Abstract
BACKGROUND: Cerebrovascular diseases are the second most common cause of death worldwide and one of the major causes of disability burden. Advancements in artificial intelligence have the potential to revolutionize health care delivery, particularly in critical decision-making scenarios such as ischemic stroke management. OBJECTIVE: This study aims to evaluate the effectiveness of GPT-4 in providing clinical support for emergency department neurologists by comparing its recommendations with expert opinions and real-world outcomes in acute ischemic stroke management. METHODS: A cohort of 100 patients with acute stroke symptoms was retrospectively reviewed. Data used for decision-making included patients' history, clinical evaluation, imaging study results, and other relevant details. Each case was independently presented to GPT-4, which provided scaled recommendations (1-7) regarding the appropriateness of treatment, the use of tissue plasminogen activator, and the need for endovascular thrombectomy. Additionally, GPT-4 estimated the 90-day mortality probability for each patient and elucidated its reasoning for each recommendation. The recommendations were then compared with a stroke specialist's opinion and actual treatment decisions. RESULTS: In our cohort of 100 patients, treatment recommendations by GPT-4 showed strong agreement with expert opinion (area under the curve [AUC] 0.85, 95% CI 0.77-0.93) and real-world treatment decisions (AUC 0.80, 95% CI 0.69-0.91). GPT-4 showed near-perfect agreement with real-world decisions in recommending endovascular thrombectomy (AUC 0.94, 95% CI 0.89-0.98) and strong agreement for tissue plasminogen activator treatment (AUC 0.77, 95% CI 0.68-0.86). Notably, in some cases, GPT-4 recommended more aggressive treatment than human experts, with 11 instances where GPT-4 suggested tissue plasminogen activator use against expert opinion. For mortality prediction, GPT-4 accurately identified 10 (77%) out of 13 deaths within its top 25 high-risk predictions (AUC 0.89, 95% CI 0.8077-0.9739; hazard ratio 6.98, 95% CI 2.88-16.9; P<.001), outperforming supervised machine learning models such as PRACTICE (AUC 0.70; log-rank P=.02) and PREMISE (AUC 0.77; P=.07). CONCLUSIONS: This study demonstrates the potential of GPT-4 as a viable clinical decision-support tool in the management of acute stroke. Its ability to provide explainable recommendations without requiring structured data input aligns well with the routine workflows of treating physicians. However, the tendency toward more aggressive treatment recommendations highlights the importance of human oversight in clinical decision-making. Future studies should focus on prospective validations and exploring the safe integration of such artificial intelligence tools into clinical practice.