The promise of Artificial Intelligence in healthcare is continuing its rapid growth throughout the industry. Amongst all the hype, there remains a sense of hope that AI technology will make a significant impact in medicine. There are some striking examples of progress, for example, in radiology, in disease prevention, and in drug discovery. However, there are also some questionable claims. The lack of clarity around what AI is does not help with understanding about both the possibilities and the potential misgivings. This is true for the clinical community and for the public. How do we distinguish between advanced science and soundbites of the latest trend?
In order to start to answer this question, we need to define what AI is and isn’t. So, what’s the definition we should work with? Well, the original is still the best, when talking about healthcare. In 1955, John McCarthy organised a conference at Dartmouth College, New Hampshire. His aim was to develop ideas about thinking machines, for which McCarthy coined a new term, “Artificial Intelligence”. The choice of term was deliberate; it perfectly encapsulated the aim of the new field but was broad enough to cover the many different approaches they discussed. The conference is now widely acknowledged as the birthplace of AI. Amongst the attendees was Claude Shannon, the father of information theory. Without his work there would be no internet. In digital health, Shannon’s legacy extends across cryptography, data compression, biology, and genetics. The conference worked with the following definition: “For the present purpose, the artificial intelligence problem is taken to be that of making a machine behave in ways that would be called intelligent if a human were so behaving.”
Its brilliance is in its simplicity. It takes us away from thoughts about robotics and cybernetics – still many people’s first reaction to questions about AI – and moves us to thinking about tasks. If a machine can perform a task that would require intelligence if a person were doing it, then it’s AI. If that task is playing high-level chess, it’s AI. If the task is safely driving a car, it’s AI. If it’s suggesting differential diagnoses from a patient’s presenting symptoms, it’s AI.
Thinking this way leads us to a diverse and complex landscape for AI technology. There are different problems to consider – understanding language, image recognition, finding patterns in data, and improving logistics, for example. There are also different technical approaches – knowledge-based, probability-based, logic-based, and machine learning; and it is a landscape that is still evolving.
Unleashing the power of Deep Learning
Most of the major breakthroughs in AI over the last few years have been due to a single approach: Deep Learning. The concept of Deep Learning was first introduced in the 1980’s but the core technology that underpins it was discovered decades before that. It’s recently undergone a revolution. Deep Learning has enabled self-driving cars, speech recognition, and smart web services. It has impacted almost every industry, from finance to defence, and is the reason that Google DeepMind’s AlphaZero algorithm can now beat all human competition at the Chinese game of Go, something experts previously believed was over 50 years away. Deep Learning remains the fast-growing field in AI.
If the idea is decades old, then what sparked the Deep Learning revolution? In short, mathematics, computational power, and data. During the late 1980’s a major new technique was discovered that unleashed the power of Deep Learning, albeit restricted at the time by a lack of data and low computational power. Of course, no more needs to be said about the explosion of Big Data, but the final piece of the jigsaw comes, surprisingly, from video game consoles. The mathematics needed for great computer graphics is the same mathematics needed to power Deep Learning algorithms – it’s the calculations with matrices you might remember from high-school algebra. The development of graphical processing units, to power Fortnite and Mario Kart, inadvertently triggered an AI revolution.
In medicine, Deep Learning is certainly showing its value, swiftly infiltrating many areas within the industry. A major focus has been on medical imaging. Research teams across the world have developed algorithms for helping to detect diabetic retinopathy, pneumonia, breast cancer, and even to grade cardiovascular disease risk. These algorithms work solely on medical images and are backed up by impressive statistics. Radiology will be the first medical profession whose workflow is radically enhanced by AI.
It’s reasonable to ask why the focus has been on medical imaging. Firstly, it’s important; medical imaging is a key part of many diagnostic pathways, certainly for some serious, acute conditions. Secondly, intelligent image recognition spans many different industries, with many research teams working in this area. Thirdly, Deep Learning lives and dies by the data it can train on. Take a look at the United States; for every 100 Medicare recipients over 65 years of age, there are over 50 CT scans, 50 ultrasounds, 15 MRIs and 10 PET scans, per year. Every year. It’s an extraordinary amount of testing – most of which is almost certainly unnecessary – but it yields an extraordinary amount of data. If you’re working in a healthcare AI team using data from the United States, then medical imaging is almost certainly one of your major projects.
