Predictive Modeling of the Need for Tracheostomy after Traumatic Cervical Spinal Cord Injury Using Machine Learning.

PURPOSE: Traumatic cervical spinal cord injury (TCSCI) frequently necessitates mechanical ventilation, and early tracheostomy has been shown to reduce complications in patients requiring prolonged ventilation. This study aimed to develop a machine learning model to predict the need for tracheostomy in TCSCI patients, utilizing early clinical data to improve patient outcomes. MATERIALS AND METHODS: This study was conducted using data from 2017 to 2024, obtained from the single institution database. A total of 267 TCSCI patients were included, of whom 49 underwent tracheostomy. Variables selected for the model included demographics, comorbidities, injury level, medical research council motor grading, Glasgow Coma Scale (GCS), and treatment details. This study also implemented SHapley Additive exPlanations analysis to interpret the predictive model and identify significant risk predictors contributing to the outcomes. RESULTS: The CatBoost model outperformed other models, achieving the highest performance metrics. The model that incorporated GCS and the American Spinal Injury Association (ASIA) impairment scale (AIS) yielded the highest area under the curve (AUC) score. Following variable selection, the CatBoost model, utilizing age, GCS, AIS score, surgery, and injury levels (C3/4, C4/5, and C2/3), achieved an AUC score of 0.8166 and an accuracy of 0.8652. CONCLUSION: Our machine learning model effectively predicted the need for tracheostomy in TCSCI patients. Important predictors were age, GCS, AIS score, cervical surgery, and injury level. This model may improve outcomes for patients requiring tracheostomy.