The Field Attributes May not Accurately Predict the Need for Early Tracheostomy Tube Insertion in Severe TBI Patients: A New Insight With the Help of AI Algorithms.

Objective: Traumatic brain injury (TBI) patients often require prolonged intubation, and tracheostomies are often performed in intensive care units (ICUs) for patients who require prolonged ventilator support. Tracheostomy tube insertion can facilitate weaning in patients who require prolonged ventilation, leading to a decrease in mechanical ventilation duration and the length of stay in the intensive care unit. Additionally, it minimizes complications associated with extended tracheal intubation. There is a lack of data on determining which individuals will necessitate tracheostomy. Our objective was to predict tracheostomy requirements using patient data at arrival. Methods: We used a retrospectively collected data set of a large number of patients, who had been admitted to neuro-ICU Emtiaz Hospital, a large tertiary center of trauma. We trained several machine learning (ML) models in conjunction with initial predictors such as age, Glasgow comma scale (GCS), Rotterdam score, pupil response, first blood sugar, shift, and intracranial hematoma. The ML methods we used and compared were Logistic regression, random forest (RF), Gradient boosting machines (GBT), and multilayer perceptron (MLP). Results: 546 patients including 282 negatives (who did not receive tracheostomy) and 264 positives (who did receive tracheostomy) were included in our study. We randomly divided the data set into 70% for training and 30% for testing. The logistic regression predicted tracheostomy with a lower AUC (0.61) as compared to RF (AUC 0.64), MLP (AUC 0.65), and GBT (AUC 0.66). Conclusions: Our study aimed to create an ML model for predicting TTI in severe TBI patients. Despite using recognized predictors, the best model achieved an AUC of 0.66, indicating the inherent complexity in prediction. The findings serve as a valuable starting point for developing a predictive model and reassessing factors in the clinical setting that may not reliably predict tracheostomy needs at admission.