可解释机器学习识别关节镜下半月板后根部撕裂的膝关节形态阈值：一项回顾性队列研究
简介
本研究回顾性分析了608例膝关节镜手术患者（281例经关节镜证实的内侧半月板后根撕裂，327例对照），基于MRI形态学参数构建10种机器学习模型。关键发现：多因素分析显示年龄增大、内侧胫骨后倾角增大、内侧胫骨平台深度加深是MMPRT的独立危险因素；GBM模型在独立测试集中AUC最高，SHAP分析显示年龄贡献最大，其次为内侧胫骨后倾角和内侧胫骨平台深度。依赖图…
英文摘要
BACKGROUND: Medial meniscus posterior root tear (MMPRT) is clinically important because disruption of posterior root function compromises hoop tension and load sharing, accelerating medial-compartment degeneration. Early or subtle posterior pathology may be underrecognized on MRI, highlighting the potential value of morphology-based risk awareness during MRI interpretation. METHODS: We retrospectively analyzed 608 patients who underwent arthroscopic surgery for knee joint injuries, including 281 patients with arthroscopically confirmed MMPRT and 327 controls without MMPRT. Demographic, clinical, and MRI-based morphologic parameters were compared in the training set. Variables identified from training-set comparisons and clinical/biomechanical relevance were used to develop 10 machine learning models, including CatBoost, Decision Tree, GBM, LightGBM, LASSO, Naive Bayes, Neural Network, Random Forest, Support Vector Machine, and XGBoost. Models were evaluated with 10-fold cross-validation and an independent testing set. Explainability was assessed using SHapley Additive exPlanations (SHAP), including global importance and dependence plots. RESULTS: Training-set multivariable analysis identified older age, greater medial tibial slope (MTS), and deeper medial tibial plateau depth (MTPD) as independent factors associated with MMPRT. GBM achieved the highest AUC in the independent testing set and was selected for SHAP-based interpretation. SHAP analysis ranked age as the dominant contributor, followed by MTS and MTPD. Dependence plots suggested non-linear, threshold-like patterns in model contribution. Age showed an apparent transition from negative to positive SHAP contributions around midlife, MTS showed a threshold-like increase between approximately 6° and 8°, whereas MTPD shifted toward positive SHAP contributions around approximately 2.2-2.5 mm. CONCLUSION: Age was the dominant model contributor, and MTS and MTPD were independently associated with MMPRT, exhibiting non-linear patterns in SHAP-based model interpretation. These findings may help raise suspicion for MMPRT in symptomatic patients undergoing MRI, particularly when age-related and tibial plateau morphologic risk patterns are present.