GPT-4 shows promise in locating brain lesions after stroke

Trending 2 weeks ago
ARTICLE AD BOX

Artificial intelligence (AI) whitethorn service arsenic a early instrumentality for neurologists to thief find wherever successful nan encephalon a changeable occurred. In a caller study, AI processed matter from wellness histories and neurologic examinations to find lesions successful nan brain. The study, which looked specifically astatine nan ample connection exemplary called generative pre-trained transformer 4 (GPT-4), is published successful nan March 27, 2024, online rumor of Neurology® Clinical Practice, an charismatic diary of nan American Academy of Neurology.

A changeable tin origin semipermanent disablement aliases moreover death. Knowing wherever a changeable has occurred successful nan encephalon helps foretell semipermanent effects specified arsenic problems pinch reside and connection aliases a person's expertise to move portion of their body. It tin besides thief find nan champion curen and a person's wide prognosis.

Damage to nan encephalon insubstantial from a changeable is called a lesion. A neurologic exam tin thief find lesions, erstwhile paired pinch a reappraisal of a person's wellness history. The exam involves denotation information and reasoning and representation tests. People pinch changeable often person encephalon scans to find lesions.

Not everyone pinch changeable has entree to encephalon scans aliases neurologists, truthful we wanted to find whether GPT-4 could accurately find encephalon lesions aft changeable based connected a person's wellness history and a neurologic exam."

Jung-Hyun Lee, MD, study author of State University of New York (SUNY) Downstate Health Sciences University successful Brooklyn and personnel of nan American Academy of Neurology

The study utilized 46 published cases of group who had stroke. Researchers gathered matter from participants' wellness histories and neurologic exams. The earthy matter was fed into GPT-4. Researchers asked it to reply 3 questions: whether a subordinate had 1 aliases much lesions; connected which broadside of nan encephalon lesions were located; and successful which region of nan encephalon nan lesions were found. They repeated these questions for each subordinate 3 times. Results from GPT-4 were past compared to encephalon scans for each participant.

Researchers recovered that GPT-4 processed nan matter from nan wellness histories and neurologic exams to find lesions successful galore participants' brains, identifying which broadside of nan encephalon nan lesion was on, arsenic good arsenic nan circumstantial encephalon region, pinch nan objection of lesions successful nan cerebellum and spinal cord.

For nan mostly of people, GPT-4 was capable to place connected which broadside of nan encephalon lesions were recovered pinch a sensitivity of 74% and a specificity of 87%. Sensitivity is nan percent of existent positives that are correctly identified arsenic positive. Specificity is nan percent of negatives that are correctly identified. It besides identified nan encephalon region pinch a sensitivity of 85% and a specificity of 94%.

When looking astatine really often nan 3 tests had nan aforesaid consequence for each participant, GPT-4 was accordant for 76% of participants regarding nan number of encephalon lesions. It was accordant for 83% of participants for nan broadside of nan brain, and for 87% of participants regarding nan encephalon regions.

However, erstwhile combining its responses to each 3 questions crossed each 3 times, GPT-4 provided meticulous answers for 41% of participants.

"While not yet fresh for usage successful nan clinic, ample connection models specified arsenic generative pre-trained transformers person nan imaginable not only to assistance successful locating lesions aft stroke, they whitethorn besides trim wellness attraction disparities because they tin usability crossed different languages," said Lee. "The imaginable for usage is encouraging, particularly owed to nan awesome request for improved wellness attraction successful underserved areas crossed aggregate countries wherever entree to neurologic attraction is limited."

A limitation of nan study is that nan accuracy of GPT-4 depends connected nan value of nan accusation it is provided. While researchers had elaborate wellness histories and neurologic exam accusation for each participant, specified accusation is not ever disposable for everyone who has a stroke.

Source:

Journal reference:

Lee, J-H., et al. (2024) GPT-4 Performance for Neurologic Localization. Neurology Clinical Practice. doi.org/10.1212/CPJ.0000000000200293.