An Artificial Intelligence-Based Framework for Enhanced Clinical Decision-Making in Minimizing Repetitive Pediatric Brain CT Scans

Abstract

Background: Children frequently receive repeat brain CT scans without strong justification. Rapid bedside estimation of cumulative radiation risk is challenging, and no validated clinical rule exists for repeat scans in pediatric head trauma. We developed a mobile clinical decision support tool integrating PECARN-informed criteria with an artificial neural network (ANN) that approximates ICRP-defined cancer risk from cumulative CT exposure.

Methods: We retrospectively trained KNN, SVM, and ANN models to approximate ICRP risk equations using data from 4025 pediatric brain CT scans (age 0–16 years). The best-performing model was embedded with PECARN-informed criteria into a mobile app. Eleven clinicians completed a usability survey, and repeat CT request rates were tracked over six months using statistical process control.

Results: The ANN achieved the lowest approximation error (MAPE 6.92%; R² 0.917) compared with KNN (11.24%; 0.842) and SVM (8.76%; 0.879). Over six months, repeat CT requests decreased by a mean 57.0% (range 50.5%–63.0%). A downward trend met Nelson rule criteria for non-random pattern (p < 0.01). Pilot usability: 72.8% found navigation easy or very easy.

Conclusions: A mobile tool using PECARN-informed criteria and real-time ICRP risk estimation reduced repeat pediatric brain CT requests by 57% in a Saudi emergency department. Larger prospective validation is required.