Classification of vertical jump performance categories in futsal using machine learning algorithms.

Background: Futsal requires high-intensity explosive actions, including vertical jumps. Force-time analysis from dual force plates provides multidimensional biomechanical data that traditional statistical methods struggle to analyze effectively. Objective: To develop and evaluate a supervised classification framework capable of reproducing internally-derived vertical jump performance categories in futsal athletes using independent force-time metrics from dual force plates. Methods: Fifty-one male athletes performed countermovement jumps (CMJ), squat jumps (SJ), and drop jumps (DJ) on VALD ForceDecks dual force plates (1,000 Hz), yielding 148 valid observations after exclusion of records with missing PCA-input variables. Four internally derived performance categories were constructed using Principal Component Analysis of six biomechanical variables. The first two retained components explained 75.6% of the total variance. We compared four machine learning algorithms using an approximate 75/25 athlete-level stratified group split and stratified 5-fold cross-validation with group constraints to prevent leakage from repeated jump records from the same athlete. Robustness was assessed through reduced-predictor sensitivity analysis, jump-test ablation, learning curves, calibration, and nested cross-validation. Results: Logistic Regression achieved strong performance on the independent test set (F1-Score = 0.830, AUC-ROC = 0.977). Grouped 5-fold cross-validation yielded more conservative estimates (F1 = 0.770 ± 0.069, AUC = 0.941 ± 0.039), reflecting a more realistic estimate of internal generalization to unseen athletes. Coefficient-based and permutation importance analyses consistently identified CMJ peak power as the most influential predictor, while DJ-derived variables, particularly eccentric mean force, concentric impulse, and peak power, also showed substantial discriminative relevance. Ablation analysis confirmed the critical role of drop-jump variables: removing them reduced F1-score from 0.830 to 0.604, corresponding to an absolute decrease of 0.226 and a relative reduction of approximately 27%. Conclusions: Machine learning algorithms, particularly Logistic Regression, supported accurate classification of internally derived PCA-based vertical jump performance categories in futsal athletes. The combination of PCA-based category construction and supervised classification, with rigorous athlete-level validation and comprehensive robustness diagnostics, provides a reproducible methodological framework that may support talent identification, individualized training, and longitudinal monitoring after external validation.