This retrospective cohort study developed machine learning models to predict sarcopenia in community hospital patients. This study was approved by Institutional Review Board of Changshou Community Health Center and the Ethics Committee of Putuo People’s Hospital of Shanghai (Number:2025006), registered at the Chinese Clinical Trial Registry (Registration No.: ChiCTR2500099224, https://www.chictr.org.cn/) and filed in the Chinese Medical Research Registration and Filing System (MR-31-25-022873). Meanwhile, the patients’ informed consent was acquired. This study adhered to the TRIPOD reporting standards for transparent reporting of multivariable prediction models in individual prognosis or diagnosis.

Sarcopenia was diagnosed based on the 2019 consensus criteria of the Asian Working Group for Sarcopenia (AWGS) (Chen et al., 2020). The diagnosis required the co-presence of low muscle mass and low muscle strength. Low muscle mass was defined as a skeletal muscle mass index (SMI) below the sex-specific AWGS 2019 cut-offs: <7.0 kg/m² for men and <5.7 kg/m² for women. Low muscle strength was defined as handgrip strength <28 kg for men and <18 kg for women.

The dataset was obtained from the Changshou Community Health Center, Six patients with incomplete or missing records were excluded, 1650 patients were included in this study. The Synthetic Minority Oversampling Technique (SMOTE) was employed to address class imbalance (Du et al., 2024), and patients were randomly allocated into training and test sets at a 7:3 ratio using computer-generated randomization.

The variables selected to predict the occurrence of sarcopenia included various demographic and clinical variables. These 35 variables were based on domain expertise, and features with potential discriminative power to predict sarcopenia in clinical practice. Demographic variables included gender, age, weight, height, BMI, marriage status, educational experience, living condition, pension, sleep duration, outdoor activity time, living floor, living area, smoking, drinking, building with elevator or not, coffee and tea, physical exercise and teeth. Clinical variables included hypertension, diabetes, COPD, hyperuricemia, osteoarthritis, osteoporosis, tumor, CKD, other disease, chronic disease number, drug amount, MNA-SF score, GLIM phenotypic criteria, GLIM etiologic criteria, GLIM-malnutrition, and SARC-Calf score.

Twelve machine learning models: random forest learner (RF), k-nearest neighbour learner (K-NN), support vector machine learner (SVM), extreme gradient boosting learner (XGBoost), light gradient boosting machine learner (LightGBM), CatBoost learner, naïve bayes learner, Decision tree learner, gradient boosting decision tree learner, Adaboost, logistic classification learner, multilayer perceptron learner were constructed, and in model development and comparison (Du et al., 2025b), we utilized 5-fold cross-validation, which provides a robust and reliable method for evaluating model performance metrics.

Each model was evaluated using multiple performance metrics: precision, recall, accuracy, F1 score, prevalence, specificity, false negative rate (FNR), false positive rate (FPR), matthews correlation coefficient (MCC), and the area under the receiver operating characteristic curve (AUROC). Accuracy represented the proportion of correctly classified instances among all observations. Precision denoted the ratio of true positive predictions to all positive predictions, while recall indicated the ratio of true positives to all actual positive instances. The F1 score was the harmonic mean of precision and recall. Prevalence referred to the proportion of actual positive instances in the dataset. Specificity measured the ratio of true negatives to all actual negative instances. FNR represented the proportion of actual positives incorrectly predicted as negatives, while FPR indicated the proportion of actual negatives incorrectly predicted as positives. MCC provided a balanced measure of classification quality, calculated as a correlation coefficient between observed and predicted binary classifications. Finally, the AUROC curve graphically represents the trade-off between the true positive rate (recall) and the false positive rate (1 - specificity) across different classification thresholds.

“Shapviz” is an R package designed to interpret machine learning model predictions by providing visual explanations based on SHAP (SHapley Additive exPlanations) values. SHAP values quantify the contribution of each feature to individual predictions, indicating either positive or negative effects. The package includes a feature importance plot, which ranks features by the average magnitude of their SHAP values, thus highlighting those with the greatest influence on the model’s output. In this study, SHAP was used to identify which variables play an important role in the prediction of sarcopenia.

