Work

Multilayer Perceptron versus Direct Guideline Rule Application for Four-Class Sarcopenia Classification: A Development and External Validation Study

Year: 2026
Type: preprint
Abstract: Background Sarcopenia phenotypes are defined by cut-off values for muscle mass, grip strength, and physical performance. When reference labels are generated from the same variables assigned to a model, the aspects contributed by a machine learning classifier to the direct application of cut-off values remain unclear. Methods We developed a multilayer perceptron (MLP; 5–36–28–9–4) to classify the states of “normal,” “sarcopenia,” “functional sarcopenia,” and “severe sarcopenia” based on five variables: sex, height, skeletal muscle index, grip strength (higher of the left- and right-hand means), and Short Physical Performance Battery score. The development cohort comprised 339 participants from one hospital (data collection: September 2023–August 2025) and was evaluated using stratified five-fold cross-validation (out-of-fold predictions). An external cohort of 151 participants from three institutions (148 women) was used for validation. The MLP was compared with the direct application of guideline rules (Rules A and B), logistic regression, a decision tree, and a random forest. Confidence intervals (CIs) were obtained through bootstrap resampling. Results Reference labels agreed with Rule A in 93.5% (317/339) of the development cohort and 94.0% (142/151) of the external cohort and with the originally stated definition (Rule B) in 91.7% and 92.1%, respectively. The MLP achieved out-of-fold accuracy, balanced accuracy, and macro F1-score values of 0.870, 0.819, and 0.819, respectively, in the development cohort and 0.901, 0.923, and 0.897, respectively, in the external cohort. Rule A, applied without training, achieved 0.935 (95% CI 0.909–0.959), 0.926, and 0.909 internally and 0.940 (95% CI 0.901–0.974), 0.939, and 0.926 externally, higher than the MLP for every point estimate. The random forest matched Rule A externally (accuracy 0.940) but not internally (0.882); logistic regression and the decision tree performed worse. Conclusions The MLP reproduced the reference labels but did not outperform the direct application of the guideline rules. When labels are defined by the cut-off values of the input variables, the model performance largely reflects rule reproduction, and any added value must be tested against independent reference standards or clinical outcomes. As 98% of the external cohort comprised women, generalizability to men could not be assessed.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7236455851
Cites: 28
Cited by:
Related to: 10