Article: AI-Generated Draft Replies Integrated Into Health Records and Physicians’ Electronic Communication

Electronic healthcare record (EHR) systems are a leading cause of physician fatigue and burnout. EHR patient-to-provider inbox messaging, which gained popularity during the COVID-19 pandemic, has further contributed to and compounded this problem. In recent years, generative artificial intelligence (GenAI) and large language model (LLM) based chatbots, such as Open AI’s Chat GPT and Microsoft’s Bing Chat have demonstrated the ability to synthesize large volumes of information and generate relevant, intelligible text. In this quality improvement study, published in JAMA Network Open, Small and colleagues aimed to see if GenAI can aid in responding to patient EHR messages by comparing provider and GenAI-generated drafts.  

The authors identified in-basket patient messages (medical advice, medication refill requests, laboratory result inquiries, and paperwork requests) from their institution’s outpatient internal medicine clinics that fit the study’s inclusion criteria. An EHR-compatible version of Open AI’s GPT-4 was used to draft response messages. Sixteen primary care providers (PCPs) were recruited and together reviewed 344 message-response pairs. Of the 344 pairs, 175 responses were drafted by GenAI models, and 169 responses were actual responses to patients’ messages by the healthcare provider (HCP). The PCPs were tasked with evaluating each response’s content quality (accuracy, completeness, relevance) and communication quality (understandability, tone, verbosity) using a 5-point Likert scale (1 indicates strongly disagree, while 5 indicates strongly agree). The PCPs were also asked to evaluate if the response drafts were usable or unusable for patient communication.

The authors found no differences in information content quality between responses generated by GenAI and HCPs (mean [SD], 3.53 [1.26] vs 3.41 [1.27]; p=0.37). Interestingly, subgroup analysis of responses rated as having inadequate content quality demonstrated that AI-generated responses were more likely to include irrelevant content, while HCP-generated responses were more likely to be incomplete. GenAI drafts performed particularly poorly when answering questions related to laboratory results. In terms of communication quality, PCPs rated GenAI responses higher than HCP-written responses (mean [SD], 3.70 [1.15] vs 3.38 [1.20]; p=0.01). Subgroup analysis of responses rated as poor for communication quality found that HCP responses were more likely to have inappropriate tone and jargon, while AI responses were more likely to be verbose. Overall, GenAI and HCP-generated responses were found to be useful at similar rates (0.69 [0.48] vs 0.65 [0.47]; p=0.49). PCPs noted drafts were useful because it would take them less time to edit an existing draft compared to starting from scratch.

Importantly, while some PCP raters were able to identify GenAI drafts, they more frequently classified GenAI drafts as personalized (45 [52.3%] vs 30 [38.0%]) and empathetic (32 [37.2%] vs 13 [16.5%]). Linguistic analysis of the drafts supported this finding as GenAI drafts were more likely to use positive language. However, the linguistic analysis found that GenAI drafts were 38% longer (not statistically significant) and required higher educational levels to understand (mean [SD] Flesch-Kincaid grade level, 8.1 [1.6] vs. 6.5 [2.3]; p<0.001).

This study by Small et al. demonstrates that generative AI technology may prove a useful tool in reducing EHR-related messaging and documentation burden. GenAI drafts demonstrated similar rates of usability compared to HCP drafts. Importantly, AI chatbots are not affected by work-related fatigue, which may explain why the AI-generated drafts were found to be more empathetic compared to HCP responses. Further fine-tuning and training of the chatbots may be able to address some of the shortcomings of AI-generated drafts, including verbosity and information relevance. The chatbots can be instructed to limit length and reduce the lexical complexity of their responses. In addition, model training may aid the models in reducing irrelevant information in their responses.

Like any new technology, particular attention and care must be paid to ensuring data and protected health information security. Healthcare institutions, including this study’s institution, are already addressing this concern by securely integrating AI chatbots and LLMs into their EHR systems. Using such technology, physicians may one day be able to use GenAI-drafted for patient communication and EHR notes, saving them time and reducing EHR-related burnout.

SES Journal Club
Anthony DiGiorgio, DO, MHA
Attending physician

GenAI is an important use case for this rapidly emerging technology.  Student Doctor Ambati provides a wonderful review of the piece.  However, I’m concerned that the original manuscript is another case of losing the forest for the trees.  This article is important but fails to discuss a few salient points.

First, we should be examining what led to this unmanageable inbox load for physicians and the concurrent dissatisfaction of patients.  This doesn’t happen as often with independent private practices.  The success of the direct primary care (DPC) model, both with patients and physicians, shows that practice models exist which can minimize the inbox overload.  When physicians have agency and control over the practice, there is little need for a generative AI to answer inbox messages.  This underlying issue, restoring physician independence, must be addressed rather than patching the problem with generative AI.

Which brings us to the next point: we are missing the true potential of AI for an inbox answering gimmick.  Large language models excel at managing dispersed, unstructured data.  This potential make these models perfect for providing clinical summaries, selecting billing codes, answering risk adjustment queries, handling prior authorizations, simplifying order entry and improving processes in hundreds of other annoyances which are slowly destroying the practice of medicine.  I’ll get excited when JAMA publishes a study showing an AI chatbot that takes a clinical interaction and automates the note, order entry, billing codes and prior authorization.  I’m hoping that day is soon.