The top five contributing variables selected by SHAP in CatBoost, LightGBM and GBDT models were involved in logistic regression model and established a nomogram to predict sarcopenia. Calibration was employed to evaluate the performance of the aforementioned nomogram models. A calibration plot, which visually represents the relationship between predicted and actual risk, was constructed using bootstrap resampling methods. On these plots, a 45-degree diagonal line indicates excellent absolute risk estimation. Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) were both employed to compare the discriminative ability between the models based on CatBoost, LightGBM and GBDT contributing variables.

Continuous variables with a normal distribution were presented as mean ± standard deviation and compared using the independent samples t-test, whereas non-normally distributed variables were expressed as median with interquartile range (IQR) and analyzed with the Mann-Whitney U test. Categorical variables are presented as frequency (percentage) and analyzed with the chi-square test to assess the demographic characteristics and clinical pathological data of participants. Comprehensive analyses and data visualization were conducted using Python Version 3.10.6 and R Version 4.5.1. Statistical significance was set at a P-value threshold of less than 0.05 (two-sided).

A total of 1,656 patients from Changshou Community Health Center were initially enrolled in this study. After excluding six patients due to incomplete or missing medical records, 1,650 eligible participants were retained and balanced by SMOTE method. The patient selection process was illustrated in Figure 1. The cohort was randomly partitioned into a training set (N = 672) and a testing set (N = 288). The normality of continuous variables was assessed using the Kolmogorov-Smirnov test, with results detailed in Table 1. All tested variables—including age, height, weight, BMI, living floor, MNA-SF score, and SARC-Cal score—showed significant deviations from normality (all p < 0.001), indicating non-parametric distributions. The density distributions of above seven variables are visually presented in Figure 2.

The baseline characteristics of participants in the training and testing cohorts were summarized in Tables 2, 3, respectively. Overall, the two cohorts exhibited broadly comparable distributions across key demographic and clinical profiles, supporting the robustness of the subsequent modeling approach.

In the training cohort (N = 672), 331 participants (49.3%) were diagnosed with sarcopenia, compared to 149 (51.7%) in the testing cohort (N = 288). Significant differences between the sarcopenic and non-sarcopenic groups were observed in both cohorts for core anthropometric measures—including age, height, weight, and body mass index (BMI)—as well as for functional and nutritional assessments such as the MNA-SF score and SARC-Cal score (all p < 0.001). Numerous socio-demographic, comorbidity, and lifestyle variables (e.g., gender, marriage, living status, pension level, physical activity, and several chronic diseases) also showed statistically significant differences between groups (p < 0.05). In contrast, the prevalence of certain conditions, including diabetes mellitus and hyperuricemia, did not differ significantly between the sarcopenic and non-sarcopenic groups in either cohort (p > 0.05). Similarly, the variable “Drinking” showed no significant difference in either cohort (training: p = 0.327; testing: p = 0.119), indicating balanced baseline characteristics for these specific parameters. The overall similarity between the training and testing sets enhanced the reliability and generalizability of the model developed in this study.

Multiple machine learning models were developed to predict risk factors associated with the occurrence of sarcopenia. Based on their superior discriminative performance, CatBoost, LightGBM, and Gradient Boosting Decision Tree (GBDT) were selected for final evaluation. All three models achieved exceptional area under the receiver operating characteristic curve (AUROC) values above 0.995, with CatBoost reaching a near-optimal score of 0.999 (Table 4). This high discriminative ability was visually reflected in their ROC curves, which occupied the upper-left corner of the plot, confirming excellent diagnostic accuracy (Figure 3A). Beyond AUROC, these models exhibited balanced performance across other key metrics: CatBoost and LightGBM attained perfect sensitivity (Recall = 1.000), while all three maintained high precision (>0.967), F1-scores (>0.976), and substantial Matthews correlation coefficients (>0.951). Decision curve analysis (Figure 3B) demonstrated that using any of the three models yielded substantially greater net clinical benefit across a wide range of threshold probabilities compared to “treat-all” or “treat-none” strategies, underscoring their potential utility for risk stratification. Furthermore, calibration curves for all models closely followed the ideal diagonal (Figure 3C), indicating well-calibrated and reliable probability estimates. Together, the near-perfect discrimination, meaningful net benefit, and strong calibration support the robustness of these ensemble models. This suggests that the models hold potential clinical utility for guiding decision-making in sarcopenia risk intervention.

