Call to action on medical AI

‘Why do we see so few machine learning tools being used in medical practice?’
 
That was how research scientist Dr Martin Seneviratne began his GP19 plenary speech on the implementation gap in medical artificial intelligence – AI.
 
Why, he asked, given the huge amount of research into this space, are doctors not currently able to use machine learning?
 
Dr Seneviratne argued that researchers and doctors should not lose sight of the end goal with AI.
 
‘Any technology must help the clinician to help the patient. Technology must be fit for clinical purpose,’ he said.
 
The Google clinical informatician then outlined key challenges in his GP19 plenary speech.
 
‘Will AI mean the end of doctors? Most researchers couldn’t disagree more,’ he said. ‘Instead, the view is doctors who use AI will replace doctors who don’t.
 
‘What I like about that is the idea that it’s a tool to be used, just like an MRI. When they came out, they were forecast to replace radiologists – instead, it spawned a new discipline.’
 
Dr Seneviratne said that AI could outperform humans in a few narrow areas – including image-based diagnostics. He listed advances in areas such as augmented reality microscopes, where AI contours around suspicious lesions in real-time.
 
He said symptom-checking tools could be very helpful, particularly in low-resource settings with minimal healthcare available.
 
But Dr Seneviratne pointed to a major implementation gap between the promise and early tests of medical AI tools and their widespread deployment.
 
‘There’s a huge leap from high-performing algorithms in papers to going into practice,’ he said.
 
‘In the last 12 months, there have been over 3800 papers published in this area – but only a small minority have reached deployment.’
 
Dr Seneviratne said there is a key opportunity for GPs to be involved.
 
‘Clinicians can play a pivotal role, highlighting clinical pain points needing to be solved,’ he said.
 
Another challenge is gathering large data sets of sufficient quality to be able to train machine-learning tools.
 
‘If you care about a problem, collect data on it,’ he said. ‘The more primary care data we collect, the more opportunities there will be for AI to help.’
 
Dr Seneviratne said there are outstanding issues around actionability, correcting for biases in the data, questions around the black box issue of how these AI algorithms actually come to conclusions, and generalisability. He called for a framework governing the use of medical AI, similar to that used in pharmaceutical clinical trials.
 
‘We need controlled use in real work settings, using a similar culture as for pharmaceuticals or medical devices, with ethical reviews, regulatory oversight, and sandboxing – testing in a controlled environment,’ he said.
 
AI ultimately has to fit into the lives of people, Dr Seneviratne said, so it is invisible to users.
 
He finished with a call for help.
 
‘We need your help in four grand challenges – on collecting data, framing problems in a way that’s clinically meaningful, robust monitoring and patient and clinician involvement in the design process,’ he said.
 
‘You’re in a wonderful position to shape a fledgling field and make AI useful.
 
‘There will be mistakes and overhyped claims, but it will have definite use.’

Login below to join the conversation.

AI artificial intelligence GP19