Abstract
BACKGROUND: High tibial osteotomy (HTO) is a key joint-preserving procedure for medial compartment osteoarthritis and varus deformity, but conventional planning is operator-dependent. Machine learning (ML) offers potential to automate radiographic assessment and predict surgical outcomes. This systematic review evaluated the accuracy, efficiency, and generalizability of ML models applied to HTO planning and prediction. METHODS: Following PRISMA 2020 guidelines (PROSPERO CRD420251122187), PubMed, EMBASE, and Web of Science were searched to August 2025. Studies using ML for HTO planning, alignment measurement, or outcome prediction were included. Two reviewers independently extracted data and assessed risk of bias using the PROBAST tool across four domains: participants, predictors, outcomes, and analysis. RESULTS: From 43 retrieved records, 11 studies met inclusion criteria. Most were retrospective, single-center analyses with heterogeneous datasets. Convolutional neural networks and deep learning systems achieved sub-2° mean absolute error for alignment parameters such as hip-knee-ankle, medial proximal tibial and lateral distal femoral angles. Ensemble ML models predicted lateral hinge-fracture risk and postoperative alignment with area-under-curve values exceeding 0.80. Intraclass correlation coefficients for AI-derived measures were consistently >0.90. Automated analysis was markedly faster than manual measurement-0.2 s versus 1-2 min per radiograph. Only a single study performed multicenter external validation; others remained internally tested. CONCLUSION: ML demonstrates excellent precision and efficiency in radiographic analysis and complication prediction for HTO. Nonetheless, evidence is constrained by single-center data, small cohorts, and lack of functional validation. Future multicentric, prospective, and explainable AI studies are required to confirm clinical applicability and patient-reported benefit.