Exploring Artificial Intelligence Models for the Detection and Prediction of Sarcopenia: A Scoping Review

Background/Objectives: Sarcopenia is a progressive skeletal muscle disorder associated with adverse health outcomes, including disability, hospitalization, and mortality. Despite the availability of consensus-based diagnostic criteria, its identification in clinical practice remains limited due to reliance on specialized assessments and resource-intensive measurements. Artificial intelligence (AI) and machine learning (ML) approaches have emerged as promising tools to enhance sarcopenia detection; however, evidence remains methodologically heterogeneous and has not been comprehensively synthesized. The objective was to systematically map the existing literature on AI/ML models developed to detect or predict sarcopenia, describing their methodological characteristics, input features, model performance, validation strategies, and adherence to reporting standards. Methods: A scoping review was conducted following PRISMA-ScR guidelines. Searches were performed in five databases (MEDLINE, Scopus, SciELO, Web of Science, and LILACS) from inception to 28 February 2026. Studies were included if they developed and/or validated AI/ML models for binary classification of sarcopenia using recognized consensus definitions. Two reviewers independently screened studies, extracted data, and assessed adherence to TRIPOD+AI reporting recommendations and appraised methodological quality using the NHLBI Quality Assessment Tool. Findings were synthesized descriptively. The review protocol was registered in the Open Science Framework. Results: Of 2910 records, 28 studies (n = 34,498 participants in the primary analytical cohorts, with an additional 2800 participants in independent external validation cohorts) were included, predominantly from Asia (China n = 14, South Korea n = 7). Random Forest was the most frequent algorithm (n = 6), followed by voting classifiers, MLP, and XGBoost (n = 4 each). Clinical tabular data, including anthropometric and laboratory variables, were the most common input features. Model performance was generally high, with AUC values ranging from 0.710 to 0.975 (mean ≈ 0.869), although heterogeneity across populations and validation strategies was substantial. External validation was performed in only 7 of the 28 studies (25%). No study fully adhered to TRIPOD+AI recommendations; mean compliance was moderate, with critical gaps in code sharing, patient/public involvement, external validation, and subgroup performance reporting. Conclusions: AI/ML models, particularly ensemble methods using accessible clinical variables, showed high discrimination in model development settings, with performance estimates that derive predominantly from internal validation and are not comparable across studies. However, important methodological limitations, including lack of external validation, limited generalizability, and suboptimal adherence to reporting standards, were observed. Future research should prioritize external validation in diverse populations, transparent reporting following TRIPOD+AI, and open sharing of code and data to facilitate clinical translation.