Abstract
Proximal junctional failure (PJF) is a significant mechanical complication following corrective surgery for adult spinal deformity (ASD), often resulting in structural failure at the uppermost instrumented vertebra and necessitating revision surgery. Early identification of patients at high risk for PJF remains clinically important but challenging due to the multifactorial and nonlinear nature of risk factors. This study aimed to evaluate the predictive performance of five machine learning (ML) models-Random Forest, Logistic Regression, Support Vector Machine (SVM), Decision Tree, and Naive Bayes-in identifying patients at risk for PJF using both preoperative and postoperative spinal alignment parameters. A retrospective cohort of 92 ASD patients who underwent two-stage corrective surgery, including lateral lumbar interbody fusion (LLIF), was analyzed. Radiographic parameters were measured preoperatively and approximately six weeks postoperatively. Six alignment-related features were selected through a combination of univariate statistical testing and Random Forest-based feature importance ranking: preoperative PI-LL, preoperative PT, postoperative PI-LL, postoperative and preoperative TK, and change in lumbar lordosis (ΔLL). Each ML model was trained and tested using five independent stratified 80:20 train-test splits. Among the models, Random Forest achieved the highest mean accuracy (78.4%) and area under the curve (AUC = 0.704). The predicted probabilities for PJF were significantly higher in the PJF group compared to non-PJF cases (0.306 ± 0.181 vs. 0.186 ± 0.164, p = 0.0057). Cross-validation confirmed model robustness (fivefold: 79.4%, tenfold: 77.3%). These findings suggest that Random Forest can serve as a reliable tool for preoperative and early postoperative PJF risk stratification based on alignment correction. Future work should incorporate bone mineral density, comorbidities, and multicenter validation to enhance clinical applicability.