Artificial intelligence (AI) could potentially be a valuable tool for neurologists in identifying the location of a stroke in the brain. In a recent study, AI was utilized to analyze text from health histories and neurologic examinations to pinpoint brain lesions. The study focused on the generative pre-trained transformer 4 (GPT-4) model and was published in the March 27, 2024, online edition of Neurology® Clinical Practice, an official journal of the American Academy of Neurology.
A stroke can result in long-term disability or even death. Understanding the precise location of a stroke in the brain can help predict the long-term effects, such as speech and language problems or impaired movement. Additionally, it can aid in determining the most effective treatment and overall prognosis.
Brain tissue damage caused by a stroke is known as a lesion. By combining a review of a person’s health history with a neurologic exam, lesions can be identified. The exam involves evaluating symptoms and conducting cognitive assessments. Individuals who have experienced a stroke often undergo brain scans to locate lesions.
Study author Dr. Jung-Hyun Lee from State University of New York (SUNY) Downstate Health Sciences University and an American Academy of Neurology member explained, “Not all stroke patients have access to brain scans or neurologists, so we aimed to assess whether GPT-4 could accurately identify brain lesions post-stroke based on health history and a neurologic exam.”
The study utilized 46 published cases of stroke patients. Researchers extracted text from participants’ health histories and neurologic exams, which was then input into GPT-4. The model was tasked with determining whether participants had lesions, the side of the brain where lesions were located, and the specific brain region affected. Results from GPT-4 were compared to brain scans for validation.
The study revealed that GPT-4 successfully identified brain lesions in many participants, including the side of the brain and the specific brain region, except for lesions in the cerebellum and spinal cord. The model showed a sensitivity of 74% and a specificity of 87% in identifying the side of the brain with lesions, as well as a sensitivity of 85% and a specificity of 94% in determining the brain region.
Although GPT-4 demonstrated consistency in identifying the number of brain lesions, the side of the brain, and brain regions for a majority of participants, it provided accurate answers for 41% of participants when combining responses to all questions and times.
Dr. Lee noted, “While not yet ready for clinical use, large language models like generative pre-trained transformers have the potential to assist in locating lesions post-stroke and may help reduce healthcare disparities, especially in underserved areas with limited neurologic care access.”
One limitation of the study is that the accuracy of GPT-4 relies on the quality of information provided. Detailed health histories and neurologic exam data may not always be available for all stroke patients.
