DOI: 10.1097/bn9.0000000000000113 ISSN: 3066-8905

Early Prediction of Prolonged Hospitalization After Penetrating Spinal Cord Injury: A Machine Learning Analysis of the National Trauma Data Bank

Wedam Nyaaba, Alexa Lauinger, Neema Darabi, Helen Kemprecos, Samuel Blake, Prasanna S. Padmanabham, Gregory M. Polites, Paul M. Arnold
Show PDF Cite

Study Design:

Retrospective cohort study using the American College of Surgeons National Trauma Data Bank (NTDB) from 2017 to 2023. Prolonged hospital length of stay (LOS) was defined as exceeding the 75th percentile of the cohort distribution (>26 d).

Objective:

To develop and internally validate machine learning (ML) models for predicting prolonged LOS after penetrating spinal cord injury (pSCI) and to identify clinically available admission-stage predictors that retain useful performance in a reduced-feature model.

Summary of Background Data:

pSCI is a distinct subset of spinal cord injury with high morbidity and resource utilization, primarily affecting younger, socioeconomically vulnerable populations. pSCI-specific LOS prediction models remain limited.

Materials and Methods:

We studied 2912 adults with pSCI; 665 (22.8%) had prolonged LOS. We trained logistic regression, Naïve Bayes, decision tree, random forest, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), support vector machine, and deep neural network (DNN) models with five-fold cross-validation and an independent hold-out test set. The four strongest full-feature models were retrained on a clinically motivated set of 10 early-available predictors. SHapley Additive exPlanations (SHAP) analyses provided model interpretation. Sensitivity analyses tested robustness across alternative LOS thresholds.

Results:

Patients with prolonged LOS had markedly higher rates of pressure ulcer (26.2% vs . 2.5%), deep vein thrombosis (12.0% vs. 1.6%), sepsis (5.1% vs . 0.8%), and pulmonary embolism (6.0% vs . 1.0%; all P ≤0.001). Full-feature models achieved hold-out ROC AUCs of 65% to 72%, with XGBoost performing best (72.44%). A 10-variable reduced model retained comparable discrimination (DNN-reduced ROC AUC: 67.48%). Discrimination was consistent across alternative LOS threshold definitions (ROC AUC: 0.66–0.70).

Conclusion:

ML models using routinely available admission-stage variables identify patients at risk for prolonged hospitalization after pSCI with modest but clinically meaningful discrimination. A parsimonious 10-variable reduced model retains useful performance, supporting early risk stratification, discharge planning, and resource allocation. External validation is needed before clinical implementation.

Level of Evidence:

Level III [OCEMB Levels of Evidence Working Group 2011].

More from our Archive