So what about medicine and healthcare? The application of ‘big data’ using AI in health has been around for many decades. Machine learning however is a newer application of AI where computers are trained on data sets but critically can then learn how to perform tasks and make predictive decisions without explicit programming. And some of the really exciting advances in machine learning are examples of deep learning; here very large data sets are stratified into many different layers and can start to learn or in other words improve the probability of what is correct or incorrect in comparison to the training sets. A further difference is scale; healthcare is becoming increasingly data intensive: cheaper and more widespread genome screening, smart phone use, electronic health records – its potential is growing exponentially. The difficulty is trying to apply it appropriately to a healthcare system that tends to grow incrementally.
With big data comes big responsibilities. Consent and providing a genuine understanding of data use to patients is certainly one of the most important of these. We need to balance not stifling innovation with the need for an accepted evidence base of benefit and the very real safety risks of poorly designed interventions.
So what are some of the hot areas where AI is starting to impact meaningfully in medicine?
- Earlier diagnosis. Accurate diagnosis of and exclusion of melanoma is already an area where AI can improve diagnostic specificity and sensitivity compared to an international reference group of dermatologists. While it may not yet achieve the level of the ‘super-specialist’ quite yet, it relentlessly continues to improve.
- Smart algorithms often fronted by an avatar that provide auditory or text conversations are growing rapidly, particularly in primary care. Some also focus on providing information in particular disease areas such as cancer. Many still struggle to convey accurate personalised information particularly with rare diseases and co-morbidities but this is changing rapidly.
- Workforce challenges. The 2017/ 2018 RCP census highlighted the alarming finding that 45% of advertised consultant physician posts went unfilled due to a lack of suitable applicants. Much of the work we undertake is administrative or requesting and checking large amounts of routine diagnostic tests. Applying AI to some of these pathways is likely to be more efficient and cost saving by freeing up expert time for more challenging tasks that require a greater understanding of clinical context and relevance.
Despite clear opportunities there remain significant challenges which provide opportunities for forward-thinking clinicians.
- A knowledge/expertise gap between clinicians and companies that develop AI: AI companies don’t have the staff that are sufficiently clinically connected to inform their algorithms; people with clinical knowledge don’t know about algorithms. We need to find ways to span that gap. There needs to be a greater emphasis on how we engage and educate physicians (and indeed the whole NHS workforce) so they can contribute meaningfully to what is in essence a fourth industrial revolution. The NHS Digital Academy is a start but largely is recruiting the technological early adopters.
- The issue of who is diligently assessing companies developing AI. Some work is happening here in primary care. Could the RCP develop a similar framework in secondary care?
- The huge task of data cleansing. A very significant and costly task is not the AI development itself but the work to collate and cleanse the data, and the fact that the NHS at the moment is not set up to do this. Unless we have a framework for agreed data entry across health and social care systems we will continue to struggle to use these tools in the best ways that benefit our patients and populations.
A final thought. I am convinced that AI will become an indispensable tool across many aspects of healthcare. Will it put us out of a job? AI will not replace physicians. However physicians who use AI will replace those who don’t.