What can you do with AI and big data today and what will be possible within three years? Why has adoption been slower than you might have expected? And what sort of role should health care providers look to play? HBI gets an expert view from digital, technology and business strategy expert Klaus Boehncke, partner at L.E.K. Consulting, who is chairing the Best Practice Workshop: Delivering Digital Excellence: A Framework for Success at HBI 2020 in two weeks.
HBI: Our impression is that as with self-driving cars, adoption in health care is somewhat slower than might have been anticipated. Why and what are the hot spots for you?
Klaus Boehncke (KB): One of the most important aspects to understand in the Digital World is the principle of exponential growth. Once anything moves to bits and bytes, it develops according to Moore’s law, which means the technology doubles in performance per dollar every 12-18 months. This means things move ‘gradually and then suddenly’. Of course, people understand exponential development better now due to coronavirus.
With AI, it’s very similar. Much of the movement is happening under the radar, it’s still in the ‘gradual’ phase. But if you look closely, you can see it everywhere. Let’s look at some leading indicators. The number of academic studies on deep learning and AI in medical specialities has increased from 600 p.a. in 2010 to over 12,000 p.a. in 2019. Investments into digital health venture funding have increased nearly ten times in the same time period, from just under €1bn to nearly €7bn, and there are well over 300 such deals every year. Also, note that the number of FDA-cleared AI algorithms has passed 70.
You can see AI being used in the entire patient pathway, from stratifying the population by risk to diagnosis and treatment. AI can also help make all this even more personal and precise for the individual at a genome or biome level. Of course, you then have other applications in areas such as workflow management and the development of new drugs and medtech.
Some specific cases have emerged in the last 12 months, from primary care through acute care and even including the fertility space:
Take the market of apps in primary care that help patients understand and manage their disease symptoms, like Ada Health. Ada has become a leading AI-based primary care platform since the global launch of its personal health assistant in 2016. Ada currently has nearly 11m users and has completed over 20m assessments globally. In our own clinical evaluation, Ada was found to materially enhance the diagnostic capability of GPs, not only when assessing common but especially also uncommon conditions.
A further example is from acute care, stroke: There is an AI-based stroke care is a solution called RapidAI. This software is already used in over 1600 hospitals in more than 50 countries. Based on clinical trials that used RapidAI imaging software to select stroke patients for late-window treatment, the American Heart and American Stroke associations revised guidelines in 2018, extending the treatment window for acute stroke patients from 6 to 24 hours. This obviously has a significant impact on patient outcomes and provider economics.
A third example is leading fertility provider Virtus Health, which has teamed up with a company called Harrison.ai, who have developed a deep learning algorithm to support an embryologist’s decision on which embryo is most likely to result in a pregnancy. Currently undergoing a randomised clinical trial across Asia and Europe, the technology is demonstrating that it can discriminate between embryos that will result in a pregnancy with a predictive power of 0.95, which is superior to existing methods.
So we can see some really impactful examples of AI if we look in the right places.
HBI: So why does overall adoption seem slow?
KB: I think the above examples show that the adoption is actually quite high for some application areas – but that this is not widely known outside of these specialities, or even within the broader healthcare professional space.
As you’ve written about at HBI, we have very innovative digital legislation now in Germany, where apps can be prescribed and will be reimbursed by public health insurance. The issue, of course, is that family doctors know pills and they don’t know these digital apps. Put yourself in the position of a doctor who has prescribed pills for 30 years! “If I prescribe digital therapeutics will it lead to better results? How much time do I need to devote to researching this? My patients are elderly, how likely are they to use these things? Will I need to answer their questions? How much time will that take?” In a recent survey by the Barmer Statutory Health Insurance, more than 55% of GPs stated that they were “ill-prepared” for digital apps. And of course, most of the start-ups behind these digital tools are small, without the sales and marketing skills enjoyed by big pharma (who regularly go out and engage and educate healthcare providers about their solutions).
We currently have an online automated assessment which asks users about their digital agenda and capabilities and assesses how they measure up against industry standards.
HBI: Yes they say that the latest medical advances take 20 years to reach the front line, don’t they?
KB: As I mentioned above, we are on this exponential technology path. The best illustration is probably the global development of the Covid vaccine, a process that used to take many years. As a result of data, software and information sharing, this has now been compressed at an incredible rate. Since these solutions can be very powerful, doctors may need to continually redefine what they do and how they treat patients in the future. But the profession is rightly conservative and careful about what they do with the patients.
