GPT-5 shows high sensitivity but 16.4% false-positive rate for acute ischemic stroke on DWI

GPT-5 detected acute ischemic stroke on preselected DWI slices with 100% sensitivity (95% CI 96.6–100%) and 83.6% specificity (95% CI 75.6–89.4%), yielding a 16.4% false-positive rate. Accuracy for vascular territory was 90.9%.

• Stroke detection sensitivity was 100% and specificity 83.6% (κ=0.836) on isolated DWI images without ADC maps.
• Conditional MCA segment territory accuracy was 71.6%, and end-to-end MCA segment accuracy was 66.3%, showing reduced performance in fine localization.
• False positives arose mainly from artifact misinterpretation (72.2%), chronic lesion misclassification (27.8%), and terminology issues; the authors emphasize that radiologist confirmation is needed and the tool does not yet have a clinical decision-support role.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.