心理因素在ACL重建后继发损伤风险的多域机器学习模型中增量预测价值最大
简介
该研究纳入487例ACL重建患者，在术后6个月标准化随访时采集人口学、MRI、步态分析、等速肌力及心理5个领域的30个预测变量，构建机器学习模型预测二次损伤（同侧移植物再断裂或对侧ACL断裂）风险。关键发现：64例（13.1%）发生二次损伤；逻辑回归模型表现最优（AUC=0.739，Brier评分=0.106）；SHAP分析显示TSK评分（运动恐惧症）和PH…
英文摘要
BACKGROUND: Secondary injury after anterior cruciate ligament (ACL) reconstruction, defined as ipsilateral graft rerupture or contralateral ACL rupture, remains a clinical challenge. Current prediction models predominantly fail to capture this multifactorial risk. In this study, we developed a multi-domain machine learning model to predict the risk of secondary injury. METHODS: This retrospective cohort study included 487 patients who underwent primary ACL reconstruction. Thirty predictor variables spanning demographic, magnetic resonance imaging (MRI), gait analysis, isokinetic strength, and psychological domains were collected at a standardized 6-month postoperative follow-up. Five machine learning algorithms were evaluated using a nested cross-validation scheme, with SHAP analysis and domain ablation applied for interpretability. RESULTS: Sixty-four patients (13.1%) sustained secondary injuries (ipsilateral graft re-rupture or contralateral ACL rupture). Logistic regression achieved the best discriminative performance (AUC = 0.739, 95% CI: 0.672-0.806) and calibration (Brier score = 0.106), although no statistically significant difference was observed between logistic regression and random forest (corrected paired t-test, p = 0.72). SHAP analysis of both the random forest and logistic regression models identified the TSK score and PHQ-9 as the most influential individual predictors; bootstrap resampling indicated moderate stability of these rankings. Domain ablation confirmed that the psychological domain provided the largest incremental predictive contribution, with its removal producing the greatest performance decrement (AUC decline: 0.031, 95% CI: 0.007-0.058), whereas the removal of the MRI or gait domains did not reduce model performance. All models demonstrated high negative predictive values (0.89-0.94) but limited precision (0.21-0.26), indicating that most patients flagged as high risk would not sustain a secondary injury. Decision curve analysis indicated a net clinical benefit in the 0.05-0.15 threshold range, supporting a low-threshold screening rather than a diagnostic application. CONCLUSION: A multi-domain machine learning model identified patients at elevated secondary injury risk with acceptable discrimination and calibration. Kinesiophobia and depressive symptoms showed the largest incremental predictive contributions among the domains examined, suggesting that systematic psychological screening within postoperative rehabilitation warrants further investigation in prospective and externally validated studies.