Machine learning algorithms performed no better than regression models for prognostication in traumatic brain injury

Area of Science:

• Neuroscience
• Medical Informatics
• Biostatistics

Background:

• Traumatic brain injury (TBI) is a leading cause of death and disability.
• Accurate outcome prediction is vital for patient management and resource allocation.
• Machine learning (ML) offers potential for improved prediction models.

Purpose of the Study:

• To evaluate the added value of common ML algorithms for predicting outcomes in moderate to severe TBI.
• To compare the performance of ML algorithms against traditional regression models.
• To assess the generalizability of prediction models across different cohorts.

Main Methods:

• Utilized logistic regression (LR), lasso, and ridge regression.
• Applied ML algorithms: support vector machines, random forests, gradient boosting machines, and artificial neural networks.
• Trained and validated models on IMPACT-II and CENTER-TBI databases, assessing calibration and discrimination.

Main Results:

• Discrimination and calibration varied across studies but were similar among algorithms.
• ML algorithms did not consistently outperform traditional regression models.
• Mean area under the curve was 0.82 for mortality and 0.77 for unfavorable outcomes in the CENTER-TBI study.

Conclusions:

• ML algorithms may not offer superior performance over regression in low-dimensional TBI outcome prediction.
• Rigorous external validation is essential for ML and regression-based prediction models.
• Ensuring applicability to new populations is critical for reliable TBI outcome prediction.