AI AND HEALTHCARE

The following is a summary of a Frimley Park Hospital Grand Round I gave on Artificial Intelligence (AI) and Healthcare on 21 June 2019

AI and Healthcare

There is growing interest in AI and its role in UK healthcare. The Topol review (2019) was commissioned by the UK government to look at how the NHS can benefit from technology and describes the UK as well positioned to be a world leader in healthcare technology. A section on AI and robotics highlights a national shortage of AI specialists, and recommends a programme of industry exchange networks. NHSX has been created as a specialised unit within the NHS to drive digital transformation and lead policy, implementation and change of healthcare technology from July 2019. Simon Stevens has announced that technology experts will advise on the adjustment of reimbursement frameworks by April 2020 and states that there is clear evidence that clinicians could be substituted by AI for clinical tasks. The government has announced a £250 million investment for AI development within the NHS.

What is AI?

AI is the attempt to understand intelligence, then build machines that recreate human intelligence, by which we mean a number of brain functions: thinking, learning, creating, and problem solving. AI is already integral to many routine computer functions from internet search engines to email spam filters. However, five trends have fuelled the recent surge of interest in AI: the development of powerful computers, an explosion of data (information overload), cloud computing (data infrastructure services available via the internet to smaller technology companies), developments in neuroscience, and the development of more sophisticated machine learning algorithms.

Very early visions of AI were of a system that would answer questions or engage in dialogue. In the way that a calculator obeys logical rules to perform sums faster than a human counterpart, traditional programming issues rules to a computer (if this, then that) allowing it to cope with many potential outcomes.

A series or pipeline of algorithms can lead to more complex systems, for example the combination of voice recognition and search engine functions allows Siri or Alexa to perform simple tasks to verbal command. Pattern or image recognition algorithms try to interpret an ECG in the most basic ECG machines (although my current lectures on ECGs begin by asking users to ignore these interpretations).

The more modern view of AI is that of a self-learning autonomous system or agent, capable of improving itself through experience without being explicitly programmed on how to do so. In machine learning, machines learn from raw inputs or training data and are not pre-programmed. For example, a feedback loop between the AI and its environment (or training data) is the basis of reinforcement learning based on trial and error, reward and punishment.

Thus software is ‘learnt’ from data instead of being written by programmers and is more able to cope with uncertainty or variability. These AI systems can devise novel insights or solutions to problems we may not even realise exist.

An example of this was the development of an AI by DeepMind in 2014, teaching itself to play Atari like computer games. In the game called Breakout where the AI teaches itself how to break through a wall by hitting a ball, the AI was missing the ball 3 out of 4 times after 30 mins (200 games). By 60 mins, however the AI was never missing a ball and by the 600th game, the AI had devised a novel way of maximising its score, tunnelling through the side of the wall and getting the ball to bounce behind wall. Another example of machine learning from DeepMind is Alpha Go, an AI capable of playing Go, a board game more complex than chess. AlphaGo was programmed with millions of moves of past Go masters but also able to improve with experience. The next version Alpha Go Zero relied on no human instruction. Starting with random play, the AI taught itself by playing itself millions of times, and within 3 days was able to beat Alpha Go. Within 40 days it was able to work out principles of Go that had taken humans 1000 years to work out and eventually defeated the world champion Lee Sedol in a stunning contest depicted in the Netflix movie, AlphaGo (2016).

Deep learning refers to a type of machine learning which mimics brain function, exploits many layered artificial neural networks, blossomed 5 years ago, removes the need for manual sorting of data and identifies patterns in disorganised datasets. Error rates in speech recognition, for example, were unchanged for a decade but dropped by 30% after the development of deep learning systems.

The holy grail is artificial general intelligence where system can operate over a wide range of tasks (compared to narrow AI which focuses on specific tasks).

AI and Healthcare

Data is the fuel that drives all AI, essential to build a training model. Once validated, the AI can then be deployed but continues to improve through experience. In the model above, the AI interprets patterns such as symptom clusters to make a prediction or diagnosis.

Healthcare data on diverse populations are fundamental to this process, making the UK and in particular the NHS an ideal training ground for future healthcare AI systems. Huge NHS datasets and other national databases like the UK Biobank are the fuel that will drive the training of healthcare AI. The US equivalents of such datasets are smaller and less diverse, for example a network of VA hospitals run by the US Department of Veteran Affairs delivers a dataset of 700,000 adults.

The role of AI in healthcare will range from the use of voice and pattern recognition systems to populate electronic health records, to pattern recognition systems that triage large quantities of data and identify patients in clinical need, to autonomous clinical decision making systems that complement and perhaps even replace healthcare professionals (as suggested by Simon Stevens).

AI will give rise to novel approaches to clinical problem solving, may free clinicians up to do clinically useful tasks (as emphasised by the Topol review), lead to keyboard free environments, and improve accuracy of diagnosis and timing of treatment. Mundane administration tasks will be delegated to AI, thus potentially reducing administration costs.

There are caveats. Machine learning is only as good as the data that goes in (Garbage in, garbage out). Real world data is messy and training data has to be cleaned up. We need to be wary of biases in AI systems that stem from biased training data, such as racial biases in facial recognition systems. Healthcare data will largely flow from electronic patient records, which has been a key driver for a paperless NHS (rather than any concerns about climate change).

There are concerns about data privacy and transparency of data sharing. We need to be wary of automation bias, the tendency for humans to favour automated decision making systems, a growing problem in intensive care units and aircraft cockpits.

Without doubt, AI will help us to answer big healthcare questions. However the roll-out of all technologies including AI in the NHS needs to be done safely with rigorous evaluation, putting patient safety at the centre of all transformation. Prospective clinical trials which rigorously test these technologies are important before AI is deployed on the frontline. This will require active participation by practising clinicians.

Vinod Achan is a consultant cardiologist and consultant lead for the heart attack service at Frimley Park Hospital (now part of the Frimley Health NHS Foundation Trust)

Twitter: @surreyHEART