Predicting walking recovery after traumatic cervical spinal cord injury with machine learning: a multicenter cohort study

This study aims to investigate the effectiveness of various ML models—Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), Logistic Regression (LR), Least Absolute Shrinkage and Selection Operator (LASSO), and Decision Tree (DT)—in predicting walking recovery after TCSCI. The goal is to develop a robust, personalized predictive model using demographics, injury characteristics, peripheral inflammatory blood biomarkers, and neuroimaging factors metrics. A retrospective cohort of 468 TCSCI patients treated between 2016 and 2022 was included. Patients were divided into training (235), testing (99), and validation (134) groups. Key prognostic features, including demographic data, injury characteristics, peripheral inflammatory blood biomarkers, and MRI features were collected. Several ML models were developed and validated using the training cohort, and performance was assessed in both internal (testing cohort) and external (validation cohort) validation cohorts. Model performance was evaluated using area under the curve (AUC) with 95% DeLong confidence intervals, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Brier score for calibration. The study demonstrated that all six ML models successfully predicted walking recovery after TSCI. Among the models, the RF model exhibited the highest performance in the training cohort, achieving perfect AUC, sensitivity, specificity, PPV, and NPV. However, its performance slightly decreased in the validation cohort, AUC of 0.990 (95% CI 0.977–1.000). The LR model showed stable and reliable performance across all cohorts, making it a solid option for predicting recovery. The SVM, DT, and LASSO models also exhibited strong predictive abilities, though with slight variations in performance across cohorts. However, the KNN model showed the lowest overall accuracy and NPV. ML models, particularly RF and LR, offer robust tools for predicting walking recovery in patients with TCSCI. While the RF model performed best in the training cohort, the LR model demonstrated consistent and reliable results across all cohorts, suggesting its potential for clinical application. Meanwhile, the SVM, DT and LASSO models also showed strong performance. These models could aid in personalized rehabilitation planning, improving clinical decision-making and patient prognostication.