The search strategy was developed and subsequently refined in collaboration with a medical librarian specialist (P. X.) and a senior expert in systematic reviews (J. Y.). Using the Ovid SP platform, we searched four major biomedical databases: MEDLINE, Embase, the Cochrane Central Register of Controlled Trials, and the Science Citation Index Expanded. The database-specific search strategies are provided in Supplementary Table S1. An updated search was conducted up to 29 September 2025. Additionally, reference lists of included studies were manually screened to ensure comprehensive coverage of relevant literature.

This review covered all relevant studies that have developed multivariable prediction models, tools, indices, or scoring systems (defined as models containing at least two predictors) aimed at assessing the risk of sarcopenia in older adults. We specifically searched for studies related to the development of sarcopenia prediction models. Studies that solely evaluated the diagnostic accuracy or external validation of existing screening tools or scores, without model development, were not included in the primary synthesis. We structured the inclusion criteria for this review based on the PICOTS framework (Population, Intervention, Comparator, Outcomes, Timing, and Setting), commonly used in systematic reviews of prediction models (Supplementary Table S2): (1) Population: Studies involving older adults from diverse settings, including community-dwelling individuals, nursing homes, hospitals, and those with chronic diseases or cancer; (2) Intervention: Models predicting the presence of sarcopenia (diagnostic models) or the future risk of sarcopenia development (prognostic models); (3) Comparator: No specific model comparison was required; (4) Outcomes: The outcome of interest was the presence of sarcopenia at a given time point or the risk of developing sarcopenia over a future time period; (5) Timing: No specific time restrictions were imposed on the measurement of predictors or outcomes; (6) Setting: Community, primary care, long-term care facilities, or hospital settings.

To reduce heterogeneity across studies, diagnostic criteria for sarcopenia were restricted to widely accepted definitions, including AWGS (2014/2019) (2, 18) and EWGSOP (2010/2018) (7, 19), which defined sarcopenia as low muscle mass (LMM) combined with low muscle strength (LMS) and/or low physical performance (LPP). Studies were excluded if they: (1) Lacked clear diagnostic criteria for sarcopenia or defined it using a single parameter; (2) Systematic reviews, meta-analyses, commentaries, or methodological studies; (3) Focused on genetic/biological markers of sarcopenia or sarcopenia screening scales; (4) Predictive equation or the full text was unavailable; (5) Non-English publications; (6) Duplicated publications.

Two reviewers (T. L. and N. S.) independently screened the titles and abstracts to identify potentially relevant studies. The study selection process followed three main phases: (1) Title and Abstract Screening: All retrieved references were imported into EndNote software, where duplicates were automatically removed. Two independent reviewers screened the titles and abstracts of the remaining studies against predefined inclusion criteria to exclude irrelevant articles. A third reviewer cross-checked the results to ensure consistency. Discrepancies were resolved through facilitated discussions led by the third reviewer. When consensus could not be reached, the issue was escalated to the full research team for resolution. (2) Full-Text Screening: Studies passing preliminary screening underwent detailed full-text review according to the inclusion and exclusion criteria. Two reviewers independently assessed the main texts and Supplementary materials to confirm eligibility. The extracted data were cross-verified for accuracy. Inter-rater agreement during full-text screening was quantified using the Kappa statistic, which evaluated observer agreement while accounting for chance. The Kappa statistic, categorized as poor (<0), mild (0.20–0.40), moderate (0.41–0.60), high (0.61–0.80), and substantial (0.81–1.00), confirmed strong agreement among reviewers. (3) Handling Incomplete Data: For studies with missing or incomplete data, corresponding authors were contacted via email or phone. Any disagreements during the process were resolved through discussion with a third researcher (J. Y.).

Data extraction adhered to a standardized form developed using the CHARMS checklist (14). Two reviewers (H. L. and P. X.) independently extracted data, and a third reviewer cross-verified the extracted information. The key elements captured were: (1) General information: First author, year of publication, journal, study type, and geographic location. (2) Study details: Data source, setting, population, and sample size. (3) Model specifics: Predictors, sarcopenia definition, model development/validation approach, performance metrics, and presentation format. (4) Methodological details: Handling of missing data and use of continuous variables. To guard against overly optimistic evaluations of prediction model performance, data from internal or external validation datasets were prioritized. If validation data were unavailable, performance metrics were conservatively extracted from the lowest-performing subgroup in the training dataset.

