Spinal Cord Injury and Neurorehabilitation Deep Learning–Based Prediction of Functional Recovery Trajectories After Traumatic Spinal Cord Injury

Abstract

Traumatic spinal cord injury produces substantial neurological impairment and multidimensional limitations in mobility, self-care, sphincter management, and community participation. This quantitative, multicenter, retrospective longitudinal cohort study developed and validated a multimodal deep learning model for predicting functional recovery trajectories following traumatic spinal cord injury. Of 3,184 patient records screened across seven specialist centers, 948 were excluded, leaving 2,236 eligible adults for analysis. Demographic characteristics, International Standards for Neurological Classification of Spinal Cord Injury scores, American Spinal Injury Association Impairment Scale grades, magnetic resonance imaging biomarkers, medical complications, and rehabilitation exposures were collected from acute admission through rehabilitation discharge and follow-up at three, six, and twelve months. The primary outcome was the longitudinal Spinal Cord Independence Measure III score. Secondary outcomes included Functional Independence Measure motor scores, Walking Index for Spinal Cord Injury II progression, AIS conversion, and independent ambulation. The adjusted mean SCIM III score increased from 30.8 at rehabilitation admission to 69.5 at twelve months, representing a 38.7-point improvement. The largest increase occurred between rehabilitation admission and discharge, with a mean gain of 18.5 points, whereas the six-to-twelve-month gain was 3.7 points, indicating deceleration in functional recovery. Among 1,607 participants assessed at twelve months, 649 (40.4%) demonstrated AIS conversion and 963 (59.9%) achieved independent ambulation. In geographic external validation, the deep learning model predicted twelve-month SCIM III scores with a mean absolute error of 6.9 points, root mean squared error of 9.1 points, R² of 0.80, and calibration slope of 0.90. Compared with XGBoost, the strongest benchmark, the model reduced mean absolute error by 0.9 points or 11.5% (95% CI [0.4, 1.4], p = .001). Independent walking was predicted with 88% sensitivity, 85% specificity, an F1 score of 0.89, and an AUROC of 0.93. Baseline SCIM III, AIS grade, total motor score, neurological level, age, preserved sensory function, and rehabilitation intensity contributed most strongly to prediction. The findings demonstrated that multimodal deep learning provided accurate, calibrated, and clinically informative estimates of nonlinear functional recovery trajectories across internal, temporal, and geographic validation samples.