Intersection of Artificial Intelligence and Medical Education (Student Abstract)

Keefer P. Wu, Patricia C. TsangAAAI 2024 · DOI

TL;DR

This study evaluates ChatGPT-3.5’s ability to answer complex, board-style pathology questions, finding an overall accuracy of 75%. While the AI shows promise as a fast, interactive learning tool—particularly in informatics and molecular pathology—its tendency to "hallucinate" facts and struggle with clinical judgment underscores the need for curated medical training data.

What Problem Does This Paper Address?

Pathology education is a rigorous 7-to-10-year journey requiring mastery of diverse fields, from tissue diagnosis to complex laboratory analysis. As Large Language Models (LLMs) become more common, researchers are investigating whether general-purpose AI like ChatGPT-3.5 can effectively serve as a medical tutor. The core challenge is determining if these models can handle high-stakes, specialized knowledge or if their limitations—specifically "hallucinations" and a lack of clinical nuance—render them a liability in medical training.

Methods

The researchers conducted a performance audit of OpenAI’s ChatGPT-3.5 using a dataset of 543 board-style multiple-choice questions. Because ChatGPT-3.5 was a text-only model at the time of the study, any questions requiring image interpretation were excluded. Responses were recorded and analyzed for accuracy, with incorrect answers manually categorized into five specific error types. To ensure factual reliability, the authors cross-referenced the AI's claims against peer-reviewed literature via PubMed.

Datasets

The study utilized a custom evaluation set of 543 questions spanning five major pathology disciplines:

  1. Surgical Pathology
  2. Clinical Pathology
  3. Hematopathology
  4. Informatics & Digital Pathology
  5. Molecular Pathology

The questions were randomly selected from several online pathology question banks.

Key Insights and Ideas

  • The "Informatics" Edge: The AI performed exceptionally well in Informatics & Digital Pathology (95%) and Molecular Pathology (91%). This suggests that LLMs excel in domains governed by clear rules and structured data.
  • The Clinical Judgment Gap: Performance dipped in Hematopathology (70%) and Clinical Pathology (71%), fields that require synthesizing laboratory values with clinical scenarios. This suggests AI struggles when medicine moves from "fact retrieval" to "clinical art."
  • Contextual Blindness: A significant finding was the AI's failure to recognize the importance of provided laboratory values, sometimes ignoring abnormal results even when reference ranges were provided in the prompt.
  • The Hallucination Problem: 41% of all errors were "hallucinations"—instances where the AI confidently presented false information as fact, such as misidentifying genetic causes of specific diseases.

Key Results

The AI achieved an overall accuracy of 75%, answering 406 out of 543 questions correctly. Responses were generated near-instantaneously, far exceeding human speed.

Table 1: Performance of ChatGPT-3.5 across various pathology disciplines.

The error analysis revealed that factual inaccuracies and hallucinations were the most common issues (41%), followed by ignoring lab values (20%), rank ordering errors (17%), management decision errors (16%), and testing limitations (6%).

Limitations and Future Work

The authors identify several hurdles for AI integration in medical residencies:

  • Lack of Multimodality: Future research must evaluate multimodal models (like GPT-4V) that can interpret histology slides, as pathology is a visual-heavy discipline.
  • Training Data Quality: The authors argue for models trained on curated, high-quality medical literature rather than uncurated internet scrapes to reduce hallucinations.
  • Passing Thresholds: It remains unclear if a 75% score constitutes a "passing" grade on actual board exams, as these thresholds vary.
  • Expert Collaboration: The paper calls for closer collaboration between AI developers and medical experts to create medically rigorous "tutor" bots.

Summary updated

BibTeX
@inproceedings{wu2024intersection,
  author = {Keefer P. Wu and Patricia C. Tsang},
  title = {{Intersection of Artificial Intelligence and Medical Education (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30525},
}
Figure