Abstract
BACKGROUND: This study utilized artificial intelligence (AI)-based machine learning algorithms, alongside the shapley additive explanations (SHAP) framework, to identify lower-limb muscle force patterns associated with recurrent patellofemoral pain (PFP) in the anterior and posterior patellar (APP), medial border of the patella (MBP), and lateral border of the patella (LBP) regions. The goal was to inform region-specific strength training strategies. METHODS: A total of 299 patients with prior PFP underwent baseline biomechanical assessments, during which lower-limb and trunk muscle forces were estimated using OpenSim modeling. Participants were then prospectively followed for six months and categorized into pain-free, APP, MBP, or LBP groups according to PFP recurrence and pain location. Machine learning models were subsequently applied in conjunction with the SHAP framework to identify region-specific associations between muscle force patterns and PFP incidence. RESULTS: APP recurrence was linked to gracilis force < 0.055 N/kg, adductor longus force > 0.110 N/kg, tibialis anterior force < 0.678 N/kg, tensor fasciae latae force > 0.144 N/kg, and internal oblique force < 0.699 N/kg. MBP recurrence was associated with rectus femoris force > 0.800 N/kg, gracilis force > 0.054 N/kg, gluteus maximus force > 0.379 N/kg, adductor longus force > 0.711 N/kg, and semitendinosus force < 0.037 N/kg. LBP recurrence corresponded to rectus femoris force < 0.530 N/kg, adductor longus force > 0.194 N/kg, tensor fasciae latae force < 0.082 N/kg, gracilis force > 0.040 N/kg, and gluteus maximus force < 0.151 N/kg. CONCLUSIONS: Machine learning analyses revealed region-specific muscle force patterns predictive of PFP recurrence, offering a biomechanical foundation for targeted strength interventions in APP, MBP, and LBP cases.