Seven Opportunities for Artificial Intelligence in Primary Care Electronic Visits: Qualitative Study of Staff and Patient Views

Unfamiliarity With Artificial Intelligence

For some patients, the term AI was off-putting and carried negative connotations. They felt AI to be intimidating and otherworldly. The suggestion of AI in the context of health care was therefore alarming:

Well, something like ET I think, twiddling a finger and thinking, what shall I tell this human? No, I just, no. I just…no (patient 23, practice 44, female, age 75 years).

Staff felt they knew little about AI and believed they had no experience in using AI technologies in their clinical practice. This led to some trepidation regarding the incorporation of AI, which was balanced with acknowledgment of their initial skepticism about the use of eVisits:

I think, probably like a lot of clinicians, I’m probably initially sceptical…it just comes back to safety. But then, if you’d have asked me a year ago, or a bit longer, what do I think about eVisits, again, I would have been so sceptical, but, you know, now I’m quite a fan (staff 13, practice 1, female, general practitioner [GP]).

Overestimation and Underestimation of Artificial Intelligence Capabilities

Unfamiliarity with AI led patients to overestimate the level of input and responsibility AI would have, believing that AI could diagnose and prescribe with no input from their primary care provider:

My concern would be, if it’s possible for the bot to straight away prescribe you medication or make a diagnosis (patient 4, practice 2, female, age 38 years)?

Patients feared that if AI proved too successful, it would eventually replace doctors and lead to replacement of the workforce:

One concern is that maybe some people, like they will lose their jobs, this is the only thing, that the machine will make the decision and maybe after we have more data and you teach the model better, so when the time comes that it makes knowledgeable decisions than a human, then it will replace the human (patient 7, practice 1, male, age 36 years).

Although few staff identified this as a real concern, they also expressed initial confusion as to what features of the eVisit could be automated. Some were therefore unable to specify in advance how they thought AI features could improve their workflow:

I suppose, I mean I don’t really know, I’m not entirely sure kind of what the ability of the AI is going to be (staff 7, practice 6, male, GP).

For staff and patients who could provide examples of AI, prior experiences they could recall were largely limited to frustrating encounters with online customer service chatbots or virtual assistants. Despite an initial overestimation of AI capabilities, interviewees described chatbots as rudimentary and unable to provide adequate and relevant information. These negative encounters contributed to concerns regarding the use of AI in health care, given the higher stakes involved:

I don’t like it from [energy supplier]. I wonder how I would find it in terms of my healthcare requirements (patient 10, practice 6, female, age 57 years).

Once the plain English explanation of AI had been presented (Supplemental Appendix 2), most participants were able to discuss the potential benefits and risks of AI during eVisits in more depth.

Workload Reduction

The potential to ease clinical staff workload was the most commonly cited benefit. Patients readily acknowledged the strain placed on primary care and believed that AI could make certain tasks more efficient and cost effective:

I’m extremely comfortable with it, and obviously I recognise the benefits of it, because we are an ageing population, and healthcare is becoming more and more expensive, and it is eventually going to become unaffordable unless we introduce technologies like AI (patient 11, practice 1, male, age 69 years).

Primary care staff hoped that AI would help address the perceived supply-induced demand¹⁴ witnessed since adoption of eVisits:

Because I think it’s good that patients can access GPs a bit easier, but hopefully…you’re dealing with the stuff that needs to be dealt with rather than wasting your time dealing with trivial stuff, you know, when someone else with a more urgent problem could be being dealt with…I think hopefully it would free us up but still sort the patients out (staff 5, practice 2, female, GP).

Safety Improvements

Some interviewees identified potential safety benefits from the incorporation of AI. They believed that AI would decrease the risk of human fatigue and error and welcomed its ability to access and process large amounts information quickly:

I could see perhaps a potential benefit from the wealth of information that would be available to it. It can assess things with a huge amount of information that has come in, can’t it. So it should be pretty accurate really, I would have thought (patient 18, practice 5, male, age 73 years).

Depersonalized Service

Several patients questioned if the inclusion of AI would lead to less in-person contact. Patients feared a depersonalized service and a disconnection between themselves and their primary care practice:

I feel uncomfortable if I’m being really honest. Knee-jerk reaction is I feel uncomfortable [about] any kind of AI because it’s…I just think certain things, you still need some face-to-face interaction, some interaction with another human (patient 15, practice 6, male, age 58 years).

Effect on Patient Use of eVisits

Staff were eager for patients to continue to use eVisits. They worried that AI incorporation could jeopardize patient acceptability:

So, I kind of think, taking that little bit away from them, they’ll then go back to the phones (staff 4, practice 3, female, receptionist).

