Free Chapter: Introduction

Free Chapter 10 Jun 22

This book describes a holistic approach to maximising the potential of AI solutions in healthcare, drawing upon academic and other published work, and practical insights and lessons learnt from industry and clinical experts. It brings together information into one place in a factual manner with supporting interviews and real-life examples.

By the end of this book, you will understand more about three main areas:

  • The potential of AI to add value in healthcare and improve patient outcomes
  • Where AI implementation has worked and lessons learned from where it has not
  • A new approach to consider when designing, selecting and implementing AI solutions in healthcare to increase the likelihood of success and adoption at scale.

This book is aimed at:

  • Funders and policymakers including senior management and directors within government health departments, public sector funding organisations, venture capital firms, angel investors, and central public sector health organisations, such as NHS England and Improvement (NHSE&I), including the Transformation Directorate which will incorporate the former NHSX and NHS Digital.

You will be able to make better-informed decisions on policy design and investment in the ecosystems and wider enablers (for example, data infrastructure) that will lead to more successful implementation, adoption and scaling of AI solutions. Private investors will make more informed decisions on investments and how to support those selected innovators to be more successful.

  • Buyers and operators including staff in provider hospitals and other healthcare delivery organisations including operational and procurement managers, senior finance managers and clinicians.

You will make better-informed decisions about AI solutions on offer. You will be more confident in being involved in the co-design, on-the-ground implementation, monitoring and continuous improvement of these solutions.

  • Innovators and industry including founders and senior management in start-ups and other healthcare suppliers (technology-focused or otherwise).

You will make better strategic decisions about solutions (and related functionality) to develop and invest your resources in. You will design and execute implementation strategies that avoid common pitfalls and that will result in higher rates of success and wider adoption of the solutions.

The time is now

AI in healthcare deserves our close attention. Let’s look at the reasons why.

Healthcare systems are under enormous strain

The World Health Organisation (WHO) estimates that the percentage of the world’s population over 60 years of age is due to almost double, from 12% up to 22% between 2015 and 2050, and reach a total of 2.1 billion[1]. WHO also highlights the diversity of needs among the population and the increase in certain conditions such as dementia, osteoarthritis and diabetes, as well the chances of having more than one long-term condition at the same time (multimorbidity).

A US study has shown that patients with multimorbidities have been shown to have poorer quality of life, health and physical functions, with patients with three or more conditions shown to have a significantly worse outcome that those with one or two conditions[2]. A 2018 review of 300,000 people in England revealed that of patients admitted to hospital as an emergency, the percentage who have five or more conditions increased from one in ten in 2006/7 to one in three in 2015/6[3]. It also estimated that over the five-year period from 2018, patients with multimorbidities will increase total hospital activity by 14% and costs by £4 billion.

The backdrop is an increasing need to identify cost efficiencies in delivering healthcare, and often in the context of a reduction in available funding in real terms.

Covid-19 has added fuel to the fire

Covid-19 has had a significant impact across different countries on how health systems operate, elective surgical procedures, and workforces.

The waiting list for treatment on the English NHS reached 6.18 million in February 2022, a 46% increase since March 2020[4]. A review of waiting time data on elective surgeries in Finnish hospitals revealed that in 2020, compared to between 2017 and 2019, the waiting time increased on average by 8%. In certain specialities, the waiting times increased up to 34%.

A US study in 2020 modelled the projected elective orthopaedic procedures (defined as total knee and hip replacements, and spinal fusions), to estimate when the health system will be back at full capacity to perform these procedures and the size of the backlog that would have accumulated[5]. With the assumption that elective surgeries would resume in June 2020, the study estimated (in the optimistic scenario) that the health system would be at 90% of capacity after 7 months and still have more than a one million procedure backlog in two years’ time in May 2022. Another US study highlighted the wider adverse effects of deferring knee replacement surgeries[6]. Apart from the continuing discomfort of osteoarthritis and inconvenience related to rescheduling surgeries, it explains that patients will suffer muscle wasting, making rehabilitation more difficult and that it will worsen comorbidities such as depression and lead to a reduction in the overall quality of life.

The healthcare workforce was already stretched before Covid-19. WHO estimated that we will need 18 million more healthcare workers globally by 2030[7]. It is difficult to recruit healthcare workers. Despite increased demand and government plans to increase the number of general practitioners (GPs) in England, there are in fact 1,565 fewer fully qualified full-time equivalent (FTE) GPs in February 2022 than there were in 2015[8].

The American Hospital Association reports that America will have a shortage of up to 124,000 physicians by 2033[9]. There will also be shortages of other healthcare workers, especially in certain rural and urban locations. The article also highlights a survey done by the Washington Post-Kaiser Family Foundation in 2021 which showed that nearly 30% of the surveyed American healthcare workforce is considering leaving, and that nearly 60% reported an impact on their mental health from work related to the Covid-19 pandemic.


