Multidomain Physical Performance Profiling Discriminates Amateur from Professional Soccer Players: An Explainable Machine Learning Analysis

The physical qualities that separate amateur from professional soccer players are usually studied one domain at a time. This study examined a multidomain physical performance battery with explainable machine learning. Cross-sectional data from 201 male players (101 amateur, 100 professional) were reduced from 45 variables to 24 curated features spanning anthropometry/training history, isokinetic strength, vertical jump, Wingate anaerobic power, reactive agility and sprint. Five classifiers were tuned with nested cross-validation and examined with resampling, permutation, SHapley Additive exPlanations and domain-restricted analyses. All variables except the hamstring:quadriceps ratios, countermovement–squat jump difference and stretch-shortening-cycle contribution differed between groups after false-discovery-rate correction (Cohen’s d = 0.49–2.10). All classifiers achieved near-ceiling discrimination (test area under the receiver operating characteristic curve [AUC] 0.998–1.000; nested cross-validated AUC 0.994–0.999; label-permutation test of the logistic-regression pipeline p = 0.001), and the unsupervised composite index reached AUC = 0.969. Training age, jump height, 30 m sprint time and reactive agility contributed most, whereas body-mass-normalised isokinetic torque discriminated least (AUC 0.77–0.80); jump-plus-Wingate and sprint-plus-agility subsets alone reached AUC 0.95–0.98. Discriminatory information was distributed across domains rather than concentrated in isokinetic strength. These cross-sectional, cohort-specific findings describe established groups and do not support individual talent selection without external validation.