Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum

“Some patient messages are unsolicited questions seeking medical advice, which also take more skill and time to answer than generic messages (eg, scheduling an appointment, accessing test results). Current approaches to decreasing these message burdens include limiting notifications, billing for responses, or delegating responses to less trained support staff. Unfortunately, these strategies can limit access to high-quality health care. For instance, when patients were told they might be billed for messaging, they sent fewer messages and had shorter back-and-forth exchanges with clinicians. Artificial intelligence (AI) assistants are an unexplored resource for addressing the burden of messages. While some proprietary AI assistants show promise, some public tools have failed to recognize even basic health concepts.

ChatGPT represents a new generation of AI technologies driven by advances in large language models. [..] The system was not developed to provide health care, and its ability to help address patient questions is unexplored. We tested ChatGPT’s ability to respond with high-quality and empathetic answers to patients’ health care questions, by comparing the chatbot responses with physicians’ responses to questions posted on a public social media forum. [..]

The original question, physician response, and chatbot response were reviewed by 3 members a team of licensed health care professionals working in pediatrics, geriatrics, internal medicine, oncology, infectious disease, and preventive medicine [..]. The evaluators were shown the entire patient’s question, the physician’s response, and chatbot response. Responses were randomly ordered, stripped of revealing information (eg, statements such as “I’m an artificial intelligence”), and labeled response 1 or response 2 to blind evaluators to the identity of the author. [..]

The sample contained 195 randomly drawn exchanges with a unique member-patient’s question and unique physician’s answer. The mean (IQR) length of patient questions in words averaged 180 (94-223). Mean (IQR) physician responses were significantly shorter than the chatbot responses (52 [17-62] words vs 211 [168-245] words; t = 25.4; P < .001). A total of 182 (94%) of these exchanges consisted of a single message and only a single response from a physician. [..]

The evaluators preferred the chatbot response to the physician responses 78.6% (95% CI, 75.0%-81.8%) of the 585 evaluations. [..]

Evaluators also rated chatbot responses significantly higher quality than physician responses (t = 13.3; P < .001). The mean rating for chatbot responses was better than good (4.13; 95% CI, 4.05-4.20), while on average, physicians’ responses were rated 21% lower, corresponding to an acceptable response (3.26; 95% CI, 3.15-3.37). The proportion of responses rated less than acceptable quality (<3) was higher for physician responses than for chatbot (physicians: 27.2%; 95% CI, 21.0%-33.3%; chatbot: 2.6%; 95% CI, 0.5%-5.1%). This amounted to 10.6 times higher prevalence of less than acceptable quality responses for physicians. Conversely, the proportion of responses rated good or very good quality was higher for chatbot than physicians (physicians: 22.1%; 95% CI, 16.4%-28.2%; chatbot: 78.5%; 95% CI, 72.3%-84.1%). This amounted to 3.6 times higher prevalence of good or very good responses for the chatbot.

Chatbot responses (3.65; 95% CI, 3.55-3.75) were rated significantly more empathetic (t = 18.9; P < .001) than physician responses (2.15; 95% CI, 2.03-2.27). Specifically, physician responses were 41% less empathetic than chatbot responses, which generally equated to physician responses being slightly empathetic and chatbot being empathetic.

[..] as tested, chatbots could assist clinicians when messaging with patients, by drafting a message based on a patient’s query for physicians or support staff to edit. This approach fits into current message response strategies, where teams of clinicians often rely on canned responses or have support staff draft replies. Such an AI-assisted approach could unlock untapped productivity so that clinical staff can use the time-savings for more complex tasks, resulting in more consistent responses and helping staff improve their overall communication skills by reviewing and modifying AI-written drafts.

In addition to improving workflow, investments into AI assistant messaging could affect patient outcomes. If more patients’ questions are answered quickly, with empathy, and to a high standard, it might reduce unnecessary clinical visits, freeing up resources for those who need them. Moreover, messaging is a critical resource for fostering patient equity, where individuals who have mobility limitations, work irregular hours, or fear medical bills, are potentially more likely to turn to messaging. High-quality responses might also improve patient outcomes. For some patients, responsive messaging may collaterally affect health behaviors, including medication adherence, compliance (eg, diet), and fewer missed appointments. Evaluating AI assistant technologies in the context of randomized clinical trials will be essential to their implementation, including studying outcomes for clinical staff, such as physician burnout, job satisfaction, and engagement.”

Full article, JW Ayers, A Poliak, M Dredze et al., JAMA Internal Medicine, 2023.4.28