Mistakes and missed diagnoses occur during the delivery of healthcare services

A 2014 US study collated results from three studies and extrapolated the rates to the US adult population[10]. It showed a 5.08% diagnosis error rate in an outpatient setting and estimated that approximately 50% of these errors could lead to patient harm. A review of over 2,000 English primary care consultations revealed missed diagnostic opportunities in 4.3% of consultations[11]. 72% of these instances had two or more contributing factors from within the processes of taking a medical history, examining the patient, or ordering, interpreting or following up on investigations. It was estimated that 37% of these instances led to moderate to severe avoidable patient harm.

A review of emergency department (ED) patient safety reports in England and Wales from 2013 to 2015 to identify diagnostic errors revealed that 86% were delayed, and that 14% were incorrect diagnoses[12]. Bone fractures were the most common diagnoses involved (44%) with myocardial infarctions (heart attacks) the second-most common (7%).

It has also been estimated that 237 million medication errors are made every year in England, leading to 1,700 patient deaths and costing the National Health Service (NHS) £98 million[13]. The highest proportion (51%) occurred during administration of the drug, with 21.3% occurring during the prescribing stage.

AI can help

AI has huge potential to add value in healthcare across a number of areas. This is across the spectrum of health, from basic biomedical sciences, drug discovery and clinical trials to the provision of healthcare services through primary and secondary care, and preventative and self-management services.

It’s interesting to reflect on how AI could help healthcare professionals deliver care. The following Lynda Chin quote is included in Eric Topol’s book Deep Medicine[14].

“Imagine if a doctor can get all the information she needs about a patient in two minutes and then spend the next 13 minutes of a 15-minute office visit talking with the patient, instead of spending 13 minutes looking for information and two minutes talking with the patient.”

Studies and anecdotes from general practitioners bring to life the challenges primary care clinicians have in delivering care, often within a very short period of time, juggling to fit in several activities as well as actually talking to the patient. A systematic review of primary care physician consultation durations considering 28.5 million consultations revealed a wider range of consultation lengths, from 48 seconds in Bangladesh to 22.5 minutes in Sweden[15]. They are reported to last an average 9.2 minutes in the UK, which also needs to include arranging and/or reviewing investigations, making specialist referrals and administrative tasks to enable quality-related payments[16]. The average number of problems patients present with during these consultations also differs, with 2.5 being the average in England[17] and family physicians in the US reporting managing 3 problems on average[18].

As Atul Gawande points out, the expansion of medical knowledge means that ‘doctors can no longer know and do everything’ and they ‘must specialise in a field to absorb all the relevant information to treat a certain kind of illness’[19].

Gawande warned new doctors that[20]:

“the volume and complexity of the knowledge that we need to master has grown exponentially beyond our capacity as individuals”

Dr Gawande

If AI can help healthcare professionals gather and make sense of relevant information to help them have more efficient consultations and build more meaningful relationships with their patients, that, surely, can only be a good thing.

We need to separate the hype from the reality

AI is often seen as a clever quick fix for the thorny issues related to the delivery of safe, effective and efficient healthcare.

Audiences heard Vinod Khosla, a Sun Microsystems co-founder and Silicon Valley Investor, make a controversial speech in 2012, at the Health Innovation Summit hosted by Rock Health in San Francisco[21]:

“Machines will replace 80% of doctors in a healthcare future that will be driven by entrepreneurs, not medical professionals”

Predictably, Khosla’s comments sparked outrage from the medical profession. I will revisit later in the book the more nuanced comments Khosla has made elsewhere, and how AI could contribute to Lynda Chin’s visionary GP visit. For now, let’s consider the broad outcomes. The real-life results of AI in healthcare have not been impressive, and AI is nowhere near replacing 80% of doctors.

An article published by Massachusetts Institute of Technology’s Review in July 2021 highlights two review papers and a report by the Turing Institute that considered the impact of AI tools developed to predict and support management of the Covid-19 pandemic[22]. The conclusions were damning. The Turing report revealed the minimal impact of AI tools; the two review papers assessed 647 tools, and concluded that none were fit for clinical use, and only two warranted further evaluation of potential.

Translation from theoretical and testing results to success in front-line clinical settings can also be a challenge. The team at Google developed a Deep Learning (DL) algorithm that analysed photos of the retina (back of a patient’s eye) to identify signs of diabetic retinopathy in patients in Thailand, damage caused by high blood sugar levels[23]. A successful solution could help mitigate the shortage of specialist doctors in Thailand who can review these images. The algorithm showed impressive levels of accuracy, (displaying more than 90% sensitivity and specificity, that is, confidence identifying disease and confidence when no disease identified).