The clinical utility and robustness of the top-performing models—CatBoost, LightGBM, and GBDT—were further assessed through decision curve analysis (DCA), calibration plots, and precision-recall (PR) curves, as summarized in Figure 4. These evaluations were conducted across all five test folds of the 5-fold cross-validation, with corresponding accuracy results provided in Table 5.

As shown in Figures 4A–C, the DCA indicated that all three models yielded a higher net benefit over a wide range of risk thresholds compared to the “treat all” or “treat none” strategies, supporting their potential clinical value for identifying individuals at risk of sarcopenia. The calibration plots showed close agreement between predicted probabilities and observed outcomes, with low mean absolute errors (e.g., 0.001) across test folds, confirming excellent calibration reliability (Figures 4D–F). The PR curves further affirmed the models’ strong performance, with the area under the PR curve (AUC-PR) approaching or reaching 1.0 in several folds (Figures 4G–I). For example, CatBoost and LightGBM attained perfect AUC-PR values (AUC = 1) in at least three of the five folds, while GBDT also maintained consistently high AUC-PR scores, reflecting a robust balance between precision and recall despite class imbalance.

Overall, these results align with the high cross-validation accuracy reported in Table 5 (mean scores >0.98), confirming that the models are not only accurate but also well-calibrated and clinically applicable. LightGBM, which achieved the highest mean accuracy (0.992), also demonstrated the most consistent and superior performance in both calibration and PR analyses.

Figure 5 illustrated the SHAP analysis results for the three top-performing models- CatBoost, LightGBM, and GBDT-elucidating the contribution of individual features to sarcopenia risk prediction. Figures 5A–C depicted the summary plots of feature importance, quantified by the mean absolute SHAP value, for each respective model. The SARC-CalF score and BMI consistently emerged as the most influential predictors across all models. Other prominent features included weight, Gender, GLIM_Phenotypic criteria, MNA_SF_score, osteoporosis and age, though their relative rankings varied. For example, Gender ranked highly in LightGBM and CatBoost but was less dominant in GBDT, which assigned greater importance to weight.

Figures 5D–F present beeswarm plots that visualize the distribution and directional influence of each feature. Elevated SARC-CalF score, advanced age, male, and lower body weight were consistently associated with increased sarcopenia risk (positive SHAP values). Conversely, elevated MNA_SF_score, higher BMI and female were generally correlated with reduced risk (negative SHAP values). The directional agreement of these key predictors underscores their robustness.

In summary, the SHAP analysis not only confirms the strong effects of known sarcopenia risk factors but also affords model-specific interpretability, revealing both shared and distinct feature contribution patterns among the ensemble methods.

The performance of the final tree-based ensemble models (CatBoost, LightGBM, and GBDT) was evaluated against logistic regression (Table 6; Figure 6). Decision curve analysis (Figure 6A) indicated that all three ensemble models provided greater net clinical benefit across most decision thresholds than both the “treat-all” or “treat-none” strategies and the logistic regression model, underscoring their superior potential for guiding clinical interventions. Although receiver operating characteristic (ROC) curves confirmed high discriminative ability for all classifiers (AUC range: 0.995 to 1.000; Figure 6B), logistic regression failed to identify any positive cases (sensitivity = 0), rendering its F1-score incalculable (Table 6). In contrast, the tree-based models maintained a balanced performance profile: CatBoost and LightGBM achieved perfect sensitivity (1.000) with high precision (>0.973), and all three models demonstrated strong overall accuracy. Furthermore, calibration analysis (Figure 6C) showed that all models were well calibrated, with mean absolute errors below 0.015, confirming the reliability of their predicted probabilities. Together, the demonstrated net clinical benefit, balanced classification performance, and reliable calibration supported the clinical applicability of these ensemble models over traditional logistic regression for this predictive task.

