Abstract
AbstractWe investigated the ability of large language models (LLMs) to answer anesthesia related queries prior to surgery from a patient’s point of view. In the study, we introduced textual data evaluation metrics, investigated “hallucinations” phenomenon, and evaluated feasibility of using LLMs at the patient-clinician interface. ChatGPT was found to be lengthier, intellectual, and effective in its response as compared to Bard. Upon clinical evaluation, no “hallucination” errors were reported from ChatGPT, whereas we observed a 30.3% error in response from Bard. ChatGPT responses were difficult to read (college level difficulty) while Bard responses were more conversational and about 8thgrade level from readability calculations. Linguistic quality of ChatGPT was found to be 19.7% greater for Bard (66.16 ± 13.42 vs. 55.27 ± 11.76;p=0.0037) and was independent of response length. Computational sentiment analysis revelated that polarity scores of on a Bard was significantly greater than ChatGPT (mean 0.16 vs. 0.11 on scale of −1 (negative) to 1 (positive);p=0.0323) and can be classified as “positive”; whereas subjectivity scores were similar across LLM’s (mean 0.54 vs 0.50 on a scale of 0 (objective) to 1 (subjective),p=0.3030). Even though the majority of the LLM responses were appropriate, at this stage these chatbots should be considered as a versatile clinical resource to assist communication between clinicians and patients, and not a replacement of essential pre-anesthesia consultation. Further efforts are needed to incorporate health literacy that will improve patient-clinical communications and ultimately, post-operative patient outcomes.
Publisher
Cold Spring Harbor Laboratory
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献
1. Analyzing the Future of ChatGPT in Medical Research;Artificial Intelligence Applications Using ChatGPT in Education;2023-09-15