We live in an age of wide-reaching population health problems, such as diabetes, cardiovascular disease, and obesity, that have a huge physical, mental, and financial cost for patients and providers. However, new forms of artificial intelligence (AI) could be the answer to solving these population health problems in an efficient and cost-effective way.
According to the Centers for Disease Control (CDC), 60% of adults in the U.S. have a chronic disease and 40% have at least two. While there are many such conditions, heart disease, diabetes and cancer lead the field affecting the most people and accumulating the highest healthcare costs.
Overall, annual healthcare costs for U.S. citizens with chronic disease or mental health conditions account for 90% of the healthcare spend, which before the pandemic in 2019 was $3.8 trillion. Each year, 868,000 Americans die from heart disease or stroke alone and the numbers are increasing for most chronic conditions.
AI technology has improved exponentially in recent years. With an initial focus on imaging technologies for use in pathologist-led diagnostics in fields such as cancer, more recently AI is now being used in a variety of different medical fields including drug development and to develop models to improve population health.
IBM Watson’s disappointing entry into the field of AI and clinical decision support tools ten years ago was not a great start for AI and population health, but things have improved since then with a number of different companies and researchers working to target different aspects of population health.
One such company is IQVIA, one of the world’s largest healthcare data science companies. “Properly designed AI has the potential to make our health care system more efficient and less expensive, ease the paperwork burden that has more and more doctors considering new careers, fill the gaping holes in access to quality care in the world’s poorest places, and, among many other things, serve as an unblinking watchdog on the lookout for the medical errors that kill an estimated 200,000 people and cost $1.9 billion annually in the U.S. alone,” Joanne Hackett, Head of Genomic and Precision Medicine at the company, told Inside Precision Medicine.
Finding patterns in language
There are many types of data that need to be analyzed to put into practice effective population health management strategies. An approach taken by a number of different companies and researchers using AI to tackle medical issues is to use natural language processing (NLP) to search medical documents for information.
A newly formed startup, Realyze Intelligence, a spinout from the University of Pittsburgh health system, UPMC, is taking this approach but is combining difficult to analyze ‘unstructured’ language data with more standardized information found in electronic medical health records (EHR).The company has already applied this technique to help chronic kidney disease patients. “All of these detailed clinical findings, even things like social determinants of health, things like they live alone, or they have food insecurities, all of these things really matter when you’re trying to figure out how likely this patient is to progress on to the next stage,” commented Aaron Brauser, CEO and co-founder of the company, in an interview.
Brauser and colleagues were able to use this information to give clinicians the information they needed to focus in on the patients who were most in need of help. “Instead of 150,000 chronic kidney disease patients, there were 5000 that were highly likely to progress the dialysis in the next three to six months based on all of these clinical findings.”
During the COVID-19 pandemic the team also helped health authorities look for people at high risk of serious infection. “Things like, they live alone’ or ‘ they are oxygen dependent, again, things that are very clearly documented in the chart, but don’t tend to be represented in any sort of structured field.”
Brauser believes a key factor slowing uptake of tools, such as the one developed by his company, is poor data standardization across EHRs. This is not helped by the companies managing them. “They’ve created almost walled gardens where each hospital can define what certain things are, what codes they use, how they think. So, interoperability has been extremely poor.”
Predictive modeling for a healthier population
Assessing language in EHRs is just one type of data that AI algorithms can target. “Over the last 10 years the volume of healthcare data generated has gone exponential,” says Hackett. “There used to be one image of a patient per day: their morning X-ray. Now, if you get an MRI, it generates literally hundreds of images, using different kinds of filters, different techniques, all of which convey slightly different variations of information. It’s just impossible to even look at all of the images.” This is just on the imaging side. Data analytics companies Medial EarlySign and Glooko are both collecting and analyzing a range of other health data such as EHR information and medical test results using AI models.
EarlySign’s acting CEO and co-founder Ori Geva says that the company’s mission is “to keep patients healthier for longer.” Split between Israel and the U.S., the company was founded 8 years ago and has partnered with Geisinger Health System, a regional health care provider to much of Pennsylvania, for the last couple of years. As part of the collaboration, a number of machine-learning based tools that can look for early signs of disease are being implemented.
