Application of Artificial Intelligence and Surface Electromyography in the Assessment of Quadriceps Neuromuscular Fatigue and Functional Risk of Knee Injury in Professional Basketball Players: A Pilot Study

Background: Neuromuscular fatigue reduces a muscle’s capacity to generate force and is a key determinant of performance and injury risk. In basketball, the quadriceps femoris (QF) is central to jumping, acceleration, and changes in direction, so objective detection of QF fatigue is clinically valuable. Objective: To develop and validate machine-learning models for detecting QF neuromuscular fatigue in professional basketball players from surface electromyography (sEMG), and to relate fatigue-induced changes to knee biomechanics. Methods: In this pilot study, eight male athletes (n = 8; basketball; 18–40 years) performed standardized isometric and isotonic 30-repetition knee-extension fatigue protocols (isometric contraction at 60% of maximal voluntary contraction [MVC]; isotonic 30-repetition set against a fixed submaximal load of 60% of MVC), performed on both legs (16 limbs). sEMG from vastus lateralis (VL) and vastus medialis (VM) was recorded per SENIAM guidelines at baseline, during, and post-fatigue. Time-domain (RMS), frequency-domain (MDF, MNF), and non-linear features (approximate entropy, sample entropy, recurrence quantification analysis), plus the VL/VM ratio, were extracted; logistic regression (baseline), Support Vector Machine, Random Forest, and CNN–LSTM classifiers were trained; the three classical models were evaluated with participant-level (leave-one-subject-out) nested cross-validation in the six athletes with separate VL and VM recordings, whereas the CNN–LSTM was evaluated under within-subject validation only; knee loading was examined with finite-element (FEM) analysis in one representative athlete. Results: Fatigue reduced MDF and MNF and increased RMS in both muscles (p < 0.001), increased signal determinism, and shifted the VL/VM ratio (0.97 → 1.14, p < 0.01). Across 57 sEMG trials (eight athletes, 16 limbs), RMS increased in 50 and MNF decreased in 37, with the complete fatigue signature (rising RMS with falling MNF) present in 32 (56%). Under within-subject validation, the CNN–LSTM model reached an accuracy of 0.93 (AUC 0.97), but it was not re-evaluated under participant-level (leave-one-subject-out) nested cross-validation, so its ability to generalize across athletes is unknown and no head-to-head comparison with the classical models is claimed. Under participant-level nested cross-validation, carried out in the six athletes with separate VL and VM recordings, the classical models reached only modest performance (accuracy 0.68–0.70, AUC 0.73–0.74, with per-fold accuracy ranging from 0.52 to 0.82); in a single representative subject-specific model, FEM indicated increased localized cartilage stress under fatigue. Conclusions: In this pilot study, sEMG captured the canonical fatigue signature in just over half of the trials, and under participant-level nested cross-validation in six athletes the classical classifiers generalized only modestly (accuracy 0.68–0.70, AUC 0.73–0.74), well below the within-subject CNN–LSTM estimate. No injury outcomes were recorded, so the sEMG and finite-element markers cannot yet be used to estimate knee-injury risk; these preliminary findings require confirmation in a larger, adequately powered cohort.