[A machine learning-based model for predicting postoperative coronal imbalance in degenerative scoliosis]

Objective: To develop a machine learning-based prediction model for postoperative coronal imbalance (CIB) in patients with degenerative scoliosis (DS). Methods: The data of patients with DS who underwent corrective surgery at Xuanwu Hospital, Capital Medical University, between January 2018 and January 2023 were retrospectively collected. Based on the coronal alignment at the final postoperative follow-up, patients were categorized into a coronal balance (CB) group and a CIB group (defined as coronal vertical axis>3 cm). A total of 26 preoperative and intraoperative variables were considered as candidate predictors. Feature selection was performed using three machine learning algorithms, including least absolute shrinkage and selection operator (LASSO) regression, recursive feature elimination (RFE), and Boruta algorithm, and the intersection of the selected features was used as the final input variables. The dataset was randomly divided into a training set and a test set at a 7∶3 ratio using stratified sampling. Five machine learning algorithms, including logistic regression, random forest, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and multilayer perceptron, were employed to construct predictive models for postoperative CIB. Model performance was evaluated using six metrics: area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The SHapley Additive exPlanations (SHAP) method was used to assess the contribution and importance of each predictor. Results: A total of 168 patients aged (64.6±8.0) years were included in the study, with 49 males and 119 females. Among them, 115 patients achieved postoperative coronal balance (CB group), whereas 53 developed postoperative coronal imbalance (CIB group). Compared with the CB group, patients in the CIB group had a significantly higher prevalence of preoperative frailty [39.6% (21/53) vs 14.8% (17/115), P<0.001], were more likely to undergo a posterior-only surgical approach [81.1% (43/53) vs 57.4% (66/115), P=0.003], and had a higher proportion of type Ⅱ DS [71.7% (38/53) vs 41.7% (48/115), P<0.001]. In addition, the CIB group exhibited a significantly greater correction ratio between the two curves (1.3±0.6 vs 0.8±0.5, P<0.001), a larger preoperative lumbosacral Cobb angle (22.3°±5.8° vs 18.4°±6.4°, P<0.001), and greater preoperative coronal displacement [(3.3±1.0) cm vs (2.5±1.4) cm, P<0.001]. Five key predictors were ultimately selected for model development: correction ratio between the two curves, preoperative coronal displacement, DS classification, frailty, and surgical approach. Among the five machine learning models evaluated, the random forest model demonstrated the best overall performance, achieving an AUC of 0.857 (95%CI: 0.737-0.977), with an accuracy of 0.814 (95%CI: 0.674-0.903), sensitivity of 0.643 (95%CI: 0.388-0.837), specificity of 0.897 (95%CI: 0.736-0.964), positive predictive value of 0.750 (95%CI: 0.468-0.911), negative predictive value of 0.839 (95%CI: 0.674-0.929), and a Brier score of 0.152. The calibration curve showed good agreement between predicted and observed probabilities, while decision curve analysis demonstrated a positive net clinical benefit across a wide range of threshold probabilities (10%-85%). SHAP analysis revealed that the most influential predictors of postoperative CIB, ranked in descending order of importance, were a higher correction ratio between the two curves (0.157), greater preoperative coronal displacement (0.104), type Ⅱ DS (0.073), preoperative frailty (0.046), and a posterior-only surgical approach (0.042). Conclusion: The machine learning-based prediction model demonstrated good predictive performance for postoperative CIB in patients with DS, offering valuable insights to support the optimization of surgical strategies.