However, some staff believed that any patient trepidation would be part of an inevitable transition period, similar to that observed after the adoption of eVisits. They believed that over time, patients would begin to trust a system with automated features, just as they had done with eVisits:

I think, initially, there possibly might be quite a lot of resistance from it…it might be extra work in the beginning, but once they start to spot a pattern, actually, no, the doctors are agreeing with what the AI is telling us, that might then disappear (staff 10, practice 5, female, GP).

Many patients were largely apathetic about the introduction of AI as long as their user experience was not adversely affected. They found eVisits an easy way to contact their primary care provider and were happy with the introduction of AI as long as they could still get help swiftly:

I mean, they’re going to get replies, aren’t they, still? And be treated exactly the same…Is there going to be a big difference if they do it (patient 5, practice 8, female, age 49 years)?

Safety Concerns

Both patients and staff expressed safety concerns, specifically with regard to the complexity of primary care practice. Worries stemmed from the belief that AI would miss nuanced issues, leading to delays in patients receiving correct help. Some clinicians confessed that they were not always confident as to how to prioritize requests based on the information received from eVisits currently and therefore were unsure about trusting AI, given that it lacked human instinct and knowledge of their individual patients’ behavior:

I think general practices are so complicated and sometimes I can’t even work out what’s routine and what isn’t based on an initial request. It just rings a lot of alarm bells for me (staff 9, practice 3, male, GP).

The requirement for rigorous and ongoing evaluation was a prominent theme for clinicians, along with the need for clarification of clinical responsibility and oversight. However, most believed that there was minimal risk to patient safety by automating specific tasks. Patients and staff felt reassured if eVisits had contact with a human at some point after submission. This would provide an opportunity for staff to intercept any errors (eg, an eVisit sent to reception instead of a clinician) and to escalate urgent eVisits as needed:

I think there’s obviously risk involved with trusting computers and AI, so we’d want to know that it was fit for purpose…I wouldn’t want it to just do something automatically in the background and then close something off without a human taking a look at that request too at some point in the process (staff 1, practice 3, male, GP).

Patients voiced concerns that the use of AI would increase the responsibility placed on them to input their symptoms accurately and in sufficient detail. They worried that omitting key words or using incorrect terminology could result in their request being deprioritized in error or directed to the wrong person, or even in a misdiagnosis. They believed that administrative staff triaging requests over the telephone, or a clinician triaging eVisits, would note the absence of key information and probe with further questions:

If someone was having a heart attack, say, but didn’t express themselves very well, and the AI thought, oh, it’s nothing much that and said just rest at home or something (patient 18, practice 5, male, age 73 years).

Few patients raised concerns pertaining to data security and the potential vulnerability of AI systems. Most dismissed fears of data misuse, citing trust in their primary care practice to have undertaken all necessary checks and precautions before the introduction of any new software:

Well, there’s the element of trust in everything, isn’t there? If you can’t trust your GP, you’re on a loser straight away (patient 17, practice 5, male, age 58 years).

Workflow

The ability of AI to direct eVisits to the most appropriate professional within the practice was suggested, to speed up triage decisions and ensure that requests needing clinical input were reviewed more quickly:

I certainly think from an AI point of view, things like allocating tasks to the right person would be quite useful for us…so taking out those steps where we have to kind of triage things to various clinicians. And perhaps if it learned…if somebody wanted a form completing it would go to…a member of staff or if somebody needed a prescription it would go to another member of staff, it might be quite useful (staff 1, practice 3, male, GP).

Directing

Directing patients away from the practice was also proposed. Requests containing emergency symptoms, such as chest pains or suicidal intentions, could be directed by AI to emergency services to improve safety. Staff also believed that directing eVisits to community pharmacies or supporting self-care would be particularly useful to relieve unnecessary pressure on primary care practices:

I think that’ll be really useful for quite a few tricky patients, that are repeatedly coming back with the same thing that’s not requiring a doctor or obviously the problems I was saying about over the weekend, signposting to the right services. I think it’ll be really useful for that (staff 5, practice 2, female, GP).

Prioritization

Clinicians believed that AI that could highlight urgent requests would help structure their time more effectively. This would support them to prioritize their workload, speed up triage decisions, and improve patient safety:

I think triaging urgent queries and putting them to the top of the list would be really useful. Because that is something that, at the moment, we’re having to do ourselves (staff 10, practice 5, female, GP).