The team then performed an observational study of the tool being used in the clinics covering 7,600 patients. Fieldwork consisted of observation and interviews with nurses and camera technicians at a small selection of clinics before and following implementation of the solution. Low lighting caused issues with the quality of the images (21% of 1,838 were rejected) and could not be graded by the algorithm, which frustrated the nursing staff, who felt the image was of sufficient quality to be graded by a human specialist. Poor internet connectivity and speeds caused delays and reduced the number of patients that could be seen in a clinic (with a reduction from 200 to 100 patients screened due to a two-hour internet outage). The team is now working with the user to identify new workflows and to overcome the barriers identified.

It’s not easy…

Companies applying AI to health, medicines, and biotechnology raised $12 billion funding in 2020, double the $6 billion raised in 2019[24]. However, there are a very limited number of success stories at a national scale.

The design, development and implementation of AI in healthcare is demanding. The personal and high-stakes nature of healthcare means some of the challenges common to non-healthcare settings are intensified and new, knotty problems introduced.

Your way around this book

This introductory chapter explains why it’s the perfect time to fully consider AI in healthcare and provides an overview of the content I cover later in the book.

Chapter Two, Opportunities for AI in healthcare, identifies where AI can add value. We delve into more detail about AI and the elements relevant to healthcare. I also propose an approach to identify areas where I think AI can add value and illustrate these areas of opportunity with examples. The chapter considers the phases of healthcare delivery, and whether each is directly related to care delivery or to supporting back office functions.

Chapter Three, Success stories, explores successful AI implementations, both in the National Health Service (NHS) in England and in the US. I then consider large-scale AI applications within the retail and entertainment industries, and identify key success factors; here are potential insights to bring back to healthcare.

In Chapter Four, Healthcare is changing, I delve into how the Covid-19 pandemic accelerated the adoption of digital health. I explore the key challenges healthcare was facing before the pandemic before considering how Covid-19 has redesigned healthcare and with it, and the challenges we face in the future.

In Chapter Five, It’s not easy being a machine: Trust and accountability with new technology, I explore the background to and events shaping some of the specific challenges to AI in healthcare. Then I identify and discuss the challenges across three areas; people, systems, and technology.

In Chapter Six, Key themes with AI in healthcare, I group and further reflect on the areas that are less developed, impacting the successful adoption and scaling of AI in healthcare. I illustrate some of the key themes with an example and story from IBM Watson Health.

In Chapter Seven, Striving for good practice with AI, I identify and discuss the key factors needed for success, and I pull in the insights and suggestions of the experts and practitioners I interviewed.

In Chapter Eight, A roadmap for the practitioner, I propose a holistic 3-step approach to improve the likelihood of success that incorporates the challenges, key themes and the good practices I discussed previously. I stress the importance of solving the right problem, I lay out a roadmap considering the lifecycle of AI solution development and the key stakeholder involved, and I highlight the wider enablers that are needed.

Chapter Nine, Let’s do this, is a call to action. It is for everyone involved to work together on the approach I outline to maximise the potential of AI in healthcare.

I appreciate you may want to dip in and out of the book or want to go straight into the ‘how’ in Chapter Eight. Feel free to do so, because I have written each chapter so it can be read on its own.

What is artificial intelligence?

There are many definitions of AI and I have seen various people use the term in different contexts, and to mean slightly different things. Before we get stuck in, I want to spend a bit of time exploring the history of the term as well as selected subsets of AI, so it is clear when I use these terms later.

In his famous 1950 paper Computing Machinery and Intelligence[25], Alan Turing asked:

“Can machines think?”

In this paper, he described the ‘imitation game’ where an interrogator would try and distinguish between a human being and a computer based solely on the answers to a set of questions. This is now commonly known as the ‘Turing test’. He thought that within about 50 years, computers, much like a child, could be taught, programmed to play the imitation game so well that in at least 70% of tests, most interrogators would not be able to tell the difference between the computer and a human being after five minutes of questioning.

There have been many attempts to pass the test, with constrained scenarios such as pretending to be a 13-year-old Ukrainian boy called Eugene Goostman in 2014 [26] or Google’s virtual assistant “Duplex” calling a hair salon and successfully booking an appointment[27], but it is generally accepted that no software has passed the Turing test in a meaningful manner. Even recent models such as GPT-3 (Generative Pre-trained Transformer), which uses deep learning to generate text similar to a human, has failed the test[28].

However, the field of artificial intelligence (AI) has made significant progress.

It is believed that John McCarthy came up with the term ‘artificial intelligence’ to describe the subject of a conference he organised to discuss research topics including complexity theory, neuron nets, and learning machines, in Dartmouth in 1956[29].

McCarthy put together a collection of questions and answers related to AI in 2004[30]. In this, he defines AI as:

“It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.”

