Deep learning and health The hurdles machine learning must leap to aid the NHS
Startups and Silicon Valley giants are pushing into medicine with artificial intelligence and deep learning. Nicole Kobie reveals how smart use of data could save your life
Data is important in healthcare. How a chart is read, if a doctor has time to take a second look at that scan of your chest, and whether there’s enough evidence to make you book an appointment with your GP could mean the difference between life and death for you — and of lower-cost preventative measures and expensive treatments for the NHS and insurance companies.
Currently, we rely on doctors and nurses to interpret key information — but machines are already coming to their aid, scanning images for signs of cancer, analysing data for symptoms of kidney failure, and more. In the future, apps will allow you to ask Alexa for medical advice, tools will assist GPs with triage and hunt for signs of cancer in medical scans, and chatbots could help treat mental illness.
The NHS is working with Googleowned DeepMind on projects to analyse patient data to predict kidney failure, spot head and neck cancers, and to read complicated eye images. However, those projects are under scrutiny because of data-sharing agreements with Google’s parent firm Alphabet, problems that could delay the life-saving — and budget-rescuing — potential of deep learning and other forms of machine learning.
Three months ago, we tapped the wisdom of ten health experts to see if tech could save the NHS ( see issue 270, p30). Here, we dig deep into the data.
Necessary data
The coding and engineering that goes into deep learning is only the start. No matter how a machine-learning system operates, it follows the old IT axiom: garbage in, garbage out. They need to be trained on high-quality, accurate data before they can read our medical charts and interpret images. “Without data there can be no application,” explained Natalia Simanovsky, business development lead at data management platform Cvedia, speaking at ReWork’s Deep Learning in Healthcare Summit in London. “You need lots of data.” That’s complicated by privacy concerns — and not helped by previous missteps by technology companies entering healthcare, which have had a chilling effect on the ease of getting data. Plus, when data is made available, it’s often in legacy
formats that must be converted, which adds yet another barrier.
Even when data is accessible and properly formatted, it needs to be of good quality. Simanovsky noted that training a skin-scanning system requires not only a large number of high-quality medical images, but correct labelling to avoid passing mistakes on to machine-learning apps.
“We need to work with dermatologists to ensure these images are correct,” she said. If an image is labelled incorrectly, the mistake could be learned by the algorithm — which means doctors’ errors are passed on to the machines.
Algorithmic assessment
Diagnosing disease will remain the job of human doctors for a long time to come, but AI could help narrow down potential causes of symptoms, acting as an assistant to human physicians, explained Daniel Nathrath, co-founder and CEO of Ada Health, speaking at ReWork. Doctors have the disadvantage of being human, meaning they have biases, make mistakes and suffer limited brain capacity. “It’s impossible to have it all in your head; there’s where AI has an advantage,” said Nathrath. “We’re not replacing doctors; we’re augmenting the intelligence of a doctor.”
Machine-learning analysis is particularly useful for people with multiple health issues – known as comorbidities – and that’s applicable to most older people, notes Marzieh Nabi, director of business strategy at Xerox PARC. Doctors aren’t diagnosing your complaints one by one: they need to consider multiple causes of symptoms and think about how medications will interact – something that isn’t often tested in drug trials. Deep-learning tools could be applied to pre-existing data sets to analyse how drugs interact and how diseases progress, Nabi noted, making it easier for doctors to decide how to address all that’s wrong with your health– without the treatments conflicting. Such diagnostic support could help health services with stretched budgets, assisting medical staff to work faster and better, but it could also help people in areas where doctors are scarce. “There are one billion people with no access to doctors,” Nathrath said. “Even in developed countries, people wait weeks.”
Treatment techniques
We often judge doctors on their bedside manner — on how well they listen and communicate in return, and how they make us feel. It’s one area machines struggle, since understanding natural language isn’t easy. “One thing human doctors are still better at is empathy,” said Valentin Tablan, principal scientist at Ieso Digital Health. “This will be something good doctors do more of in the future.”
But, once speech is unpicked, deep-learning tools could also be used to offer treatment to those suffering with mental health issues — and that’s as many as a quarter of us, with suicide the leading cause of death for younger men, noted Tablan. “Unlike cancer, which we don’t always know what to do about, we do know what to do with mental health,” he explained, as cognitive behavioural therapy (CBT) is proven to offer respite for many people.
However, it’s a human-intensive process that’s high cost, leaving those in rural or poor areas unable to access the therapy. Furthermore, the remaining stigma surrounding mental health issues means some people fail to seek treatment in person.
Tablan said deep-learning systems are already assisting with triage, by suggesting diagnoses and tracking indicators to prevent relapses. Following recent trials, his firm has posted a 44% accuracy rate in triage — which is on par with human doctors. The aim is to extend such technologies beyond diagnosis into digital therapists, with patients speaking to bots as part of their treatment.
Doctors’ jobs
Medical staff have difficult jobs — and that’s even before budgets start shrinking. “It’s not about replacing doctors; it’s about helping them do tough jobs,” said Tablan. Technology such as deep learning could help them work more quickly and accurately, but it remains early days. The companies and research mentioned here are discussing trials; they’re not finished, usable products. That’s partly because the data to train deep-learning systems isn’t easily available, with privacy a key concern, but also because the technology is still imperfect — for example, anyone with an Amazon Echo in their home will be aware of the limits of AI for nuanced conversation, which is at the heart of any diagnosis.
“It’s important not to oversell what AI is able to do,” explained Aureli Soria-Frisch, head of R&D for AI firm Starlab. “Any kind of mechanical, artificial system makes errors. 100% performance doesn’t exist in any system.”