# Optimizing pediatric “Mild” traumatic brain injury assessments: A multi-domain random forest analysis of diagnosis and outcomes

## TL;DR

This study uses machine learning to determine the best clinical assessments for diagnosing and predicting outcomes in children with mild traumatic brain injury.

## Contribution

The study introduces a multi-domain random forest analysis to optimize clinical assessments for pediatric mild TBI.

## Key findings

- Self-reported clinical ratings were more effective than performance-based metrics for diagnosing and predicting outcomes.
- Somatic complaints showed the highest predictive validity across all visits.
- Emotional disturbances predicted poor outcomes up to four months post-injury.

## Abstract

Despite advances in imaging and fluid-based biomarkers, the care for pediatric “mild” traumatic brain injury (pmTBI) remains primarily dependent on clinical evaluation. However, the optimal clinical assessments for diagnosing pmTBI and predicting outcomes remain debated, including which individual test or combinations of assessments are most effective, and how this evolves as a function of time post-injury.

Random Forest models were used to identify the most effective assessments for diagnostic (pmTBI vs. healthy controls) and outcome (pmTBI with favorable vs. poor outcomes, based on persisting symptoms) classification accuracy across a comprehensive battery including domains of self-reported clinical-ratings, paper-and-pencil cognitive tests, computerized cognitive tests, symptom provocation during neurosensory tests, and performance-based neurosensory measures. Assessments were conducted within 11-days, at 4-months and 1-year post-injury to examine acute and long-term recovery trajectories. A total of 323 pmTBI (180 males; age 14.5 ± 2.8 years) and 244 HC (134 males, 14.0 ± 2.9 years) were included (∼75 % 1-year retention) in final analyses.

Self-reported clinical-ratings outperformed performance-based metrics across all visits in both models, with somatic complaints demonstrating the highest predictive validity. Cognitive tests of memory aided diagnostic classification, while emotional disturbances were predictive of outcome classification up-to 4-months. Retrospective ratings, reflecting trait-like characteristics, were more predictive for identifying individuals at risk of poor outcomes. Computerized cognitive and neurosensory tests had limited predictive value beyond 1-week post-injury.

Clinicians should adopt a tailored approach for clinical assessments across different post-injury intervals to enhance clinical care, shorten assessment batteries, and better understand recovery in children with “mild” TBI.