To evaluate the predictive ability of each model, we extracted the key indicators related to the performance of discrimination, calibration and classification. To evaluate how well each model distinguishes between individuals who develop sarcopenia and those who do not, we used the area under the receiver operating characteristic curve (AUC) as the primary measure of discrimination (15). In the absence of validation data, we extracted the minimum reported AUC value as the most conservative estimate. For research that proposed multiple models, priority should be given to the model with the most complete information, including the complete prediction equation. The AUC values were categorized as poor (0.50–0.59), average (0.60–0.69), good (0.70–0.79), very good (0.80–0.89), and excellent (≥0.90) (15). Calibration assessed the consistency between the predicted and observed results, using calibration intercepts and slopes for evaluation (20). Extract the classification performance when available to provide a comprehensive understanding of the diagnostic accuracy of the model, such as sensitivity, specificity, accuracy, positive predictive value, and negative predictive value.

We systematically applied the Prediction model Risk Of Bias ASsessment Tool (PROBAST) to evaluate both the risk of bias and the applicability of each prediction model (Q. X. and J. L.). PROBAST assessed four key domains—participants, predictors, outcomes, and analysis—using 20 signaling questions to evaluate methodology related to study design, conduct, and data analysis (11). Each question was categorized as “Yes,” “Probably Yes,” “Probably No,” “No,” or “No Information.” The overall risk of bias for each domain was classified as low, high, or unclear. A domain was judged to be at high risk of bias if one or more signaling questions are answered as “No” or “Probably No.” A low-risk domain required all questions to be “Yes” or “Probably Yes.” A prediction model was rated as having low overall risk of bias only if all four domains are assessed as low risk. If any domain was rated as high risk, the overall judgment is high risk. If at least one domain had an unclear risk of bias, and all other domains were rated as low risk, the model was categorized as having an unclear overall risk of bias. In addition to the bias assessment, the applicability of the prediction model was evaluated based on the first three domains—participants, predictors, and outcomes. Applicability concerns were classified as low, high, or unclear based on expert judgment, without the use of specific signaling questions. Any discrepancies in judgments were resolved through discussion among the reviewers to ensure consensus. By utilizing PROBAST, it was possible to critically appraise the methodological rigor and applicability of prediction models, offering a scientifically robust foundation for interpreting the reliability and generalizability of the research findings.

Considering the significant heterogeneity of the included studies in terms of the study population, study design, model development and validation methods, and the final included predictor subsets, we did not conduct quantitative synthesis of the reported predictive performance (AUC values), such as summary effect size estimation, subgroup analysis, sensitivity analysis, or publication bias assessment (15, 21). Instead, we used Stata/SE version 15.1 (StataCorp LP, College Station, TX, USA) to summarize the reported AUC values of the prediction model in the original study and generate forest plots to represent the model performance comprehensively (15). Furthermore, we used Microsoft Excel to construct a bar chart based on the PROBAST assessment to illustrate the overall risk of bias and applicability ratings of the included prediction models. The bar chart was also used to show the most common predictors in the model. These findings provided practical references for the selection of variables when developing and validating high-quality sarcopenia prediction models in the future. Furthermore, we summarized the baseline characteristics of all the included models. Continuous variables were reported as the median of the quartile range (IQRs), while categorical variables were expressed as frequency and percentage.

A comprehensive literature search on August 4, 2023 identified 13,866 records. After removing 5,802 duplicates, 8,064 records were screened, resulting in 64 full-text reviews, of which 18 studies met the inclusion criteria (Figure 1) (9, 22–38). Additionally, the literature search was updated on 29 September 2025 and another 8 studies were included (Figure 1) (39–46). Exclusion reasons for full-text reviews were detailed in Supplementary Table S3. Inter-reviewer agreement was quantified using the Kappa statistic, resulting in a value of 0.80, indicating strong agreement between reviewers.

The included studies, summarized in Table 1, involved 33,273 participants across China (n = 21), Japan (n = 3), South Korea (n = 1), and Iran (n = 1). Thirteen models (50%) were developed in community-based populations, while the remaining studies were hospital-based patients with chronic disease. Similarly, half targeted general older adults, while others focused on specific populations, including dialysis patients (n = 4), diabetes mellitus (n = 5), stroke (n = 1), rheumatoid arthritis (n = 1) and cancer patients (n = 1). Most models (n = 23) were derived from cross-sectional studies for diagnostic purposes, and only three cohort-based models focused on prognosis. Sample sizes ranged from 137 to 4,610, with sarcopenia prevalence between 5.8 and 22.5%. Except for three studies, all sampled participants were aged 60 years or older.

