283 Machine Learning–Based Prediction of Transfer Decisions in Mild Traumatic Brain Injuries: A Telemedicine-Guided Triage Strategy

Résumé fourni par la source

INTRODUCTION: Complicated mild traumatic brain injury (cmTBI) often leads to neurosurgical consultation and inter-hospital transfer, despite rare surgical intervention. to reduce unnecessary transfers and support local management, our community hospital network implemented a telehealth-based neurosurgical consultation program (tele-TBI). METHODS: We conducted a retrospective study across four community hospitals implementing tele-TBI for cmTBI patients. Over two years, patients were grouped by tele-TBI status. an XGBoost classifier was trained on clinical and radiological variables to predict transfer, with performance evaluated using standard metrics. Pairwise Pearson correlation assessed collinearity and bias among predictors. Highly collinear variables (r > 0.9) were reviewed for redundancy. RESULTS: 117 patients received tele-TBI consultations, while 62 eligible patients did not. Only 15 tele-TBI cases (13%) required transfer; the remaining 102 (87%) were managed at community hospitals. the XGBoost model achieved an AUC of 0.92, with 88.9% accuracy, 84.6% sensitivity, 91.3% specificity, 85% precision, and an F1-score of 0.85 at a threshold of 0.30. It identified teleconsultation (OR 0.0012, 95% CI [0.0001–0.02]), SDH (OR 3.65, 95% CI [1.95–6.82]), age (OR 0.22, 95% CI [0.10–0.47]), and GCS (OR 1.92, 95% CI [0.88–4.19]) as predictors of transfer. Correlation analysis showed a strong inverse association between teleconsultation and transfer (r = –0.84), and moderate associations for SDH (r = +0.31), age (r = –0.31), and GCS (r = +0.12). Older patients were more likely to receive teleconsultation (r = +0.47), while SDH was linked to reduced teleconsultation likelihood (r = –0.35). No strong collinearity was observed (all r < 0.9). CONCLUSIONS: Tele-TBI reduced unnecessary transfers. The XGBoost model (AUC = 0.92) identified teleconsultation, SDH, age, and GCS as key predictors. Teleconsultation was highly protective, with older patients more likely to receive it but less likely to be transferred.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.