Artificial intelligence in the biomechanics of knee osteoarthritis: biomechanical phenotyping, risk stratification, diagnosis, treatment and clinical translation-a narrative review.

Structural degeneration, pain and functional limitation in knee osteoarthritis (KOA) often diverge, which suggests that radiographic grading alone cannot explain the heterogeneous clinical course of the disease. Lower-limb alignment, gait strategy, muscle function, joint contact forces and real-world activity exposure jointly shape KOA onset, progression and treatment response. Artificial intelligence (AI) methods provide tools for integrating multimodal imaging, gait, wearable-sensor and clinical data, and can be combined with finite element analysis (FEA) and other computational biomechanical models. Conceptually, we propose an evidence-to-action framework that distinguishes AI-supported measurement, longitudinal prediction and intervention or decision utility, while separating AI-native methods from non-AI digital and computational biomechanical technologies. This structured narrative review synthesizes and critically interprets evidence on artificial intelligence in knee osteoarthritis from a biomechanical perspective, with an emphasis on biomechanical phenotyping across imaging, gait, wearable-sensor and clinical data, together with longitudinal risk stratification, diagnosis, progression prediction, rehabilitation feedback, perioperative management and clinical translation. Available studies indicate that artificial intelligence can support cartilage and meniscus segmentation, imaging-based grading, biomechanical load estimation, preoperative planning, and rehabilitation monitoring, although the strength and clinical maturity of the evidence vary substantially across these applications. Multimodal AI models have identified imaging, biomechanical, sensor-derived, and clinical features associated with structural progression, pain, and functional outcomes, but these associations should not be interpreted as causal or clinically actionable without longitudinal, external, and prospective validation. However, limited external validation, heterogeneous data acquisition, label bias, limited causal interpretability, limited use of patient-centered outcomes and poor workflow integration continue to constrain clinical use. Future AI-enabled KOA assessment should be grounded in biomechanical mechanisms and oriented toward patient outcomes and actionable clinical decisions. Translation will require standardized multimodal data, prospective validation and real-world utility evaluation.