This site is intended for health professionals only

How artificial intelligence is changing the GP-patient relationship

‘Alexa, what are the early signs of a stroke?’

GPs may no longer be the first port of call for patients looking to understand their ailments. ‘Dr Google’ is already well established in patients’ minds, and now they have a host of apps using artificial intelligence (AI), allowing them to input symptoms and receive a suggested diagnosis or advice without the need for human interaction.

And policymakers are on board. Matt Hancock is the most tech-friendly health secretary ever, NHS England chief executive Simon Stevens wants England to lead the world in AI, and the prime minister last month announced £250m for a national AI lab to help cut waiting times and detect diseases earlier. Amazon even agreed a partnership with NHS England in July to allow people to access health information via its voice-activated assistant Alexa.

Little surprise then that private developers see now as a good time to develop AI to guide patients through their various ailments. Babylon last month announced a £450m R&D investment, partly for AI technology to manage chronic conditions, while the likes of Ada and Your.MD also offer patients the chance to check symptoms. This is on top of the NHS App’s own symptom checker.

Yet the evidence in support of algorithms – and AI – is still lacking, and a Pulse analysis has shown potential drawbacks, such as overreaction to mild conditions and potentially unsafe advice.

If patients are using symptom-checker app I would worry the app will give a false positive. Even worse, a false negative

Dr Rebecca Fisher

Dr Rebecca Fisher, a GP and senior policy fellow at the Health Foundation says: ‘If patients are using symptom-checker apps, I would have two main worries. The first is that the app will give a false positive, with the risk that the patient becomes anxious and also potentially generates unnecessary use of NHS resources.

‘Even worse, there a risk of an app giving a false negative, meaning you might not seek help you actually need.’

Dr Nick Mann, a London-based GP with an interest in AI, says he is already seeing this sort of impact: ‘People will come in with headache and be convinced they’ve had a brain bleed whereas I know, talking to them, they haven’t.

‘I’ve had a lot of requests in the past couple of years, which I never used to have, from people wanting investigations for symptoms they have diagnosed on Google, which are inappropriate.’

With this in mind, Pulse tested some of the available symptoms checkers. We found the apps were successful in offering appropriate advice in the case of a heart attack, but problems also emerged. In one case, a 26-year-old female with acute pyelonephritis was told her condition would clear up on its own.

Dr Roger Henderson, a sessional GP who is also medical director of Liva Healthcare, a digital healthcare company that supports the management of patients with diabetes and who tested the apps for Pulse, says: ‘In this tiny snapshot there are worrying features where everyday complaints were marked as emergencies and potentially severe ones were underplayed.

‘Symptom checkers use a linear algorithm approach and depend on the information provided to them, rather than being able to follow the more nuanced process that GPs use. It is this black-and-white computer reasoning that causes problems, since diagnosis tends to be shades of grey in the real world.’

He says the fact that symptoms checkers encourage people to include all symptoms to give the fullest possible picture can lead to anxiety: ‘If you give a patient a range of diagnoses ranging from minor to very serious, it is natural to focus on the serious even if this is incorrect, causing worry and anxiety.’

This black-and-white computer reasoning….causes problems, since diagnosis tends to be shades of grey in the real world

Dr Roger Henderson

Lincolnshire GP Dr Phillip Williams, who also tested the apps for Pulse, agrees patients don’t always present as textbook cases. ‘Often real patients don’t present with the symptoms we think they should. As these apps become more sophisticated, they may flag key symptoms which aren’t on our radar. For example, we’re taught motor neurone disease presents with fasciculations, whereas, in real life, a common first symptom is fatigue.’

The shortage of relevant research is a problem for many GPs (see box). Dr Benjamin Brown, a senior academic GP and health informatician in Manchester, says: ‘The NHS should only bring in routine care systems that have an evidence base. In the case of model-driven triage, the models may be too conservative. I have anecdotally heard that one of the well-known providers modified its algorithms over concerns about patient safety, which resulted in it sending many more patients to A&E.’

Perhaps the highest-profile patient-facing algorithm is NHS Pathways, used by NHS 111. A 2013 study found NHS 111 increased emergency and urgent care activity by 5-12% each month, while emergency ambulance incidents rose by 2.9%.1

Is there any evidence to support AI in healthcare?