Significant variations existed in the diagnostic methods and cutoff values for sarcopenia (Supplementary Table S4). Most studies assessed LMM using bioelectrical impedance analysis (BIA, n = 18), while four employed dual-energy X-ray absorptiometry (DXA). Muscle strength was consistently measured via handgrip strength, whereas physical performance was evaluated using diverse methods, including 6-meter (n = 14), 4-meter (n = 3), and 5-meter (n = 2) walk tests. Four studies incorporated multiple measures, such as gait speed, Short Physical Performance Battery (SPPB), or chair stand tests. Diagnostic criteria for sarcopenia definition also varied: 21 used AWGS 2019 standards, 3 used AWGS 2014, 1 used EWGSOP1, and 1 adopted EWGSOP2.

Over 50 predictors were identified and grouped into demographics, health-related risks, chronic diseases, socioeconomic factors, and body composition. Among these, eight predictors were used in multiple models, including age (n = 20), BMI (n = 16), gender (n = 7), calf circumference (n = 7) and so on. Age was the most frequently integrated predictor, reflecting its strong association with sarcopenia (Figure 2).

Nine studies did not report handling of missing data. Of the remaining 17, 12 excluded incomplete records, three applied multiple imputation, one used mean substitution, and one explicitly stated no missing data. For continuous variables, 16 models retained continuous predictors, while 8 models classified continuous variables into binary or multi-level categorical variables, potentially reducing predictive power. One model provided unclear information, and another treated all variables as categorical (Table 1).

Most models (n = 12, 46.1%) were developed using multivariable logistic regression. Other techniques included machine learning (n = 1), ridge regression (n = 1), and combined methods like LASSO and logistic regression (n = 12). Internal validation was conducted in 19 studies via bootstrapping (n = 10), cross-validation (n = 3), and data splitting (n = 6). We did not identify any separate external validation reports for already published sarcopenia models. However, 7 studies lacked internal validation. External validation was performed in only 6 studies (Table 1).

All 26 models reported discrimination via AUC values, 20 models achieving AUC > 0.80 (strong discrimination) and the remaining 6 within 0.70–0.80 (acceptable). Calibration metrics were inconsistent—19 studies reported calibration methods (e.g., Hosmer-Lemeshow test, calibration curves). Models were presented as regression formulas (n = 3), nomograms (n = 12), risk score tables (n = 3), or combined formats (n = 4); 4 studies did not report presentation formats (Table 1).

The PROBAST tool indicated high overall risk of bias in 25 studies (96.1%). For applicability, 11 studies (42.3%) exhibited high risk, while 15 studies (57.7%) were deemed low risk (Figure 3). Domain-specific bias assessments were presented as follows: (1) Participants domain: 1 study excluded participants ≥85 years and was rated as high risk of bias due to limited generalizability.; (2) Predictor domain: 5 studies were rated as high risk due to including handgrip strength as a predictor, thus introducing overestimation bias and the risk of overfitting. 1 study was rated as having unclear risk, as it lacked sufficient detail on how predictors were assessed; (3) Outcome domain: 10 studies were classified as high-risk because they failed to exclude factors such as grip strength from the definition of the results, or did not follow the international consensus on sarcopenia. 2 study was classified as having unclear risk because there was a lack of relevant information on whether the predictive factors were excluded from the result determination; (4) Analysis domain: Most models (n = 23) exhibited high risk due to inadequate sample sizes, arbitrary categorization of continuous variables, poor handling of missing data, reliance on univariate predictors, incomplete performance assessments, and lack of internal or external validation. Regarding applicability, 9 studies were rated high risk in the participant domain due to inconsistent target populations (e.g., younger participants or exclusion of the “oldest old” or combined with chronic diseases). In the predictors and outcomes domains, 8 studies exhibited high risk because predictors (e.g., handgrip strength) overlapped with outcome definitions.

In summary, while 96.1% of models showed high bias risk, 42.3% had acceptable applicability. These methodological flaws limited their performance and generalizability in clinical practice (Figure 3).

Due to significant heterogeneity in population, model design, validation strategies, and predictors, combined effect size or subgroup analyses were not performed. Instead, AUC values were summarized and visualized in a forest plot (Figure 4), with all models achieving AUCs above 0.70, indicating good discriminatory ability. However, the absence of external validation and comparative analyses hindered drawing conclusions on effectiveness. Methodological issues, including non-pragmatic predictor selection, missing data, inconsistent handling of continuous variables, and poor reporting, contributed to high bias risk and concern the reliability of AUC values. Although many models demonstrated promising performance, these results should be interpreted with caution. Future studies should focus on external validation and methodological improvements.