HBI: I guess we will see patients Googling digital stuff online, taking it to their doctors and getting them to prescribe.
KB: Yes, a recent survey in Germany by bitkom suggested that 40% of patients will actively suggest digital apps to their doctors. Incidentally, more than 40% of patients also said they would actively seek out a second opinion by AI if it were available.
HBI: So what is the role for the health care operator? Incidentally, I have lost count of the number of operators who tell me that they are sitting on a trove of data which they are about to decipher with AI and that will massively boost their enterprise value!
KB: That is my experience too. IBM’s former CEO, Ginni Rometty, called data the “new gold”, but I slightly disagree. Gold has intrinsic value, while data doesn’t. It needs to be applied with intelligence; you need to know what data elements are of interest and which are not. If you can’t use the data to, for example, diagnose or improve treatment or increase efficiency or achieve other significant benefits in the healthcare system, then it may actually be worthless.
Companies with interesting data have two main business models: you can either sell it or, if you know precisely what the data can do and you have IP, you can build on it. This typically means making a service available at a low cost or for free in order to grow the data set, and then using that to make your AI and data analytics better – like Google has done with maps and search. Primary care AI providers that we mentioned above are taking a similar approach, making their services available for patients for free and growing their data set and assessment capabilities.
The availability of large data sets to train AI is, of course, also an issue. Companies in China will have an advantage, due to the larger number of users. For example, Ping An Good Doctor carried out half a million teleconsultations per day even pre-Covid. In Germany, to counteract the large volumes of healthcare data available in China and the U.S., health minister Spahn earlier this year paved the way to make anonymised data from the Statutory Health Insurance companies (c. 73m people) available for research.
Ultimately, of course, data is a necessary element to train AI – but the other key questions are knowing what solution you are training for, and the business model you are going to build.
HBI: That sounds intrinsically difficult for health care operators.
KB: I agree. On the one hand, health care providers can simply choose to become a customer for AI algorithms that you use in the knowledge that in doing so you are adding to the data pool and expertise owned by someone else. I think that works OK for a hospital chain that is, for example, improving its stroke care capabilities by leveraging RapidAI.
But it is less good for a research-centred university hospital or specialist chain which would effectively enable third parties to benefit from the data from their patient cohort. Over time, their research and IP may become less relevant. To mitigate that risk, they can try building “digital twins” of themselves and offer their services globally to collect larger data pools for AI development themselves. This is what Memorial Sloan Kettering did when it partnered with IBM Watson Health for cancer care, for example. Diaverum, a leading renal care provider, has also announced such a strategy.
So we can see some large companies and healthcare providers looking to develop their own leading AI solutions.
HBI: Is that true of all areas?
KB: There are counter-examples as well: IBM Watson had high expectations when it bought Merge, for c.$1bn in 2015. The acquisition was really focused on the reported 30bn images from Merge which were to be used to train Watson’s AI, rather than the software capabilities. But imaging has remained a very fragmented area with lots of startups and with the highest numbers of FDA approvals. Increasingly we are seeing larger players in imaging not focused on creating their own AI solutions but building ‘app stores’ and platforms for third-party AI startups. For example, Wellbeing Software, a large radiology information system (RIS) player in the UK have their AI Connect platform, GE has launched Edison, and Zeiss has their Apeer offering.
I believe that areas in which algorithms can be trained on specific use cases quickly, and at low cost, will see faster development, more startups, and potentially slower consolidation. Imaging is such an area, because images can be piped through automated AI/deep learning platforms and the results do not need to be explained in detail, because oftentimes they will be visible to the radiologist – the software has just been used to identify the anomaly more quickly.
But in other areas that are broader and more complex, require supervised training by healthcare professionals, use lots of different data types, and need much more investment, you are likely to see fewer alternatives, faster consolidation, and less competition. Stroke care is one such example with only a handful of meaningful competitors worldwide; cancer care is also one of the most complex areas. In spite of criticism that it ‘overpromised and underdelivered’ (which is probably unsurprising given the multi-year development and ambition), IBM Watson for Oncology trained by Memorial Sloan Kettering Cancer Center, is in use by hundreds of hospitals globally. I’m not aware of any other successful cancer care platform of that magnitude.
If you would like to hear more from Boehncke, click here to learn about his Best Practice Workshop, and click here if you would like to book to attend the event.
If you would like to read L.E.K. Consulting’s Digital Excellence Framework Report, click here.We would welcome your thoughts on this story. Email your views to David Farbrother or call 0207 183 3779.