An interpretable machine learning model integrating early immune biomarkers for predicting outcomes after spinal cord injury

Abstract

Objective To characterize early systemic immune-inflammatory alterations after acute spinal cord injury (SCI), evaluate their association with 12-month neurological outcomes, and develop an interpretable model for early risk stratification. Methods This retrospective cohort study included 664 patients with acute traumatic injury. Patients were stratified according to vertebral fracture and SCI status to compare early peripheral immune-inflammatory profiles. Among them, 315 patients with SCI and complete 12-month follow-up data were included in the prognostic analyses. The primary outcome was an unfavorable neurological outcome, defined as death or an American Spinal Injury Association Impairment Scale (AIS) grade A–C at 12 months. Candidate predictors were identified using univariable logistic regression and least absolute shrinkage and selection operator (LASSO) regression, followed by multivariable logistic regression and Extreme Gradient Boosting (XGBoost) modeling. Model performance was evaluated using receiver operating characteristic (ROC) analysis, bootstrap validation, calibration analysis, decision curve analysis, and SHapley Additive exPlanations (SHAP). Results Patients with SCI exhibited more pronounced early systemic immune-inflammatory activation, characterized by elevated white blood cell count (WBC), absolute neutrophil count (ANC), and neutrophil-related inflammatory indices, accompanied by reduced lymphocyte parameters and higher incidences of pneumonia and deep vein thrombosis. Among the 315 patients with SCI, 110 (34.9%) experienced unfavorable outcomes. These patients had greater injury severity, worse baseline neurological status, more frequent cervical cord involvement, higher inflammatory marker levels, and more complications. The final model incorporated Injury Severity Score (ISS), AIS grade A, cervical cord involvement, and ANC. Compared with AIS grade A alone, the multivariable model demonstrated improved discrimination (AUC, 0.958 vs. 0.886; p < 0.001), with good calibration and minimal optimism on bootstrap validation. The XGBoost model showed comparable performance (AUC, 0.943). SHAP analysis further supported the relative importance of the selected predictors. Conclusion Acute SCI is associated with marked early neutrophil-dominant systemic inflammation, and elevated ANC was consistently associated with unfavorable neurological status at 12 months. A parsimonious model integrating ISS, AIS grade A, cervical cord involvement, and ANC provided additional prognostic information beyond baseline neurological assessment and may facilitate early individualized risk stratification using readily available clinical parameters.