For most of the world though, carrying out imaging at this scale isn’t an option (nor is it really desirable). For predictive risk and early detection, it would amount to bringing in routine medical imaging, carrying a significant price tag. It would be a set of new public health screening campaigns, accompanied by complexities of patient safety and effectiveness.
For AI to have an achievable, sustainable, worldwide impact in healthcare, we need a different approach. We need to use the data held in electronic health records. In some countries there are already comprehensive, sophisticated records available, but this is seldom the case. The problem is often that the data is fragmented and can be of poor quality. To create the data sets for Deep Learning the data silos need to be broken; interoperability and FHIR-based standards are key. To get high-quality, coded, granular data, advanced terminologies like SNOMED CT are crucial. Improving the digital record infrastructure in hospitals, clinics, and across primary care, should be a core component of every country’s healthcare AI strategy.
The rise of smartphone ownership has also presented a new opportunity for detailed personal data collection. This includes data recorded in personal health records, as well as continuous data collection from wearable devices. Linking this with electronic health records yields a compelling data set. Symptoms, signs, tests, medication, diagnoses, procedures, immunisations, contacts, discharges, lifestyle factors, pollution levels, social stability, financial security, and social deprivation. It doesn’t take much to be convinced that this data is invaluable for both population health and personalised care pathways. It is also a data set that Deep Learning thrives on.
To progress with this approach, it’s vital to frame the exact problems we want to solve; the tasks we want to use Deep Learning for. This is a key part of clinical engagement. AI needs to bring genuine benefits to doctors and nurses, benefits that can be clearly realised in day to day practice. Only by working with clinical staff can we implement complex algorithms as part of simple, clean workflow.
As an example, take the problem of early detection of undiagnosed ovarian cancer in primary care. It’s a difficult task, but if a diagnosis is missed, or delayed, it can have serious consequences. Most family physicians will see five or six cases across their entire career. The condition presents with symptoms that have many other plausible diagnoses, many of which are far more common.
Advanced clinical decision support for early cancer detection is high on the priority list for many people, from prime ministers, to charities, to patients. Deep Learning algorithms, derived from the data in comprehensive electronic health records, can help. The algorithms have solid predictive capabilities, effectively crowd-sourcing the experience of hundreds of thousands of physicians. They behave intelligently on a complex task. However, working closely with family physicians on the implementation of these algorithms is essential. For example, early indications have shown that physicians favour a retrospective approach to these warnings – at the end of the day – rather than more in-consultation pop-ups and alert fatigue.
Deep Learning has also uncovered distinct clusters of diabetic patients from health record data, clusters of people at different stages of disease progression. Even better, our analysis has pointed to more personalised biochemical targets for these patients, to minimise the risk of future diabetic complications. These are just two examples; we can use this approach for many tasks in medicine, both operational and clinical.
There are of course many other aspects to consider. Ethics and governance play an absolutely central role, whatever the approach to AI. There must be the right consent models in place, strong anonymisation standards, and appropriate publicity. Establishing trust is key. We have also seen how important it is to remain cognisant of local epidemiology, local service provision, and potential bias against gender, ethnicity, and social factors. We cannot forget core scientific principles and must provide a clear evidence-base for our developments. It is also essential to track progress from the research lab into clinical practice, gathering evidence along the way.
Returning to our earlier definition, we need to concentrate on the tasks where AI can enhance healthcare by behaving intelligently. This is not about AI replacing physicians and won’t be for the foreseeable future; it’s about helping. There are pressures on all health systems, both financial and operational. There are significant workforce shortages and skills problems in many parts of the world. The position was beautifully summarised by Antonio Di leva, writing recently in the Lancet: “Machines will not replace physicians, but physicians using AI will soon replace those not using it.”
If implemented properly, AI technology can increase precious face-to-face contact time between physicians and patients. It is never influenced by availability bias, confirmation bias, or fatigue. The solutions we put in place need to be focused on specific tasks that will bring benefit to those on the healthcare frontline. This is a make-or-break year for AI in healthcare. It’s time to start delivering on the promise.
