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

Abstract

Objective: 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.

Method: 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.

Results: 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.

Conclusions: 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.

Keywords: Clinical assessments; Diagnostic classification; Machine learning; Mild traumatic brain injury; Outcome classification; Pediatric.

Fig. 1

Participant recruitment and retention. Flowchart of enrolment, inclusion and data quality control from Visits V1 (=within 11 days of injury), V2 (approximately 4-months post-injury) and V3 (approximately 1-year post-injury) for patients with pediatric “mild” traumatic brain injury (pmTBI) and matched healthy controls (HC). The asterisk denotes the total number of participants who were eligible to return, which is a sum of participants with usable clinical data and those whose data was excluded at previous visits due to quality assurance issues.

Fig. 2

Results from the Random Forest analyses for the final diagnostic models at each Visit (V) for classifying pediatric “mild” traumatic brain injury patients versus healthy controls. The top row shows receiver operating characteristics (ROC) results including area under the curve (AUC), balanced accuracy (BA), sensitivity, and specificity for each visit. The bottom row displays the variable importance (VIMP) score for each variable in the final models at each visit. Feature selection for the final model occurred based on whether the lower bound of the 90 % VIMP confidence interval was greater than 0 % (selected = blue; not selected = red). For Visit 1, current/retrospective clinical-ratings predominated for feature selection along with memory and executive function on paper-and-pencil cognitive tests. In contrast, symptom provocation ratings from multiple neurosensory tests were eliminated in the final model. Somatic complaints (sleep and headache), overall post-concussive symptom burden, and performance on memory tests exhibited the best diagnostic accuracy at visits 2 and 3. The following abbreviations are included in the figure: Post-Concussive Symptoms (PCS), Quality of Life (QoL), Retrospective (R), Immediate (IR) and Delayed Recall (DR) from the Hopkins Verbal Learning Test Revised (HVLT-R), Stroop Inhibition (Stroop I), Symptom Provocation (SP), King-Devick (KD), Reaction Time (RT), Visual Motion Sensitivity (VMS), Near Point Convergence (NPC), Vertical Saccades (V Sac), Monocular Accommodative amplitude (MA), Vertical Vestibular Ocular Reflex (V VOR), Tandem Gait (TG), Horizontal Vestibular Ocular Reflex (H VOR), Double Dorsal Foot Stretch (DDFS).

Fig. 3

Results from the Random Forest analyses for the final outcome models at each Visit (V) for classifying pediatric “mild” traumatic brain injury patients with poor versus favorable outcomes. The top row shows receiver operating characteristics (ROC) results including area under the curve (AUC), balanced accuracy (BA), sensitivity, and specificity for each visit. The bottom row displays the variable importance (VIMP) score for each variable in the final models. Feature selection for the final model occurred based on whether the lower bound of the 90 % VIMP confidence interval was greater than 0 % (selected = blue; not selected = red). Across all three visits, current clinical-ratings, in particular somatic complaints (headache and sleep) and emotional distress, together with neurosensory symptom provocation measures, predominated for feature selection. In contrast, injury severity characteristics and performance-based cognitive measures tended to be eliminated in the final models. The following abbreviations are included in the figure: Post-Concussive Symptoms (PCS), Quality of Life (QoL), Retrospective (R), Symptom Provocation (SP), Tandem Gait (TG), Visual Motion Sensitivity (VMS), Monocular Accommodative amplitude (MA), Horizontal and Vertical Vestibular Ocular Reflex (H VOR; V VOR), King-Devick (KD), Error (Er), Horizontal and Vertical Saccades (H Sac; V Sac), Double Dorsal Foot Stretch (DDFS), Smooth Pursuit (Sm Pur), Near Point Convergence (NPC), Number of Previous Injuries (NumPrevInj), Stroop Inhibition (Stroop I), Identification (IDN), One-card Learning (OCL), Detection (DET), Reaction Time (RT), Accuracy (ac), Loss of Consciousness/Posttraumatic Amnesia (LOC/PTA).