Spine Journal - 2026-06-17 - Journal Article
Machine Learning-Based Identification of Distinct Risk Factors for Moderate vs Severe Proximal Junctional Kyphosis After Adult Spinal Deformity Surgery.
Kang DH, Park JS, Lee CS, Park SJ
Topics
Key Takeaway
Machine learning identified distinct risk factor profiles for moderate PJK (17.4%; driven by L1 tilt and relative hyperlordosis) versus severe PJK (11.0%; driven by lordosis distribution index maldistribution and construct rigidity) after ASD surgery.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study asked whether moderate (PJA ≥15°) and severe (PJA ≥28°) PJK after ASD surgery share the same risk factors or represent mechanistically distinct entities. Five ML algorithms (Logistic Regression, SVM, RF, XGBoost, AutoGluon) with SHAP analysis were applied to 374 patients; feature stability analysis identified consensus predictors across models. Moderate PJK (17.4%) was predicted by BMI, L1 tilt, and relative hyperlordosis, with iliac screws protective; severe PJK (11.0%) was predicted by maldistributed lordosis distribution index, excessive postoperative LL, high cement volume, and number of rods.
Key Limitation
The absence of an external validation cohort means the ML models' predictive accuracy (sensitivity, specificity, AUC) cannot be confirmed in independent patient populations, limiting clinical deployment of these algorithms.
Original Abstract
BACKGROUND CONTEXT
Proximal junctional kyphosis (PJK) is a well-recognized complication of adult spinal deformity (ASD) surgery. However, traditional proximal junctional angle (PJA) thresholds (15°) and revision-predictive thresholds (28°) may represent distinct clinical entities with unique etiologies.
PURPOSE
To analyze and compare predictive risk factors for moderate (PJA ≥15°) and severe (PJA ≥28°) PJK using machine learning.
STUDY DESIGN/SETTING
Retrospective study
PATIENT SAMPLE
A total of 374 patients who underwent ASD surgery with a minimum 2-year follow-up.
OUTCOME MEASURES
Development of moderate (PJA ≥ 15°) and severe (PJA ≥ 28°) PJK.
METHODS
Five machine learning algorithms (Logistic Regression, SVM, RF, XGBoost, AutoGluon) were trained to predict moderate and severe PJK. Feature stability analysis identified robust predictors across models, and SHAP analysis elucidated the feature directionality.
RESULTS
The incidence of moderate and severe PJK was 17.4% (65 patients) and 11.0% (41 patients), respectively. BMI and L1 tilt were universally selected for moderate PJK. SHAP analysis showed that high RLL (relative hyperlordosis) and L1 tilt were associated with increased predicted risk, whereas iliac screws were protective. The maldistributed LDI and number of rods were consensus predictors of severe PJK. SHAP associated maldistributed LDI, excessive postoperative LL, and high cement volume with an increased predicted risk of severe PJK.
CONCLUSIONS
Moderate PJK appears to be driven by geometric stress concentration (e.g., L1 tilt and relative hyperlordosis), whereas severe PJK stems from structural/distributional mismatch (e.g., lordosis maldistribution and construct rigidity). Prevention strategies should be stratified according to these distinct mechanisms.