Objective:

To evaluate machine learning (ML) and deep learning (DL) models for acute detection of injuries and long-term prognostication of outcomes in pediatric traumatic brain injury (pTBI).

Approach:

• Methodology: A systematic review was conducted following PRISMA guidelines, analyzing studies that included pediatric patients aged 0 to 18 years, focusing on single or multimodal data for TBI identification or outcome prediction.

Key Findings:

• Twenty studies involving 68,331 subjects were analyzed.
• ML models, including XGBoost, Random Forest, and CNNs, showed potential to outperform traditional regression methods, achieving AUROCs up to 0.98 for mortality.
• Gains in low-risk triage were marginal, with ML models not significantly surpassing the 'no-information rate.'

Interpretation:

ML models have significant potential for enhancing pTBI care through improved risk stratification and automated imaging analysis.

Limitations:

• Need for model interpretability, such as SHAP values, to enhance understanding of model decisions.
• Addressing class imbalances is necessary to improve model performance across different patient populations.
• External multicenter validation is required to ensure regional generalizability of the findings.

Conclusion:

Future implementation of ML in pTBI requires focus on interpretability, class imbalances, and validation.

Sources:

• Systematic Review Registration