Performance of ChatGPT-3.5 and ChatGPT-4 in the field of specialist medical knowledge on National Specialization Exam in neurosurgery

• Published in: Annales Academiae Medicae Silesiensis Vol. 78; pp. 253 - 258
• Main Authors: Laskowski, Maciej, Ciekalski, Marcin, Laskowski, Marcin, Błaszczyk, Bartłomiej, Setlak, Marcin, Paździora, Piotr, Rudnik, Adam
• Format: Journal Article
• Language: English
• Published: Śląski Uniwersytet Medyczny w Katowicach 15.10.2024
• Subjects: artificial intelligence (ai); chatgpt; neurosurgery
• ISSN: 1734-025X; 0208-5607; 1734-025X
• DOI: 10.18794/aams/186827

Introduction: In recent times, there has been an increased number of published materials related to artificial intelligence (AI) in both the medical field, and specifically, in the domain of neurosurgery. Studies integrating AI into neurosurgical practice suggest an ongoing shift towards a greater dependence on AI-assisted tools for diagnostics, image analysis, and decision-making. Material and methods: The study evaluated the performance of ChatGPT-3.5 and ChatGPT-4 on a neurosurgery exam from Autumn 2017, which was the latest exam with officially provided answers on the Medical Examinations Center in Łódź, Poland (Centrum Egzaminów Medycznych – CEM) website. The passing score for the National Specialization Exam (Państwowy Egzamin Specjalizacyjny – PES) in Poland, as administered by CEM, is 56% of the valid questions. This exam, chosen from CEM, comprised 116 single-choice questions after eliminating four outdated questions. These questions were categorized into ten thematic groups based on the subjects they address. For data collection, both ChatGPT versions were briefed on the exam rules and asked to rate their confidence in each answer on a scale from 1 (definitely not sure) to 5 (definitely sure). All the interactions were conducted in Polish and were recorded. Results: ChatGPT-4 significantly outperformed ChatGPT-3.5, showing a notable improvement with a 29.4% margin (p < 0.001). Unlike ChatGPT-3.5, ChatGPT-4 successfully reached the passing threshold for the PES. ChatGPT-3.5 and ChatGPT-4 had the same answers in 61 questions (52.58%), both were correct in 28 questions (24.14%), and were incorrect in 33 questions (28.45%). Conclusions: ChatGPT-4 shows improved accuracy over ChatGPT-3.5, likely due to advanced algorithms and a broader training dataset, highlighting its better grasp of complex neurosurgical concepts.