confidence intervals, and case-mix summaries. Results The common unseen-participant comparison subset included 2430 participants and 683 events. In the participant-isolated evaluation, AUROC was 0.6798 (95% CI, 0.6554–0.7027) for the parsimonious model and 0.6656 (95% CI, 0.6414–0.6900) for the enhanced model. The paired AUROC difference was 0.0116 (95% CI, − 0.0028–0.0263), showing no clear advantage from added complexity. At the 0.50 threshold, accuracy exceeded 0.71 for both models, but sensitivity remained below 0.33. Sensitivity-oriented thresholds increased sensitivity to approximately 0.75 while reducing specificity to approximately 0.47. Added anthropometric predictors accounted for 9.2% of normalised XGBoost gain importance. Original pooled AUROCs were higher, but overlap-only versus unseen-only AUROC differences were uncertain and accompanied by marked differences in age and outcome prevalence. Conclusions In unseen participants, added model complexity and anthropometric predictors did not clearly improve screening performance. Apparently acceptable accuracy masked low sensitivity at the default threshold. Participant isolation is necessary for independent evaluation, but changes after isolation should not be attributed to overlap alone when case mix also changes.</p>