Can Dr. ChatGPT improve medical decision making without introducing bias?

When it comes to the use of generative artificial intelligence in medical decision making, a new study suggests these tools have the potential to improve medical care without amplifying healthcare disparities.As described in their paper published in Communications Medicine on March 4 researchers from Stanford University invited 50 U.S. physicians to review one of two videos. The choice of videos was randomized, featuring either a white male patient or a Black female patient describing their chest pain symptoms. Both videos had the following script:“I’m so glad you were able to see me this afternoon. Ever since I retired a few years ago at 65, I’ve had time to try to get healthier. I know I’m overweight, so I’ve started to exercise more. After my walk this morning, I noticed a weird, tired feeling that made me feel a little short of breath. I sat down in my kitchen to get a sip of water and rest; it felt better a few minutes afterwards. I also felt fine when I walked up the stairs to your office. The medical assistant who took my vital signs said everything looks great, and I’ve been taking the blood pressure and cholesterol medication every day. So, I don’t think it’s a big deal, but I want to make sure since my dad had a heart attack in his early 60s.”Read: The great AI scribe rolloutThe physicians were subsequently asked to complete a four-question multiple-choice quiz on their clinical judgement of the patient in the video, with questions asking about risk assessment, triage, treatment. After providing their responses, the physicians were presented with a pre-prepared response from ChatGPT to the questions and allowed to interact with the AI tool, before being given the opportunity to modify their answers.Before consulting ChatGPT, the physicians who observed the video of the white male had an average score of 47% in their clinical decision accuracy, while those who watched the video of the Black female scored an average of 63%. However, after the AI intervention, the physicians in the white male group increased their average score to 65% while the score for the Black female group increased to 80%. For both groups, that marks an increase of around 18%.“This improvement in decision-making happened equally for all patients, regardless of their race or gender, suggesting AI tools could help improve medical care without increasing bias,” the researchers wrote. The researchers say the low initial scores are mainly due to how the physicians answered the first question, which asked how they would do an initial triage for the patient. Only 12% in both groups gave the correct answer (“Full dose aspirin and refer to the emergency department”). Meanwhile, 26% answered, “Start a daily baby aspirin and provide clear return precautions. Schedule the patient to return in one week” and 62% answered, “Start a daily baby aspirin and refer the patient for an urgent stress test within 2–3 days,”—both of which are considered under-triaged.Read: Is AI just a fad? Unravelling its real potential in medical practice“Though not specifically evaluated in this study, some of the reasons that participants commented on for initially under-triaging included reassurance by the patient’s symptom resolution and the feeling of a reasonable follow-up plan in alternative answers,” the authors wrote.The researchers also emphasized that their results shouldn’t be used to conclude the “general capability” of AI in medical decision making or bias in general, but that these findings show a “path forward for AI clinical decision support.”“Within a controlled standardized vignette around cardiac chest pain assessment, a large language model AI system was able to significantly alter physician medical decisions, improving accuracy without introducing or exacerbating existing demographic bias,” the researchers said in the study.

Can Dr. ChatGPT improve medical decision making without introducing bias?

Register for free access or login.

Related Topics