A grade-specific clinical prediction rule for independent walking after traumatic spinal cord injury.

OBJECTIVE: To identify significant predictors for individual American Spinal Injury Association Impairment Scale (AIS) grades, and develop a clinical prediction rule (CPR) with differentiated prognosis for patients with different spinal cord injury (SCI) grades. DESIGN: Selection of age, gender, ASIA and Functional Independence Measure (FIM) scores as predictive variables, to predict independent walking one year after SCI. Optimal variable combinations for individual AIS classifications were screened, and various machine learning models were established. SETTING: Analysis of National Spinal Cord Injury Statistical Centre (NSCISC) dataset. PARTICIPANTS: 2607 patients with traumatic SCI. INTERVENTIONS: Not applicable. MAIN OUTCOME MEASURES: Self-reported ability to walk both indoors and outdoors. RESULTS: Optimal variable combinations differed significantly across different AIS grades. For AIS A, C, and D, significant predictors were concentrated in motor and sensory scores at lumbosacral levels, with age also validated as a key predictor. However, for AIS B, key predictors were concentrated in lower thoracic segments, with pinprick sensation being predominant. After modelling with optimal variable combinations, accuracy, F1 score and AUC value were significantly improved, especially for AIS B. Increasing model complexity improved prediction performance across all AIS grades, though the optimal model differed by grade. We therefore developed an ensemble model integrating random forest, artificial neural network, and multi-task learning, which performed at least as well as any individual base model across all metrics. In addition, incorporating discharge variables, such as discharge ASIA and FIM scores, significantly reduced predictive discrepancy between AIS B+C and AIS A+D. CONCLUSIONS: Optimal-variable ensemble model we proposed significantly improved prediction performance for all AIS grades, especially for AIS B and C. Incorporating discharge variables significantly reduced predictive discrepancy among different grades. This grade-specific CPR holds great significance for individualized rehabilitation and precision medicine in SCI recovery.