[Paper Review] Causal-Discovery Performance of ChatGPT in the context of Neuropathic Pain Diagnosis

Ruibo Tu, Chao Ma+1 more|arXiv (Cornell University)|Jan 24, 2023

Artificial Intelligence in Healthcare and EducationMedicine16 citations

TL;DR

The paper evaluates how well ChatGPT answers causal-discovery questions in neuropathic pain diagnosis using a sampled benchmark, finding limited and inconsistent performance with notable false negatives and language/concept limitations.

ABSTRACT

ChatGPT has demonstrated exceptional proficiency in natural language conversation, e.g., it can answer a wide range of questions while no previous large language models can. Thus, we would like to push its limit and explore its ability to answer causal discovery questions by using a medical benchmark (Tu et al. 2019) in causal discovery.

Motivation & Objective

• Assess ChatGPT's ability to answer causal-discovery questions in a medical domain (neuropathic pain).
• Analyze how ChatGPT answers questions about causal relationships using meta-information rather than observational data.
• Identify limitations, consistency issues, and language/concept understanding gaps in ChatGPT's responses.

Proposed method

• Create a benchmark by sampling 50 true causal pairs and 50 false pairs from a neuropathic pain causal map.
• Format questions as 'X causes Y. Answer true or false' using variable pairs.
• Query ChatGPT on whether the stated relationship is true or false.
• Analyze responses for precision, recall, and F-score.
• Inspect qualitative and quantitative errors to identify patterns of limitations.

Experimental results

Research questions

• RQ1 Can ChatGPT reliably identify true causal relationships in a neuropathic pain causal map from meta-information alone?
• RQ2 What are the common error modes (e.g., false negatives, language/region misinterpretation) in ChatGPT's causal judgments?
• RQ3 How consistent are ChatGPT's answers across repeated queries and across languages or terminologies?

Key findings

• ChatGPT shows high precision but low recall in the tested causal judgments.
• There are no false positives in the reported confusion matrix, but several false negatives reduce overall performance.
• Answers vary across days, indicating instability and potential model-update effects on responses.
• ChatGPT struggles with regional anatomy descriptions (e.g., T12 involvement) and with non-English terms, contributing to misclassifications.
• Some queries yield reasonable or non-trivial correct responses, suggesting potential complementary use for causal discovery tasks.
• Overall, current ChatGPT performance in causal discovery is limited and not reliable for definitive causal claims.