Notably, they are finding that telling people they are at higher risk for infections such as flu or diseases such as colorectal cancer, based on results from an AI-based algorithm, seems to significantly increase uptake of preventive measures like vaccination or screening. They are also working on preventive approaches in several other areas.
“We’re able to identify people that might be at risk for a condition, for example, people with diabetes already might have some sort of renal impairment, but are having a higher propensity for faster deterioration. These are people that eventually might end up with end stage renal disease, but there is stuff that could be done on the way,” notes Geva.
Glooko is based in California and was launched in 2011. It began as AI-led software and a mobile app for patients with diabetes and their healthcare providers to help optimize outcomes for diabetes patients.
While diabetes is still a strong focus for the company, it is now expanding into different areas, according to current CMO, Mark Clements, a practicing endocrinologist at University of Missouri-Kansas City. “Where I think the future leads us and where we are investing heavily right now, is in what I would call the domain of precision engagement. How does one deliver just the right digital therapeutic, or behavioral intervention, to just the right patient at just the right time?” This might include identifying who would benefit most from better eating choices, more physical activity or a different drug regimen. The company is also exploring different areas of remote patient monitoring for those with diabetes, but also related conditions such as hypertension and kidney disease.
“I think what we’re recognizing is that whether you’re talking about diabetes, or other chronic diseases, that you have to get the care out of the clinic, in order to move the needle,” Clements told Inside Precision Medicine.
“What those of us who provide care for individuals with chronic disease have learned over time is that the physiology of the disease and the treatment with drugs and devices is only 10% of the equation. The other 90% is human behavior, and how one interacts with these therapies at home.”
Tackling complex genomic and neurological pathways
Many models either evaluate real world ‘phenotypic’ data or genomics data, but not both. Precision Life, based in Oxford in the UK, is one of the exceptions.
“There is a huge amount of data out there, we can learn much more about patients than we ever could even when we imagined the Human Genome Project back 20 years ago,” Steve Gardner, CEO of Precision Life, explained in an interview.
“We now have high quality biomedical imaging, we now have transcriptomic data, we have liquid biopsies, we have environmental monitoring, we have epidemiological data being collected in EHRs, all of this information is available. The challenge is how best to use it to benefit patients and lower the cost of health care. And that is AI, it’s analytics.”
The company has developed what it calls a ‘combinatorial platform’ that links genetic variants, genes, and other factors driving disease together to provide disease-specific information. This information can be used for applications such as improving clinical trial design, but also to uncover drivers of chronic diseases.
Precision Life is using this information for many purposes, including to find patients at higher risk for type 2 diabetes complications and to look for individuals at increased risk of experiencing coronary artery disease. “We have risk scores which are better than the existing polygenic risk scores at predicting an individual’s risk,” says Gardner. “They present with chest pain or something like that, and rather than sending them immediately to very expensive imaging, you need to know who is actually at risk of having coronary artery disease. And so, those models are a very effective first predictor.”
“We’ve been developing ways to automatically characterize the signals in your nerves, turn them into digital biomarkers –very precise measurements of the disease, and we’ve been showing how you can optimize everything from drug discovery to creating a sort of electrical stimulation that can then treat diseases,” Emil Hewage, BIOS CEO and co-founder, told Inside Precision Medicine.
This work is still at its early stages, but the company is already working on a neural interface implant designed to transmit specific electrical signal to the nerves as a treatment for different conditions.
Overcoming challenges
Although there are a lot of companies and researchers exploring ways to enhance population health by using AI, there are also a number of challenges that need to be overcome before this kind of data analysis more mainstream.
To begin with there is the sheer complexity of analyzing multiple types and sources of data using AI. This requires a different approach on the AI front to other medical applications. As outlined by Brauser the data also needs to be high quality and have a good degree of standardization to provide high quality insights.
The data demographics also need to match the population that will be analyzed by the tool. “Sometimes you get a data source that you think is going to be fantastic,” says Geva. “Then when you get the data, you suddenly understand, these guys are collecting data that comes mostly from academic medical centers. When you look at who goes to these academic medical centers, it’s usually people that are much sicker. If you want to do something that reflects the general population, maybe that’s not the right data for you.”
