Guideline-grounded feature representation for predicting incident and progressive sarcopenia in older adults: A retrospective prediction-model study

Objectives Sarcopenia is reversible when detected early, yet its trajectory is hard to predict. We tested whether encoding the EWGSOP2 and AWGS consensus criteria as model features (signed margins to the cut-offs, criterion flags and interactions) improves prediction of incident and progressive sarcopenia beyond a conventional multimodal machine-learning model, and whether any gain exceeds that of simple threshold information. Design Retrospective prediction-model study. Two feature representations were compared with the same histogram-based gradient-boosting classifier on identical inputs and folds (DeLong test, pre-planned ablation) and against threshold-aware logistic-regression benchmarks. No language-model agents were executed. Setting and Participants 880 community-dwelling and post-acute adults aged 60 years or older (Ankara, Türkiye) with paired baseline and approximately 12-month EWGSOP2 classification adjudicated by two raters; 834 were at risk of incident sarcopenia (172 events, 20.6%) and 861 of progression, any increase in EWGSOP2 category (224 events, 26.0%). Methods Discrimination (AUROC) was the primary endpoint; calibration was assessed before and after pre-planned isotonic recalibration, with decision-curve analysis, operating points and sensitivity analyses (stricter endpoint definitions, AWGS 2025 thresholds). Results For incident sarcopenia the representations were equivalent (AUROC 0.822 vs 0.811; P=0.22). For progression the guideline-grounded representation discriminated better (0.762 vs 0.734; Δ +0.028, P=0.006), but adding the baseline EWGSOP2 category and three signed margins to the conventional model reproduced the gain (0.766), and a parsimonious logistic model with those terms, age and sex discriminated at least as well (0.792 for progression, 0.844 for incident sarcopenia). Ablation localised the gain to the physical-function and body-composition blocks. Isotonic recalibration corrected over-extreme predictions (slope 0.57 to 0.82). Conclusions and Implications Proximity to the consensus thresholds carries prognostic information about progression but not about first onset. A few transparent terms capture it, and a complex learner is not required. Trajectory prediction anchored to consensus criteria is feasible, but external validation is needed before clinical use.