Machine Learning for Predicting Discharge Disposition After Traumatic Brain Injury.

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Current prognostic models for traumatic brain injury (TBI) primarily focus on severe cases, predicting mortality and Glasgow Outcome Scale.
• Mild and moderate TBI patients rarely experience severe outcomes, necessitating novel prognostic endpoints.
• Discharge disposition is an underutilized outcome that can reflect functional status across all TBI severities.

Purpose of the Study:

• To develop machine learning (ML) models for predicting trichotomized discharge disposition in TBI patients.
• To establish a new prognostic outcome that serves as a proxy for functional status, particularly in mild and moderate TBI.
• To enhance the predictive accuracy of TBI outcomes beyond traditional measures.

Main Methods:

• Utilized a large dataset of 5,292 TBI patients from a quaternary care center.
• Included 84 diverse predictors such as vitals, demographics, injury mechanism, Glasgow Coma Scale, and comorbidities.
• Trained and evaluated six ML algorithms using nested-stratified-cross-validation, followed by hyperparameter optimization and model selection.

Main Results:

• A random forest model demonstrated superior performance in predicting discharge disposition.
• The model achieved a weighted average area under the receiver operating characteristic curve of 0.84 (95% CI 0.81-0.87).
• The weighted average area under the precision-recall curve was 0.85 (95% CI 0.82-0.88).

Conclusions:

• Developed high-performing ML models capable of predicting trichotomized discharge disposition in TBI patients.
• These models offer a valuable tool for optimizing patient triage and treatment strategies.
• The findings are particularly relevant for improving care in mild and moderate TBI cases.