Development of a visualizable machine learning model for mechanical complication risk in adult spinal deformity surgery
10.3760/cma.j.cn121113-20250122-00064
- VernacularTitle:基于机器学习算法构建可视化模型预测成人脊柱畸形矫形术后力学并发症的发生风险
- Author:
Jie LI
1
;
Zhen TIAN
1
;
Zhong HE
1
;
Xiaodong QIN
1
;
Jun QIAO
1
;
Saihu MAO
1
;
Benlong SHI
1
;
Yong QIU
1
;
Zezhang ZHU
1
;
Zhen LIU
1
Author Information
1. 南京大学医学院附属鼓楼医院骨科脊柱外科，南京 210008
- Publication Type:Journal Article
- Keywords:
Adult;
Spinal curvatures;
Postoperative complications;
Machine learning
- From:
Chinese Journal of Orthopaedics
2025;45(17):1137-1146
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To predict mechanical complications (MC) following spinal deformity surgery for adult spine deformity (ASD) using machine learning models, identify key risk factors, and develop a visualizable tool for individualized risk assessment.Methods:Clinical and radiological data from 525 patients with ASD who underwent surgery in our hospital between January 2017 and December 2021 were collected. Patients were randomly assigned to a training set (70%) and a test set (30%) for model development. The cohort included 88 males and 437 females, with a mean age of 42.2±18.1 years. Variables included demographic data, comorbidities, local and systemic radiological parameters, paraspinal muscle fat infiltration (FI), and vertebral bone quality (VBQ) scores. Multiple machine learning algorithms: Random Forest (RF), Gaussian Naive Bayes (GNB), Light GBM, Support Vector Machine (SVM), XGBoost (XGB), and Logistic Regression (LR) were trained and evaluated. Model performance was compared using the receiver operating characteristic curve (ROC) and precision-recall curve (PRC). SHAP (Shapley Additive Explanations) was used to rank risk factors, while LIME (Local Interpretable Model-Agnostic Explanations) was applied to visualize MC risk in individual cases.Results:Of the 525 patients, 135 (25.7%) developed postoperative MC. Among these, 80 (59.3%) experienced proximal junction kyphosis or failure (PJK/PJF), 7 (5.2%) had distal junction kyphosis or failure (DJK/DJF), 28 (20.7%) sustained rod fractures, and 29 (21.5%) showed significant loss of correction. In the validation cohort, the RF model achieved the highest area under the curve (AUC=0.80), followed by GNB (0.77), XGB (0.76), LR (0.74), LightGBM (0.73), and SVM (0.66). The RF model also demonstrated the best PRC value (0.58), highest sensitivity (0.65), and lowest Brier score (0.20). GNB, Light GBM, and LR models achieved the highest accuracy (0.78 each), while LightGBM exhibited the highest specificity (0.93). SHAP analysis identified higher preoperative VBQ scores, larger T 1 pelvic angle (TPA), and higher paraspinal muscle FI as the main risk factors for MC. Based on the RF model, a LIME-based tool was successfully constructed for individualized MC risk estimation. Conclusion:The RF model demonstrated the best overall predictive performance for MC. A machine learning-based prediction model has the potential to provide valuable guidance for surgical decision-making in ASD patients.