He also lays out the applications of AI, which include speech recognition, understanding language, computer vision to understand images and videos, classification (for example, assess customer risk when accepting credit card payments) and ‘expert systems’ to help users carry out tasks. (For example, the MYCIN system developed at Stanford University in the 1970s, identified bacterial infections and suggested treatments).

Machine Learning (ML) is a subset of AI which enables computers to learn through data and examples, rather than being specifically programmed with instructions. There are broadly two types of ML. Supervised learning uses a labelled dataset (where the ‘correct’ answer is given) to train the algorithms that are then used to make predictions. Unsupervised learning does not require labelled datasets; instead it identifies patterns itself from the dataset. Semi-supervised ML sits in between, using a small, labelled dataset to learn and then apply those insights into a larger unlabelled dataset.

Deep Learning (DL) is a subset of ML, which draws inspiration from the human brain and the layered neural networks within it, allowing it to work with more large, raw and unstructured datasets to understand complex relationships in data and present more intricate insight [31].

One of the key parts of the AI ‘engine’ is data. To make the engine perform efficiently and accurately, the right amounts (often large quantities) of relevant and representative data is required. Furthermore, it needs to be complete, accurate and in a ‘form’ suitable for the AI technique being used.

References

  • Williams, J.S. and Egede, L.E., 2016. The association between multimorbidity and quality of life, health status and functional disability. The American journal of the medical sciences, 352(1), pp.45-52.
  • Jain, A., Jain, P. and Aggarwal, S., 2020. SARS-CoV-2 impact on elective orthopaedic surgery: implications for post-pandemic recovery. The Journal of bone and joint surgery. American volume.
  • Cisternas, A.F., Ramachandran, R., Yaksh, T.L. and Nahama, A., 2020. Unintended consequences of COVID-19 safety measures on patients with chronic knee pain forced to defer joint replacement surgery. Pain Reports, 5(6).
  • Singh, H., Meyer, A.N. and Thomas, E.J., 2014. The frequency of diagnostic errors in outpatient care: estimations from three large observational studies involving US adult populations. BMJ quality & safety, 23(9), pp.727-731.
  • Cheraghi-Sohi, S., Holland, F., Singh, H., Danczak, A., Esmail, A., Morris, R.L., Small, N., Williams, R., de Wet, C., Campbell, S.M. and Reeves, D., 2021. Incidence, origins and avoidable harm of missed opportunities in diagnosis: longitudinal patient record review in 21 English general practices. BMJ Quality & Safety.
  • Hussain, F., Cooper, A., Carson-Stevens, A., Donaldson, L., Hibbert, P., Hughes, T. and Edwards, A., 2019. Diagnostic error in the emergency department: learning from national patient safety incident report analysis. BMC emergency medicine, 19(1), pp.1-9.
  • Elliott, R.A., Camacho, E., Jankovic, D., Sculpher, M.J. and Faria, R., 2021. Economic analysis of the prevalence and clinical and economic burden of medication error in England. BMJ Quality & Safety, 30(2), pp.96-105.
  • Topol, E., 2019. Deep medicine: how artificial intelligence can make healthcare human again. Hachette UK.
  • Irving, G., Neves, A.L., Dambha-Miller, H., Oishi, A., Tagashira, H., Verho, A. and Holden, J., 2017. International variations in primary care physician consultation time: a systematic review of 67 countries. BMJ open, 7(10), p.e017902.
  • Salisbury, H., 2019. Helen Salisbury: The 10 minute appointment. Bmj, 365
  • Salisbury, C., Procter, S., Stewart, K., Bowen, L., Purdy, S., Ridd, M., Valderas, J., Blakeman, T. and Reeves, D., 2013. The content of general practice consultations: cross-sectional study based on video recordings. British Journal of General Practice, 63(616), pp.e751-e759.
  • Beasley, J.W., Hankey, T.H., Erickson, R., Stange, K.C., Mundt, M., Elliott, M., Wiesen, P. and Bobula, J., 2004. How many problems do family physicians manage at each encounter? A WReN study. The Annals of Family Medicine, 2(5), pp.405-410.
  • Gawande,  A.  2010. The Velluvial Matrix. The New Yorker. (Jan. 2010).
  • Beede, E., Baylor, E., Hersch, F., Iurchenko, A., Wilcox, L., Ruamviboonsuk, P. and Vardoulakis, L.M., 2020, April. A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (pp. 1-12).
  •  A. M. Turing (1950) Computing Machinery and Intelligence. Mind 49: 433-460.
  • Floridi, L. and Chiriatti, M., 2020. GPT-3: Its nature, scope, limits, and consequences. Minds and Machines, 30(4), pp.681-694.
  • LeCun, Y., Bengio, Y. and Hinton, G., 2015. Deep learning. nature, 521(7553), pp.436-444.

Leave a Reply

© Janak Gunatilleke 2022