A key issue is also adoption. These tools are only useful if healthcare providers and, in some cases, patients, are willing to use them. Clements believes better communication with healthcare providers and provision of educational resources on this topic can help with adoption challenges.
An approach being adopted by many developing technologies in this area to help with adoption is to move away from the ‘black box’ model of AI and make it explainable, sometimes at the cost of a small degree of accuracy, as it makes it more attractive to users if they can get a sense of why the algorithm has made the choice it has.
From an ethics standpoint, Clements advocates applying medical principles to use of AI systems. “I follow some first principles, which is that we should first do no harm when we’re implementing AI,” he explains, adding that the FDA and some institutions like Intermountain Health Institute have recently created an ethical framework to help implement AI tools in healthcare.
A major problem with these kinds of interventions or tools is that they are often not incentivized by the healthcare insurance companies. “Hospitals today are, are not incented, to prevent you from coming to hospital, they make their budgets by having so many people in the hospital,” says Brauser.
“For something like chronic kidney disease, you want to prevent them from showing up with crash dialysis in the ER, that one, costs a lot of money, two, once they get to that point it’s too late, you’re putting them on dialysis, and it’s a pretty quick downward spiral from there.”
A less obvious, but equally important challenge in this area is that of merging the two different fields of advanced computer algorithms and AI with medicine. For many years now, the majority of developers of software and applications have used Agile systems that advocate for continual innovation, early launch of products and updating or fixing of errors in a parallel process. “It’s completely contrary to how you think when you develop medical devices, you need to make sure that you’re getting out a device that works,” emphasizes Geva.
Equally, the medical regulation authorities are not currently well set up to deal with ever improving models, something built in to most machine-learning algorithms. “Do you have to go again, and re- approve it in terms of the regulatory process? What’s the regulatory approach towards ever improving models? says Geva.
Future directions
Despite the existence of many adoption challenges, AI-based tools and models to help improve population health and better diagnose and treat chronic diseases are already being implemented in many areas.
The COVID-19 pandemic has helped speed up development of digital healthcare initiatives and has highlighted that much of the regular care of chronic disease patients could be done outside of the hospital or clinic settings.
“More of the care is being done outside of the hospital… Devices to interact with people, monitor, and then alert people as things are changing, or need to be followed up with, I think those are pretty exciting,” says Brauser.
“Whether you’re leaving the hospital and going home, or even just going from one place to another, those things are where things break down. And I think AI can do a really good job of smoothing that and making sure people are getting the care that they need.”
Experts agree that wider implementation of AI-based population health tools will take time, but that things are moving in the right direction. “I think that there’s great benefit for society to apply AI. You need to pick the right battles; you can’t do everything at once. There’s a learning curve on how to deploy this,” says Geva.
Ever reducing costs for generating genetic or genomic profiles and recent legal mandates in the U.S. and elsewhere to try and standardize EHR terminology, language and systems should all aid the implementation of increasingly complex health risk prediction models, as well as those that can help the management of chronic conditions such as diabetes and heart disease.
Gardner says the ultimate goal is to try to bring precision medicine into clinical practice. “That is really what we’re trying to do. A lot of it is cultural, and a lot of it is economic, the cultural pieces are, do the clinicians believe what you’re telling them in terms of the performance of the tools and their relevance to their patients… And are you able to demonstrate the economic value across the whole health system?”
While the cost savings may not yet be evident on the large scale, over time these tools really have the potential to save large amounts of money currently spent on managing patients whose disease is not adequately controlled or treated. “There’s good economic reasoning to do that. So, we’ll see this coming I think,” says Geva.
Helen Albert is Senior Editor at Inside Precision Medicine and a freelance science journalist. Prior to going freelance, she was Editor-in-Chief at Labiotech, an English-language, digital publication based in Berlin focusing on the European biotech industry. Before moving to Germany, she worked at a range of different science and health-focused publications in London. She was Editor of The Biochemist magazine and blog, but also worked as a Senior Reporter at Springer Nature’s medwireNews for a number of years, as well as freelancing for various international publications. She has written for New Scientist, Chemistry World, Biodesigned, The BMJ, Forbes, Science Business, Cosmos magazine, and GEN. Helen has academic degrees in genetics and anthropology, and also spent some time early in her career working at the Sanger Institute in Cambridge before deciding to move into journalism.
