Risk Factors and an Interpretable Machine Learning Model for Predicting Spinal Epidural Lipomatosis: A Multicenter Study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Spinal Orthopedics, General Hospital of Ningxia Medical University, Ningxia Hui Autonomous Region, Yinchuan City, China.
- The First Clinical College of Ningxia Medical University, Ningxia Hui Autonomous Region, Yinchuan City, China.
- Department of Orthopedics, Jinchang People's Hospital, Gansu, China.
- Department of Orthopedics, People's Hospital of Ningxia Hui Autonomous Region, Ningxia Hui Autonomous Region, Yinchuan City, China.
- Department of Orthopedics, The Third People's Hospital of Yinchuan City, Ningxia Hui Autonomous Region, Yinchuan City, China.
- Department of Spinal Orthopedics, Qinghai University Affiliated Hospital, Xining City, Qinghai Province, China.
Abstract
A retrospective multicenter study. To identify independent risk factors for spinal epidural lipomatosis (SEL) and to develop and validate an interpretable machine learning-based predictive model. SEL is an underdiagnosed yet clinically significant cause of debilitating lumbar spinal stenosis. Robust tools for early identification and risk stratification of at-risk patients are currently lacking. Using data from 774 patients with low back and leg pain who underwent lumbar MRI at five institutions, we applied LASSO regression for variable selection and developed a clinically accessible nomogram. The cohort was randomly divided into training (70%) and validation (30%) sets. Four machine learning models were constructed and evaluated based on discrimination (AUC), calibration, and clinical utility (decision curve analysis). Seven independent predictors were identified: elevated random blood glucose, blood type B, atherosclerosis index, body mass index, uric acid, obstructive sleep apnea, and age. The XGBoost model demonstrated superior predictive performance in the validation set (AUC: 0.726; 95% CI: 0.547-0.904), with satisfactory calibration and positive net clinical benefit. Interpretability analysis confirmed glucose, age, and uric acid as the most consistent contributors to individualized risk predictions. We developed and validated an interpretable prediction model that integrates clinical risk factors with an XGBoost algorithm and provides an actionable nomogram. This tool demonstrates strong potential to assist clinicians in early SEL detection and risk-stratified management, potentially enabling more targeted interventions for this underdiagnosed condition.