Nomograms were constructed from the CatBoost, LightGBM, and Gradient Boosting Decision Tree (GBDT) models (Figures 6D,F,H) to serve as intuitive tools for estimating an individual’s probability of sarcopenia. Each nomogram incorporated a distinct set of predictors selected by the corresponding algorithm: CatBoost used Age, BMI, Gender, GLIM_Phenotypic criteria, and SARC_Cal_score; LightGBM included Age, BMI, Gender, MNA_SF_score, and SARC_Cal_score; and GBDT utilized Age, Weight, BMI, Osteoporosis status, and SARC_Cal_score. The SARC_Cal_score emerged as a major, consistent contributor across all three nomograms, underscoring its central role in risk assessment. Internal validation via bootstrap calibration showed close agreement between predicted and observed outcomes for the CatBoost and LightGBM nomograms (Figures 6E,G), with calibration curves aligning well with the ideal diagonal. This was quantified by low mean absolute error (MAE) values (0.016 for CatBoost and 0.018 for LightGBM); the GBDT model also demonstrated good calibration (MAE = 0.013) (Figure 6I).

In summary, while all tree-based ensembles exhibited near-perfect discrimination, they substantially outperformed logistic regression in recall and net clinical benefit. The resulting nomograms, validated by excellent calibration (all MAE <0.02), represent clinically applicable instruments for individualized sarcopenia risk assessment.

To validate the interpretability of the selected tree-based ensemble models, multivariable logistic regression models were constructed using the distinct feature sets identified by CatBoost, LightGBM, and GBDT. Their predictor associations and performance were summarized in Figure 7. Forest plots (Figures 7A–C) depicted the odds ratios (ORs) and 95% confidence intervals for key predictors within each model. The SARC_Cal_score was a consistent, highly significant positive predictor across all models (CatBoost: OR = 2.932 (2.096–4.305), p < 0.001; LightGBM: OR = 2.795 (1.899–4.51), p < 0.001; GBDT: OR = 4.475 (2.934–7.334), p < 0.001). Conversely, BMI was a significant protective factor in both the CatBoost- and LightGBM-derived models (OR = 0.525 (0.366–0.713), p < 0.001; OR = 0.305 (0.199-0.427), p < 0.001). The CatBoost-based model also identified age (OR = 1.138 (1.032–1.273), p = 0.015), female (OR = 0.069 (0.024–0.176), p < 0.001). Moreover, the GBDT-based model highlighted age (OR = 1.446 (1.268–1.699), p < 0.001). Regarding discrimination, receiver operating characteristic (ROC) curves for all three logistic models demonstrated excellent performance (Figure 7D). The LightGBM-based model achieved the highest area under the curve (AUC = 0.994), followed by the CatBoost-based (AUC = 0.990) and GBDT-based models (AUC = 0.986). The calibration curves in Figure 7E further demonstrated close agreement between predicted probabilities and observed outcomes, supported by low mean absolute errors (MAE range: 0.023–0.030).

The predictive performance of the logistic regression models derived from different tree-based ensembles was further assessed using net reclassification improvement (NRI) and integrated discrimination improvement (IDI) metrics (Tables 7–9). Statistically significant differences in reclassification accuracy were observed between the model pairs.

As shown in Table 7, the model based on CatBoost-selected features demonstrated a significant improvement over the LightGBM-based model, with an NRI of 1.1004 (95% CI: 0.9785–1.2223; P < 0.0001) and an IDI of 0.041 (95% CI: 0.0235–0.0584; P < 0.0001). Similarly, the CatBoost-based model showed significantly better reclassification than the GBDT-based model (NRI = 0.9938, 95% CI: 0.8661–1.1214; P < 0.0001), although the corresponding IDI did not reach statistical significance (IDI = 0.0057, 95% CI: −0.0151–0.0265; P = 0.591) (Table 8).

In contrast, the comparison between LightGBM- and GBDT-based models yielded mixed results (Table 9). While the LightGBM model showed a positive NRI (0.2175, 95% CI: 0.074–0.361; P = 0.003), the IDI was significantly negative (−0.0353, 95% CI: −0.0594 to −0.0112; P = 0.004), suggesting that the GBDT-based model provided marginally better discrimination in terms of predicted probability separation.

In summary, although all logistic models performed well overall, the CatBoost-based model consistently provided superior reclassification compared with the LightGBM- and GBDT-based models. The comparison between LightGBM and GBDT presented a more nuanced picture, indicating a potential trade-off between reclassification and integrated discrimination.