Using Machine Learning Models to Classify Favorability of TBI Outcomes One Year After Injury Using CT Imaging Findings and Other Patient Variables

Research Objectives: To use machine learning models to classify the favorability of patient TBI recovery one year after injury based on acute head CT scan features, patient demographics, and length of acute and rehab admissions.

Design: Cohort, 1 year follow-up. Independent variables included head CT scan features, demographic factors, and length of acute and rehab admissions. The Glasgow Outcome Scale Extended (GOSE) score was used as an outcome measure reflecting TBI recovery one year after injury. Participant GOSE scores were categorized as Favorable (GOSE 5-8) and Unfavorable (GOSE 0-4).

Up-sampling was used to equalize the number of participant data points with favorable and unfavorable scores and prevent bias in model training. Variables that highly correlated with each other were identified using Pearson correlation coefficient (threshold of 0.6) and were eliminated to prevent model overfitting. The dataset was split into 70% training and 30% testing sets for model training. Classification models were trained using these data. Their performance was evaluated based on predictive accuracy and F1 score.

Setting: Two urban hospitals, one private, and one public safety-net.

Participants: Hospital-based sample (n=301) participants eligible and enrolled in the NYU Rusk Rehabilitation Traumatic Brain Injury Model System Study through medical record abstraction and patient questionnaires.

Interventions: Not applicable.

Main Outcome Measures: The Glasgow Outcome Scale Extended (GOSE) score was used as an outcome measure reflecting TBI recovery one year after injury. Participant GOSE scores were categorized as favorable (GOSE 5-8) and unfavorable (GOSE 0-4).

Results: The five classification models trained and compared were Random Forest Classifier, Decision Tree, Logistic Regression, Linear Discriminant Analysis, and Support Vector Machine. Random Forest classifier had the highest accuracy and FI score: 81.37% and 80.41% respectively. Decision tree was the second highest classifier with accuracy of 70.5% and F1 score of 65.9%

Conclusions: Machine learning models may be used to classify patient GOSE scores – as favorable or unfavorable - one year post-TBI, using information generally available on acute admission. These models have the potential to assist in advanced assessments of TBI recovery.

Author(s) Disclosures: Authors have nothing to disclose.