Primary care providers (PCPs) face numerous challenges, including time constraints and burnout, and are often limited in their ability to collect detailed medical histories from their patients [1]. Currently, patients’ personal and family medical histories are collected via paper forms or interviewing prior to and/or during a patient visit, with manual data entry into the electronic health record (EHR). Often, data are missing or of poor quality due to lack of time or lack of training [2].

Information gaps can lead to errors or delays in diagnosis and failures to address actionable risk factors, which affect an estimated 12 million Americans each year [3]. Diagnosis in primary care is a high-risk area for errors, for several reasons. PCPs typically face high patient volumes, make decisions amid uncertainty, and must balance the risk of missed or delayed diagnoses with the stewardship of scarce resources [4]. Poor communication leading to gaps in information sharing is a major driver of diagnostic errors in primary care settings [4], while engaging and empowering patients in the task of generating data using technology aims to close those gaps and support personalized medicine [5].

Recent advances in conversational agents powered by artificial intelligence (AI) and natural language processing show promise in augmenting current human-driven methods of collecting personal and family histories as part of pre-visit planning [6]. Various types of conversational agents have emerged, including chatbots, embodied conversational agents (avatars), and voice assistants, all of which mimic human conversation using text and/or spoken language [7]. Within health care, these agents have been used for facilitating screening for health conditions, triage, counseling, self-management of chronic conditions, and training for health care professionals; reported benefits have included their potential to support populations with poor health literacy, be scaled to reach large populations, and improve patient engagement [8]. However, such tools remain largely unproven in real-world settings, and few studies have assessed the use case of collecting patients’ information before an appointment to provide tailored counseling [2,9-11]. This study aimed to assess the feasibility and acceptability of a conversational AI medical interviewing system from the perspective of primary care patients.

Twenty patients with a mean age of 50 years completed an interview with the AI system, including 12 females (60%) and 8 males (40%); 11 were White (55%), 8 were Asian (40%), 1 was Black (5%), and 19 had at least a bachelor’s degree (95%) (Multimedia Appendix 1).

The majority of participants agreed that using a conversational AI medical interviewing system to collect histories could help PCPs have a better understanding of their health (16/20, 80%) and help them stay healthy through identification of their health risks (14/20, 70%) (Figure 2). Participants who felt the system was easy to use tended to be younger. The median age for those who agreed that they were able to learn the system quickly was 41 years compared to 61 years for those who disagreed and 41 years for those who agreed that the system was clear and understandable compared to 68 years for those who disagreed (Multimedia Appendix 1). Those who reported that the tool could motivate them to share more comprehensive medical information with their PCPs tended to be older—median age of 52 years for those who agreed and 36 years for those who disagreed (Multimedia Appendix 1). Patients were split on the effort required to use the tool, and on whether AI should be used for medical interviewing (Figure 2).

In this feasibility and acceptability pilot of a conversational AI medical interviewing system, the majority of primary care patients believed that such a tool could help PCPs better understand their health and identify their health risks. However, while performance expectancy of the system was predominantly positive, results were mixed in terms of effort expectancy and attitude toward the emerging technology.

Our results are aligned with existing literature. A recent systematic review examining the effectiveness and usability of AI-based conversational agents in health care found that 67% of the 31 studies reported positive or mixed evidence supporting the effectiveness, usability, and positive user perceptions of the agents [8]. Additionally, in a study exploring the acceptability and feasibility of a virtual counselor to collect family health histories in an underserved population, a vast majority of participants found the virtual counselor easy to use and understood the questions being asked [2]. At the same time, studies reporting qualitative feedback have consistently cited the following as barriers that will need to be addressed before conversational agents can be deployed and used widely: agents having difficulty understanding users, agents being repetitive and not sufficiently interactive, and users having difficulty forming connections with the agent [8]. Our study adds to this growing body of evidence by assessing the use of conversational agents in a slightly older population in primary care.

Our study has several limitations including the small sample size and convenience sampling recruitment approach. Those who opted to participate may have had more positive notions about new technologies and may have evaluated the system more favorably than the general population. Moreover, participants in our study were highly educated, which may limit the generalizability of our findings.

The use of conversational agents to gather more detailed information about patients’ personal medical history, family medical history, and social determinants of health could aid in capturing a more holistic view of patients and identifying disease risks earlier. Our findings suggest areas for further research, such as understanding the user interface factors that influence ease of use and adoption, and the reasons behind patients’ dichotomous attitudes toward AI-assisted history-taking.