A 2013 study by the University of Sheffield1 revealed that NHS 111 increases ambulance and urgent and emergency care use. It looked at 400,000 calls, including 277,163 triaged using NHS Pathways, and found emergency ambulance incidents rose by 2.9%. It estimated this could mean an additional 14,500 call-outs for a service attending 500,000 incidents a year. In addition, emergency and urgent care activity rose by between 5-12% per month.

The study concluded: ‘The findings reflect the inherent characteristics of the NHS Pathways system such as the levels of caution and risk built into the assessment algorithms, particularly as it is designed to be used by non-clinical call handlers. There may be less flexibility to change decisions compared with assessments made by nurses and it is possible that a different call assessment system could produce different results.’

A 2015 evaluation by Harvard Medical School3 found 23 symptom checkers for self-diagnosis provided the correct diagnosis first in 34% of 45 standardised patient evaluations, listed the correct diagnosis within the top 20 diagnoses given in 58% and provided appropriate triage advice in 57% of cases. It said: ‘Overall they had deficits in both diagnosis and triage accuracy. The risk-averse nature of symptom checkers’ triage advice is a concern. In two-thirds of evaluations where medical attention was not necessary, we found symptom checkers encouraged care.’

• A 2017 evaluation by NHS England4 found patients had a very good experience of triage and assessment tools including the digital version of NHS Pathways in West Yorkshire (web interface), system in West Midlands (voice-activated avatar), Expert 24 in Suffolk (web interface) and Babylon in London. As a result of their use, fewer people were directed to primary care services and more turned to self-management than from NHS 111.

A 2018 study by Babylon5 showed the company’s triage and diagnostic system was able to identify patient conditions modelled by a clinical vignette with accuracy comparable with doctors’, in terms of precision and recall, and was on average safer than doctors. The findings, based on the MRCGP examination, showed above-average pass marks. Yet the paper was not peer reviewed, and the research team included Babylon employees.

NHS England has introduced more clinicians into the call centres but, according to 616 GPs surveyed by Pulse, an average GP still receives around six inappropriate referrals from NHS 111 a month – totalling more than three million a year. Anecdotally, GPs say they are still seeing patients referred to them for dental problems. And last month, a coroner said the lack of flexibility within the algorithm should be addressed following the death of a 17-year-old boy, whom the coroner said may not have understood what he was being asked.

Harry Longman, founder of Askmygp – an online triage and consultation tool for GPs – says: ‘We don’t use any AI or algorithms to triage automatically, we have tried that and found it doesn’t work. Many questions were irrelevant or difficult for patients, and the resulting output was not that helpful for clinicians.’

The Medicines and Healthcare products Regulation Agency says if an app is intended to influence treatment or results in a diagnosis or prognosis including future disease risk then it is a device and should obtain a CE mark before use. New EU rules, taking effect next year, will introduce more stringent requirements for device manufacturers.

We don’t use any AI or algorithms to triage automatically, we have tried that and found it doesn’t work

Harry Longman

But, as Professor Brendan Delaney, chair in medical informatics and decision- making at Imperial College London, puts it: ‘The letter of the regulation is fine, but it relies on developers to self-certificate and register – which is OK, provided entry to the market place is actually policed and purchasers insist on CE marking.’

There are positives. AI is being developed to help target patients for screening, and help doctors make decisions – uses few would argue with.

And the Topol Review2 commissioned by Mr Hancock to explore how the healthcare workforce will ‘deliver the digital future’, concluded that ‘early benefits of AI and robotics will include the automation of mundane repetitive tasks that require little human cognitive power, improved robot-assisted surgery and the optimisation of logistics.’ This would allow the workforce to focus on ‘interaction and care’.

However, the first signs are that AI will, at best, increase GP workload. It might be time for the Mr Hancock to review his championing of this new technology.

Dr Fisher says: ‘My patients often need a safe space to feel listened to so I don’t think AI is going to be a replacement for a clinician. It’s more of an add on to us.’


1 Turner J et al. Impact of the urgent care telephone service NHS 111 pilot sites: a controlled before and after study. BMJ Open 2013;3:e003451

2 Topol E. Preparing the healthcare workforce to deliver the digital future. Health Education England. February, 2019.

3 Semigran H. Evaluation of symptom checkers for self diagnosis and triage: audit study. BMJ 2015;351:h3480

4 NHS England online evaluation, December 2017.

5 Razzaki S et al. A comparative study of artificial intelligence and human doctors for the purpose of triage and diagnosis. June, 2018