Assessing and validating machine learning-enhanced imputation of admission American Spinal Injury Association Impairment Scale grades for spinal cord injury.

Objective: The American Spinal Injury Association Impairment Scale (AIS) assigned at patient admission is an important predictor of outcomes following spinal cord injury (SCI). However, nearly 80% of records in the Spinal Cord Injury Model Systems (SCIMS) database-a multicenter prospective database of patients with SCI-lack admission AIS grades. Accurate imputation of this missing data could enable more robust analyses and insights into SCI recovery. This study aims to develop and validate methods for imputing missing admission AIS data in the SCIMS database.

Methods: The study included 16,062 patients with SCI from the publicly available SCIMS database (1988-2020). Five machine learning algorithms-random forest (RF), linear discriminant analysis, K-nearest neighbors, naive Bayes, and support vector machine-were compared using performance metrics (accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and multiclass area under the receiver operating characteristic curve) using five-fold cross-validation on a training subset of 6054 patients with complete AIS admission grades. The model with the highest performance was trained on all 16,062 patients. The imputed AIS grades were validated by predicting discharge functional independence measure (FIM) scores (range 13-91) with simple and multiple linear regression models on a 1:1 propensity score-matched cohort (n = 5828). Model performance was compared using differences in root mean square error (∆RMSE) with bootstrapped 95% confidence intervals (CIs).

Results: The full cohort contained a representative distribution of AIS grades (45% grade A, 13% grade B, 18% grade C, and 24% grade D), and the propensity score-matched cohort characteristics were well balanced. The RF algorithm demonstrated the highest validation accuracy (81.7%). Predictive models showed no significant differences between models using true versus imputed AIS grades, with 95% CIs for ∆RMSE of -0.60 to 0.47 for simple regression and -0.63 to 0.46 for multiple regression models. The coefficients of AIS grades also did not significantly differ between models with true versus imputed values.

Conclusions: A data-driven approach to imputation resulted in a robust method for imputing admission AIS grades that demonstrated clinical validity in the SCIMS database. This approach extends the utility of this longitudinal database and may provide a framework for other SCI databases.