Interpretable machine learning for predicting 28-day mortality after cervical spinal cord injury: development and external validation

Background Cervical spinal cord injury (cSCI) is associated with substantial early mortality. We aimed to develop and externally validate an exploratory machine learning model for predicting 28-day mortality in critically ill patients with cSCI. Methods This retrospective study used 232 patients from MIMIC-IV for development and 210 patients from the Second Hospital of Shanxi Medical University for external validation. The development cohort was divided into training and internal test sets at a stratified 70:30 ratio. Within ten-fold cross-validation repeated five times in the training partition, recursive feature elimination (RFE) and Boruta were applied independently, with their intersection used for feature selection; synthetic minority over-sampling was confined to each training fold. Eight algorithms were compared using the area under the receiver operating characteristic curve (AUC) and exploratory decision curve analysis. An additional analysis of saved support vector machine (SVM) probability outputs assessed calibration. Shapley additive explanations (SHAP) described model predictions. Results The SVM had the highest observed AUC among the evaluated models in the internal test cohort, 0.956 (95% CI, 0.895–1.000), and external cohort, 0.738 (95% CI, 0.664–0.812). The final training-set selection comprised APACHE II, APS III, temperature, age, glucose, bicarbonate, base excess, norepinephrine use, arterial pH, heart rate, and lactate. Conclusion The SVM showed potential prognostic utility in the studied ICU populations. Prospective multicenter validation, confirmation of probability calibration, and assessment of clinical impact are needed before routine clinical implementation.