Functional outcome prediction after traumatic cervical spinal cord injury using ensemble machine learning: a three‑center validation study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Sichuan Academy of Medical Sciences & Sichuan Provincial People's Hospital, Chengdu, 610072, China.
- Department of Radiology, 905Hospital of People's Liberation Army Navy, Shanghai, 200052, China.
- North Sichuan Medical College, Nanchong, 637000, China. [email protected].
- Shijiazhuang People's Hospital, Shijiazhuang, 050000, China. [email protected].
Abstract
Traumatic cervical spinal cord injury (TCSCI) often causes severe neurological dysfunction. Accurate prediction of functional recovery is essential for clinical decision‑making and rehabilitation planning. To develop an ensemble learning model integrating baseline clinical data, neurological assessments, and cervical MRI features to predict neurological recovery and functional outcomes at one year post‑injury in TCSCI patients. We retrospectively collected data from 410 TCSCI patients across three medical institutions (2017-2025). A prediction model was constructed using a two‑layer Stacking ensemble strategy. Baseline characteristics included demographic, clinical, and radiologic features. Primary outcome was one‑year ASIA Impairment Scale (AIS) grade; secondary outcomes were Upper Extremity Motor Score (UEMS), Lower Extremity Motor Score (LEMS), Total Motor Score (TMS), and Spinal Cord Independence Measure III (SCIM III). SHapley Additive exPlanations (SHAP) analysis was performed to evaluate model interpretability and quantify the contribution of each predictor to the model output. Performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC), R², mean absolute error (MAE), and root mean square error (RMSE). Of 340 patients analyzed (mean age 54.1 ± 14.9 years; 229 males), 242 formed the training set and 98 the external test set. The model achieved AUC ≥ 0.85 for all AIS grades. For continuous outcomes, R² values for UEMS, LEMS, TMS, and SCIM III were 0.9867, 0.9880, 0.9872, and 0.9863, respectively, with corresponding MAEs of 2.0667, 7.2751, 7.9496, 3.5956. SHAP analysis identified baseline UEMS and AIS grade as the most influential predictors, followed by maximum spinal cord compression (MSCC). This externally validated model accurately predicts 1‑year functional recovery in TCSCI patients and may support early prognosis assessment and individualized rehabilitation planning.