Asking Questions

Patients voiced concerns regarding the use of AI to adapt questions during eVisit query submission, as used in some existing eVisit systems.⁸ They believed that it would lead to irrelevant or superfluous questions and feared that the system would no longer be quick and easy to use. However, staff suggested that AI could help gain more targeted information from patients after submission, which was also acceptable to patients. This could include sending relevant validated questionnaires (eg, 9-item patient health questionnaire for depression) based on the content of the request, prompting patients to upload photos (eg, of rashes), and supporting patients to add more information as necessary:

The computer system could say, right, send me a picture, rather than [me] having to telephone the patient and say, you know, your rash, send me a picture (staff 16, practice 2, male, advanced nurse practitioner).

Writing Assistance

Clinicians also felt that a significant proportion of their time was spent responding to common patient requests, for example mental health inquiries. They hoped that AI could recommend messages to send to patients that they could choose and edit based on their previous activity:

Sometimes there’s certain things that I find I’m writing over and over again. So just as an example, in the last month, the number of people that I’ve, had a discussion about mental health, gone through the options…I always have the same set advice…typing that out over and over again…it would save us a lot of time and it would save a lot of repetition in us writing the same thing (staff 10, practice 5, female, GP).

Providing Self-Help Information

The ability to provide patients with appropriate trusted information regarding their condition without interacting with a clinician was suggested by staff. This could decrease clinician workload by encouraging patient self-care and making it unnecessary for them to locate and send such information to the patient themselves:

If a patient, for example, mentions a particular condition, or a certain key word…for example, irritable bowel syndrome, or some sort of common condition like that, perhaps [AI] could flag up some already known resources, which are regulated and safe, for example, from the NHS website…they might find the answer that they’re looking for that way [without needing input from a clinician] (staff 13, practice 1, female, GP).

Face-to-Face Appointment Booking

Some clinicians hoped that AI could book face-to-face appointments for patients to speed up the process and avoid workload duplication. In addition, eVisits requiring a physical examination or that pertained to complex, vague, or multiple symptoms could be highlighted as requiring a face-to-face appointment and automatically arranged. This would avoid back-and-forth messaging or telephone calls with the patient, which result in multiple contacts for the same issue:

If it could automatically triage things that definitely need an [face-to-face] appointment, because actually that could be one way in which it could save some time, rather than having a telephone call from somebody to say, yes, you need to come in…if they needed a pill check and they had to have a blood pressure [check]…could it direct them to an appointment where we have some certain slots for the pill checks (staff 7, practice 6, male, GP)?

eVisit System Developers and Designers

Our results suggested the following 3 new AI-powered features that could be incorporated into future eVisit system designs that have not previously been reported (Figure 1): (1) asking questions (eg, sending patients detailed questionnaires about symptoms), (2) writing assistance (eg, recommending editable templates based on previous responses), and (3) providing self-help information (eg, directing patients to relevant websites). System designers could build these new features into their eVisit systems to evaluate their utility in clinical practice.

Our findings suggest that the use of AI to adapt questions posed to patients during eVisit query submission might not be acceptable to all patients. Patients had negative associations with online chatbots and believed that using AI to determine questions would be encroaching on the core doctor role. Patients were reluctant to alter the user experience of this particular eVisit system, which consisted of a few simple questions and free-text responses. A combination of multiple-choice questions and free-text responses could be considered to ensure that sufficient information is captured for AI to workflow, direct, and prioritize requests effectively. This will also address patient concerns that the use of AI could increase their responsibility to input symptoms accurately and in sufficient detail.

Future system development could include the integration of internet of things devices to supplement information received during eVisits, which could be analyzed by AI. Examples include physiologic data from wearables such as smart watches and devices that enable remote physical examinations (eg, TytoCare [www.tytocare.com]; a handheld exam kit and app that connects patients with clinicians for on-demand medical exams).

Primary Care Staff

Primary care staff should be reassured that patients are open to the use of AI during eVisits once their initial anxieties have been addressed, for example, by providing reassurance that AI will not replace clinical input from their primary care clinician and emphasizing that face-to-face appointments will still be available if necessary. Highlighting expected benefits for patients, such as improved patient safety and the ability to speed up the process of resolving their eVisit, is also important to ensure sustained use by patients.

Researchers

This study investigated preadoption views of AI during eVisits. Future research should develop and implement the suggested AI features into eVisits in practice. Discussions with staff and patients should aim to establish key performance criteria for each AI feature to help inform eVisit system design. Development and evaluation of AI features has already begun in the Patchs system, and articles are currently being prepared for publication. Given that participants in the present study were current users of eVisits, more research is required to determine if any particular groups, such as racial and ethnic minority groups or those less confident with technology, are more hesitant regarding the incorporation of AI.