Refining outcome prediction after traumatic brain injury with machine learning algorithms.

Area of Science:

• Neuroscience
• Medical Informatics
• Biostatistics

Background:

• Traumatic brain injury (TBI) outcome is typically assessed using the 8-grade Glasgow Outcome Scale Extended (GOSE).
• Traditional binary outcome prediction (favorable/unfavorable) may miss subtle TBI recovery nuances.
• Developing models for the full GOSE scale is crucial for precise TBI outcome assessment.

Purpose of the Study:

• To explore machine learning (ML) methods for predicting the full 8-grade GOSE in TBI patients.
• To compare the performance of different ML models in TBI outcome prediction.
• To validate ML models using internal and external patient cohorts.

Main Methods:

• Utilized patient data including age, GCS-motor score, pupillary reaction, and Marshall CT score.
• Developed and validated proportional odds logistic regression (POLR), random forest regression, and neural network models.
• Included 866 patients for internal validation and external validation cohorts from Leuven (369) and ProTECT III study (573).

Main Results:

• ML models achieved accuracies of 0.3-0.35 on internal data, significantly outperforming the random baseline (0.125).
• Models demonstrated satisfactory performance in the Leuven external validation cohort.
• Performance was unsatisfactory when applied to the ProTECT III multi-center study dataset.

Conclusions:

• Machine learning models show potential for predicting the full GOSE scale in TBI.
• Model generalizability varies across external datasets, highlighting the need for robust validation.
• Further research is needed to improve ML model performance and applicability in diverse TBI populations.