E-TBI: explainable outcome prediction after traumatic brain injury using machine learning.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Neurology

Background:

• Traumatic brain injury (TBI) is a common health issue requiring accurate severity assessment for effective management.
• Current machine learning (ML) approaches for TBI outcome prediction face challenges with limited data and lack of interpretability for clinicians.
• Explaining ML decisions is crucial for trust and adoption by medical professionals, especially those with less experience.

Purpose of the Study:

• To develop an explainable machine learning tool, E-TBI, for predicting TBI severity.
• To provide a user-friendly interface for visualizing the decision-making process of the ML model.
• To improve the interpretability and applicability of automated TBI outcome prediction in clinical settings.

Main Methods:

• Developed E-TBI, a web-based tool integrating feature selection and classification modules.
• Utilized multimodal patient data including demographics, clinical information, lab results, and CT findings.
• Investigated various ML models and feature selection techniques, identifying Gradient Boosting Machine and Random Forest (GBMRF) as optimal.

Main Results:

• The GBMRF model achieved high accuracy rates of 88.82% and 89.78% on two distinct datasets.
• Identified a small set of essential features, leading to a 35% reduction in patient testing costs.
• The E-TBI tool provides visualized decision rules for enhanced interpretability.

Conclusions:

• E-TBI offers a valuable, explainable ML solution for TBI severity prediction.
• The tool enhances clinical decision-making by providing interpretable predictions and reducing diagnostic costs.
• This approach addresses limitations of current ML methods in handling imbalanced data and explaining outcomes.