Laying the Foundation for Artificial Intelligence in Healthcare
Medicine is often described as both an art and a science. One might think that the “science of medicine” should be fairly straightforward, driven by findings from clinical studies and evidence-based protocols derived from such findings. Yet numerous studies show that a startlingly high percentage of medical treatment is not in conformance with evidence-based guidelines. Some of this has to do with information overload – given that 90% of the world’s data has been generated over the last two years, most clinicians on the front lines of healthcare simply aren’t aware of some key evidence-based recommendations as they may pertain to the situation at hand. Even before the recent proliferation of data, a number of studies have found that it takes 17 years for advances from medical research to become incorporated into our standards of care.
Assuming that there is a knowable answer to each and every medical question, we must also consider the “art of medicine.” This might be seen as consisting of two main domains: (1) clinical judgment, and (2) communication skills. This post will focus on the concept of “clinical judgment.”
Frequently, the reason given as to why a particular clinical protocol was not followed in a given situation is that the guidelines were not felt to apply – health care, after all, needs to be individualized, and it is the clinician who knows the patient best. So, “clinical judgment” is substituted for broader guidelines. And, in fact, this may be precisely the best thing for patients. The current level of clinical evidence is drawn from studies on populations, which may yield general insights, but protocols based on averages may not be appropriate for each and every patient.
So, what is “clinical judgment”? It is a mixture of didactic learning and cumulative experience. And the crux of applying experience has to do with pattern recognition. In my own experience, for example, I have been in clinical practice for several decades, and have had well over 100,000 clinical encounters with my patients during that time – something fairly common among physicians who have been in practice for a while. When I come across a clinical situation that doesn’t fall into a specific “little box,” my instincts try to identify a pattern – “this situation reminds me of situation x in certain ways, so doing what worked in x might be useful here.” Pattern recognition may not be conscious, or might not even be something that can be articulated (the “gut hunch” that comes with experience). And such insight works better when drawn from a larger body of experience.
Can machines do this better?
Machine learning and pattern recognition are an emerging science. A simple example is a smartphone application using geo-location pattern recognition to generate alerts that are intended to be helpful. If I commute to work regularly on certain days of the week at a certain time, I will get a prompt that says “33 minutes to work” just before I should leave, on days that I work (with a map that I can click to see traffic options). This is an example of simple pattern recognition.
Medicine, of course, is orders of magnitude more complex. But with sufficient data and sufficient clinical knowledge, machine-learned pattern recognition can provide recommendations that are ultimately more reliable than – thought built on the same premises as – the “gut hunch” of the experienced clinician.
The promise of the evolution of machine-learned pattern recognition in medicine (Artificial Intelligence [AI] in medicine) is the ability to individualize a recommendation, based on everything known about an individual patient and comparing it to patterns drawn from the world’s experience. The findings of all clinical studies, the specific observations of the individual (from individual history, to appropriate surveys, to physical examination findings, to device-based data), the knowledge of interactions of medications and the impact of lab findings (including, increasingly, genetic/DNA data), and the outcomes of different treatment approaches in other people which match the situation-at-hand with a great deal of detail – all of this can be leveraged to make recommendations (Cards) that can be useful right here-and-now. It is the science behind the clinical “gut hunch.”
Obstacles to this evolution
One of the biggest stumbling blocks in health care is the fragmentation of data into institution-centered silos. There is a lack of universal patient-centered data. The current “state of the art” is represented by attempts to share data between silos in standardized ways, though such efforts leave the silos intact. The closest we get to multi-provider population-wide data actually comes from payers, rather than providers, though it is limited to the data that appears on bills submitted to the insurers.
One attempt at understanding trends in health care has come from payer-based cost data. Business intelligence and predictive analytics have been used to identify individuals who are at risk for becoming catastrophically ill. After all, from a cost standpoint, (from an AHRQ report) 1% of the population account for 23% of healthcare spending, and 5% of the population account for 50% of spending. The problem is, the 5% of the population who are catastrophically ill in 2015 are not the same individuals who will be the 5% in 2016. The challenge is to predict who will become catastrophically ill – those with chronic conditions are more likely, but sometimes relatively “healthy” people become catastrophically ill too. AI in medicine has the potential to yield more useful predictions, which will enable both tailored treatment and tailored prevention.
In order for AI in medicine to be really useful, it needs to combine payer data (longitudinal but limited) with clinical data (more complete individual data, but fragmented into silos), as well as with consumer/patient data from surveys, devices, and self-reported information. A truly universal patient data store is needed for AI to build towards its promise. AI is not about population health as much as it is about personalizing care for individuals. Without full access to what is known about the individual, AI can not fulfill its promise.
This is what Flow Health is building: a truly universal patient-centered data store, drawing from institutional EHR data, payer data, and consumer data. This can be linked to research and protocol data, and can build the AI platform needed, so that real, useful insights – that is, specific, individualized insights – can be surfaced in real time for patients and clinicians when they are needed to support treatment decisions. On the one hand, the insights can support population needs, and identification of individuals at risk of becoming catastrophically ill, and can dovetail with traditional research (where researchers test hypotheses with data from population-based studies) – the “science of medicine.” This technology can bring the medical library to the clinic without waiting 17 years, and tailor insights from medical research to the patient in real time.
But AI can do even more than that. Rather than start with a hypothesis and test for outcomes, it starts with an observation of the results (the detailed status quo) and tries to find associations. It identifies patterns found in a patient’s data and compares that to patterns found more broadly. The relationships between observations can be incredibly complex, and include hundreds of possible causes, interactions and non-linear responses. If done properly, the result is a far more accurate predictive model that has the ability to automatically adjust and improve over time. It is the machine-learning equivalent of “clinical intuition” which is so fundamental, and previously so subjective, in clinical practice – the “art of medicine.”