Predicting Progression from Osteoporotic Vertebral Compression Fracture to Kümmell Disease: A Two-Centre Machine-Learning Study with External Evaluation.
Authors
Affiliations (5)
Affiliations (5)
- Department of Orthopedics, Shanghai Tenth People's Hospital, School of Medicine, Tongji University, Shanghai 200072, China.
- Department of Orthopaedics, The First Affiliated Hospital of Soochow University, Soochow University, Suzhou 215006, China.
- Department of Orthopaedics, The Second Affiliated Hospital of Soochow University, Soochow University, Suzhou 215004, China.
- Department of Pain Medicine, Xuzhou Renci Hospital, Xuzhou, 221000, Jiangsu, China.
- Department of Computer Science and Mathematics, Arcadia University, 450 S. Easton Road, Glenside, PA 19038, USA.
Abstract
A two-center retrospective cohort study with independent external validation. To identify factors associated with progression from conservatively treated osteoporotic vertebral compression fracture (OVCF) to Kümmell disease (KD) and to evaluate machine-learning models for predicting this progression. A subset of patients with OVCF treated conservatively subsequently develop KD, which may result in persistent pain, progressive vertebral collapse, and neurological impairment. However, factors associated with this progression remain incompletely understood, and externally validated prediction models are lacking. The development and external validation cohorts included 555 and 126 patients, respectively. Clinical characteristics and quantitative DXA, CT, MRI, and paraspinal muscle parameters were analyzed. Independent factors associated with KD were identified using multivariable logistic regression. Twelve machine-learning classifiers were developed and internally evaluated using nested stratified cross-validation and subsequently assessed in the independent external cohort. Shapley additive explanations (SHAP) were used to interpret model predictions. Among 555 patients in the development cohort, 175 developed KD and 380 achieved vertebral union. Older age, diabetes, fracture during the hot season, higher vertebral bone quality score, greater paraspinal muscle fat infiltration, and larger Cobb angle were independently associated with KD, whereas higher bone mineral density (BMD) and osteoporosis treatment were associated with lower odds of KD. Under nested internal validation, CatBoost achieved the highest AUC of 0.983. In the external cohort, LightGBM achieved an AUC of 0.993 and the highest accuracy and F1 score. SHAP analysis identified BMD as the most influential predictor, followed by Cobb angle and age. Progression from OVCF to KD was associated with skeletal fragility, vertebral deformity, paraspinal muscle quality, age, diabetes, fracture season, and osteoporosis treatment. Machine-learning models demonstrated high discriminative performance in external validation. Prospective validation of a prespecified model is warranted before clinical implementation.