Evaluation of ChatGPT-5 for CT Imaging in Canadian CT Head Rule-Positive Mild Traumatic Brain Injury: A Pilot Study.
Authors
Affiliations (2)
Affiliations (2)
- Department of Neurosurgery, University Hospital of Ioannina, 45100 Ioannina, Greece.
- Medical School, University of Ioannina, 45110 Ioannina, Greece.
Abstract
<b>Background and Objectives:</b> Mild traumatic brain injury (mTBI) is a common cause of emergency department visits worldwide. Optimal decisions are essential for optimizing patient outcomes while minimizing unnecessary imaging. The present study explores the potential utility of ChatGPT-5.1 for predicting intracranial injury and determining CT necessity in mTBI. <b>Methods:</b> We evaluated adult patients with mild traumatic brain injury who met Canadian CT Head Rule (CCHR) criteria and underwent CT imaging during a two-year period. ChatGPT-5.1 was prompted under two frameworks: ER physician simulation and rule (literature)-based. Additionally, a probabilistic multi-variable logistic regression model was developed. ChatGPT-5.1's performance was compared with that of human controls (consultants and residents) and the CCHR. Pairwise comparisons were conducted using McNemar's test. <b>Results:</b> A total of 127 patients were included. ChatGPT-5.1's recommendation for CT showed comparable performance to CCHR. For predicting CT findings, the probabilistic model achieved the highest performance (accuracy 74.0%), followed by consultants (accuracy 69.3%) and residents (66.9%). The rule (literature)-based model displayed the highest sensitivity (91.8%) but exhibited low specificity (22.2%), indicating a high rate of false positives despite its potential utility as a screening or decision-supporting tool. The ER physician simulation model demonstrated the lowest overall performance (accuracy 54.3%). Statistically significant differences were observed between the probabilistic model and both the ER physician and rule-based models (<i>p</i> < 0.05). Consultants outperformed the ER simulation model (<i>p</i> = 0.01), while their performance was similar to that of the probabilistic model (<i>p</i> = 0.44). <b>Conclusions:</b> ChatGPT-5.1 demonstrated performance comparable to the Canadian CT Head Rule in predicting the need for head CT imaging but showed lower accuracy in identifying intracranial injury. While the literature-based rule model achieved high sensitivity, its low specificity emphasizes that it acts primarily as a conservative screening aid rather than a definitive diagnostic tool in CCHR-positive mTBI patients.