[Predicting the risk of proximal junctional kyphosis after modified grade Ⅳ osteotomy for thoracolumbar old fractures with kyphotic deformity using machine learning].

Objective: To predict the risk of proximal junctional kyphosis (PJK) following modified grade Ⅳ osteotomy for thoracolumbar old fractures with kyphosis based on machine learning (ML). Methods: A retrospective analysis was conducted, including patients with thoracolumbar old fractures and kyphotic deformity who underwent modified grade Ⅳ osteotomy at department of orthopedics, the first affiliated hospital of Soochow University between January 2011 and June 2023. All patients had a minimum follow-up of two years. Patients were divided into the PJK group and the non-PJK group based on the occurrence of PJK, and the clinical indicators were compared between the two groups. After performing multicollinearity diagnosis using Least Absolute Shrinkage and Selection Operator (LASSO) regression, multivariate logistic regression was conducted to identify risk factors for PJK. Five machine learning models including optimized XGBoost, GradientBoosting, Random Forest, linear discriminant analysis (LDA) and model driven architecture (MDA) were constructed and trained, with their predictive performances evaluated subsequently. The Shapley Additive Explanations (SHAP) method was employed for model interpretation and feature importance ranking. Results: The study included 128 patients (47 men, 81 women), with a mean age of (68.8±7.2) years.The local kyphosis angle improved from 32.65°±8.17° preoperatively to 2.83°±1.99° at 2 weeks postoperatively in the non-PJK group, and from 32.97°± 8.51° preoperatively to 3.01°±1.72° at 2 weeks postoperatively in the PJK group (both P<0.01). Significant differences were found between the non-PJK and PJK groups in body mass index (BMI), bone mineral density (BMD), upper instrumented vertebra-cemented screw (UIV-CS), relative sagittal alignment (RSA), relative functional cross-sectional area of the paraspinal muscles (rFCSA), and functional muscle-fat index (FMFI) (all P<0.05). LASSO regression identified moderate multicollinearity between rFCSA (VIF=5.0) and FMFI (VIF=4.96); therefore, the variant FMFI was excluded. Multivariate logistic regression was then performed including BMI, UIV-CS, BMD, RSA, and rFCSA. The results indicated that smaller rFCSA (OR=0.990, 95%CI: 0.981-0.999), lower BMD (OR=0.680, 95%CI: 0.471-0.983), and higher BMI (OR=1.206, 95%CI: 1.036-1.403) were independent risk factors of PJK. Based on PJK risk variables, five ML models were constructed. In the test set, the optimized XGBoost model achieved the highest area under the receiver operating characteristic curve (AUC) of 0.841 (95%CI: 0.675-0.972), followed by RandomForest [0.813 (95%CI: 0.632-0.941)], LDA[0.786 (95%CI: 0.613-0.956)], GradientBoosting[0.782 (95%CI: 0.601-0.967)] and MDA[0.724 (95%CI: 0.514-0.896)]. The XGBoost model demonstrated high specificity (96.4%), positive predictive value (83.3%), and accuracy (82.05%). SHAP visualization of the XGBoost model revealed that the importance of features for predicting PJK risk was in the following descending order: rFCSA (42.1%)、BMD (33.6%)、BMI (24.3%). Conclusions: The ML-based XGBoost model can effectively predict the risk of PJK following modified grade IV osteotomy for old thoracolumbar fractures with kyphosis. Within the context of this disease model, degeneration of the paraspinal muscles, bone mineral density and BMI were identified as significant risk factors of postoperative PJK.