Enhancing fear of re-injury classification after ACL reconstruction by integrating biomechanical and electromyography data using multimodal machine learning methods

• Multimodal EMG + biomechanics improved fear classification after ACL reconstruction compared to either modality in isolation. • Best XGBoost model reached 86% accuracy with hop-level and participant-level metrics. • Explainable AI identified trunk tilt, pelvic obliquity, and hamstring–quadriceps balance as the most discriminative features distinguishing HIGH-FEAR from LOW-FEAR participants. • Dual validation (LOPOCV + GroupKFold) ensured methodological rigor and transparency. • Study reported in line with TRIPOD-AI guidelines for clinical ML prediction models. Fear of re-injury after anterior cruciate ligament (ACL) rupture often hinders return-to-sport and has been linked to movement patterns associated with increased injury risk. Identifying fearful individuals is thus critical. Self-reported questionnaires are commonly used but may be affected by underreporting in sporting and clinical contexts. We aimed to classify individuals with ACL reconstruction (ACLR) into HIGH-FEAR and LOW-FEAR groups, using machine learning (ML) models trained on unimodal or multimodal time-series movement data from side hop landings. Seventy-two participants (median: 13.0 months post-ACLR; interquartile range: 15.8 months) were dichotomized into HIGH-FEAR/LOW-FEAR groups using a single item (statement 9) from the Tampa Scale for Kinesiophobia. Participants performed one-leg standardized rebound side hops while kinematics and kinetics were recorded using 3D motion capture and force plates, respectively, and electromyography (EMG) was registered from flexor and extensor thigh muscles. We extracted time-series features from each modality and ten ML algorithms were trained and evaluated using leave-one-participant-out and grouped 3-fold cross-validation. Integrating kinematic/kinetic and EMG data improved classification accuracy compared to single modality datasets. The extreme gradient boosting model achieved the highest accuracy for fused data (86%) using the top 40 ranked features, including trunk tilt, pelvic obliquity, knee rotation, and flexor and extensor muscle activations. Kinematic/kinetic data alone achieved 83% accuracy per participant, while EMG data alone yielded 85%. This study demonstrates the potential of integrating different movement-related data to enhance the accuracy of ML models in classifying fear of re-injury post-ACLR, supporting identification of patterns associated with fear and guiding treatment.