GPs urged to consider the benefits of generative AI

It is no secret that Australia is facing a growing shortage of general practitioners (GPs), particularly in rural and remote communities and in underserved outer metropolitan populations. This shortage is putting more pressure on the GPs who remain, who also face increased practice costs. This is compounded in regional areas by workforce shortages in primary care and indeed across the whole health workforce. Given current trends, Australia will have a deficit of over 10 000 GPs by 2031.

The appropriate use of artificial intelligence (AI), and generative AI (GenAI) in particular, could be part of the answer to alleviating some of these pressures. Existing GPs are leaving the profession and 13% fewer medical students are choosing to go into general practice. Burnout and poor career progression options are often mentioned as reasons for GP dissatisfaction. One of the main elements contributing to burnout is interacting with electronic health records systems and overspending time on administrative tasks instead of devoting time to patients and clinical encounters. Other sources of frustration for GPs worldwide include longer working hours and a lack of a supportive work environment.

This makes the intelligent use of GPs and their limited time a priority. After all, primary health care provision is still the most efficient and sustainable form of providing health care.

In this context, we believe that the emergence of GenAI could have a transformational role in general practice and help ensure it has a future.

How AI could reshape primary care

With its capabilities in text generation, text processing and “understanding”, conversational AI systems such as ChatGPT should  be especially useful for clinical documentation tasks. A “digital scribe” that automates much of the documentation burden for GPs could indeed mitigate some of the reasons behind clinician burnout. This change is already happening, with several software vendors offering GenAI software to automate parts of the clinical note-taking processes.

Although documentation automation is one of the most obvious uses, this is not the most transformative one on the horizon. GenAI may truly reshape how care is delivered, especially in primary care. The appropriate use of GenAI systems could facilitate care more easily in remote areas, addressing understaffing with administrative task automation. The AI-assisted physician could address and mitigate some of the eternal tribulations of primary care, such as managing diagnostic uncertainty.

It is in this reshaping that we need to concentrate our efforts on imagining, developing and testing the future of “augmented” primary care. What do we mean by augmented? Previous work has highlighted the need to develop “learning health systems”. In a similar light, the augmented consultation could mean a departure from the usual patient–doctor–computer interaction, where, generally, there are discrete flows of information. Instead, a continuous feed from the patient to their AI assistant that then feeds and summarises this information for the doctor may appear. This flow of information won’t end with the consultation itself, as continuous ongoing exchanges between AI agents from patients and doctors will continue to provide information and guide health management.

We could argue then, that most of the primary care will indeed happen outside the consultation room itself. Likewise, we could envision an “embodied AI” consultation room, that listens and sees patient cues and provides them in concise and interpretable ways to physicians or, potentially, to other AI agents. Moreover, the record itself would be transformed into something else completely. Think of a chat interface, that either by text or speech will bring relevant patient information on the fly. You will ask with your voice, or type in a box, “Show me the last HbA_1c measurements” and a graph with the values will appear. Type a different request and the record will change. In this sense, the “AI-embodied” consultation room may not even have a computer screen or keyboards but resemble more of a comfortable lounge room, where doctors and patients interact directly, while all the other elements (text, clinical history, measurements) are subdued into the background. Much of this has been foreshadowed and now GenAI has the potential to make it happen.

AI as a training tool

GenAI systems are now multimodal, meaning they can also process and generate other sources such as sound, image or even video. The current developments of multimodal AI systems will likewise allow in-real-time or near-real-time feedback from patient and doctor communication. Think of receiving cues from how the clinical conversation is going, if the patient seems undecided about treatment or if the tone you are using to communicate information is not adequate. For instance, you could receive hints on what “stage of change” a patient is for quitting smoking based on non-verbal cues you might not have picked up on. It may also suggest follow-up questions to reach a diagnosis.

Although this real-time flow can be useful initially in training scenarios, GPs could also use this as a continuous learning and development tool. Here they would review (with the help of a GenAI assistant) past conversations and cases, practising to improve diagnostic or communication skills.

In this light, AI is also going to be decisive for medical education — akin to airline pilots, who must spend a considerable number of hours on simulators before getting “behind the yoke”. GenAI should allow students and future clinicians to develop their skills before they see their first real patient, allowing for better trained clinicians (eg, extensive training on giving bad news before facing real patients).

Caution is still required

Although there is indeed a bright future ahead, we should not race to a mindless use of these technologies. To continue with transportation analogies, just as cars require multiple quality tests before they hit the market, AI will need robust testing before it can hit the “medical roads”, and users will need training. We may need to qualify AI clinical users to know how to prompt and interact with this novel form of computation so, for instance, they won’t be confused by anomalous outputs. While it is likely that the most salient problems such as fabrication of answers (hallucination) will be overcome over the next years, GenAI departs from previous forms of deterministic computing (meaning that you will always get the same output for a given input) and into a more conversational style of interaction with computers. Although variation in expression (articulating something in different words) by GenAI is acceptable, we need to make sure that these variations don’t entail erroneous judgments. We need to make sure that GenAI variability falls within its expertise so a trusting relationship with clinicians is established and they can see the AI’s output akin to a peer.

Beyond primary care, there is a huge opportunity to rethink health care workflow from operating theatre to hospital switchboards. GPs need to be at the centre of this revolution as their expert generalism is going to be key to understanding, filtering and acting on the vast capabilities of GenAI.

We believe there is a bright future ahead, and we need to engage with it continuously. That enthusiasm should always be tempered with rational caution because we do not wish to harm for innovation’s sake. This is primary care’s chance to rethink what it means to, and how we provide care for, patients and the community.

Professor Enrico Coiera is the Director of the Centre for Health Informatics at the Australian Institute of Health Innovation, Macquarie University. Professor Coiera trained as a medical doctor and leads the Australian Alliance for AI in Healthcare

Dr David Fraile Navarro is Postdoctoral Research Fellow on Generative AI at the Australian Institute of Health Innovation, Macquarie University, and a trained general practitioner.