Predicting Proximal Junctional Kyphosis after L3, L4, and L5 Pedicle Subtraction Osteotomies: A Machine Learning-Driven CatBoost Model Analysis

Introduction: Proximal junctional kyphosis (PJK) remains a prominent complication in adult spinal deformity (ASD) surgeries. This study investigates PJK rates and predictive factors in 204 ASD patients undergoing pedicle subtraction osteotomy (PSO) at varying lumbar levels (L3, L4, and L5), hypothesizing that lower lumbar PSO levels may reduce PJK risk by improving sagittal balance.

Methods: We retrospectively analyzed 204 ASD patients who underwent PSO for fixed sagittal malalignment, with a minimum 2-year follow-up. Patients were grouped by PSO level (L3, L4, or L5). Radiographic parameters—lumbar lordosis (LL), pelvic incidence (PI), pelvic tilt (PT), and sagittal vertical axis (SVA) were collected preoperatively, postoperatively, and at final follow-up. Univariate analyses compared baseline factors between PJK and non-PJK groups, while machine learning models, including a CatBoost classifier, were developed to predict long-term PJK risk. The model was optimized using SHAP values for enhanced interpretability of predictive features.

Results: Among 204 patients (mean age 59 ± 10 years), PSO was performed at L3 (n=60), L4 (n=72), or L5 (n=72). PJK incidence was 5.9%, with a higher occurrence in L3 (6.7%) and L4 (7.6%) groups than in the L5 group (0%). Radiographic analysis showed significant improvements from preoperative to final follow-up in LL (mean increase of 20°), SVA (average reduction of 8 cm), and PT (average reduction of 6°) (all p< 0.05). Univariate analysis identified older age, higher preoperative PT, and greater PI-LL mismatch as significant predictors of PJK. The CatBoost classifier achieved a high accuracy (AUC 0.87) in predicting PJK, with SHAP analysis emphasizing preoperative PT, PI-LL mismatch, and sagittal malalignment as key predictors.

Conclusion : L5-level PSO may help minimize PJK risk by optimizing lumbar lordosis distribution and reducing stress on proximal segments. Machine learning-based predictive modeling, as demonstrated, enables effective risk stratification and enhances surgical planning. This approach holds promise for personalized treatment strategies in ASD.