(222) ChatGPT and prenatal diagnosis: Is there a role for artificial intelligence in fetal ultrasound?

Thursday, January 30, 2025

10:30 AM – 12:30 PM

Submitting Author and Presenting Author(s)

Jacqueline A. Erler, BA (she/her/hers)

Medical Student
Rutgers Robert Wood Johnson Medical School, New Jersey, United States

Coauthor(s)

Emily E. Daggett, MD

Maternal Fetal Medicine Fellow
Rutgers Robert Wood Johnson Medical School
Edison, New Jersey, United States
HP

Haylea S. Patrick, MD

Assistant Professor
Rutgers Robert Wood Johnson Medical School MFM Division
New Brunswick, New Jersey, United States

Objective: Anecdotal evidence indicates a rising popularity of generative machine learning models, like ChatGPT, as adjunct diagnostic tools in clinical medicine. Current evidence is limited for ChatGPT’s utility in diagnosing fetal syndromes when given ultrasonographic findings. We sought to assess ChatGPT’s ability to give an accurate fetal diagnosis based on standardized imaging vignettes for prenatal conditions with well-described fetal ultrasound findings.

Study Design: We wrote 50 vignettes on a broad scope of prenatal conditions. Each vignette included up to 5 characteristic ultrasound findings adapted from Callen's Ultrasonography in Obstetrics and Gynecology and peer-reviewed literature. A second-year Maternal-Fetal Medicine (MFM) fellow and board-certified MFM physician reviewed vignettes for clinical accuracy. Vignettes were entered into ChatGPT-4o, and the program was asked to generate a differential diagnosis in decreasing order of likelihood. ChatGPT responses were scored on the inclusion of an accurate diagnosis anywhere in the differential, in the top three most likely conditions, and as the single most likely diagnosis.

Results: The differential diagnosis generated by ChatGPT included the correct prenatal diagnosis in 82% (n=41) of the 50 tested cases. The correct diagnosis was in the top three most likely conditions in 76% (n=38) of cases and was selected as the single most likely diagnosis in 68% (n=34). The correct diagnosis was generated more frequently for common chromosomal abnormalities, including Trisomy 21, Trisomy 18, and Trisomy 13. Vignettes describing rare conditions with nonspecific ultrasound findings were less likely to result in a correct diagnosis.

Conclusion: ChatGPT gave an accurate prenatal diagnosis for a wide variety of fetal syndromes based on standardized descriptions of ultrasound findings. This study validates the use of ChatGPT to assist in prenatal diagnosis. Future studies aim to investigate the accuracy of ChatGPT in diagnosing real de-identified patient cases, and to compare the diagnostic accuracy of ChatGPT to that of expert clinicians.