Artificial intelligence meets population health. What data challenges need to be addressed?

Artificial intelligence has the undivided attention of many segments of the healthcare industry. This technology is regarded as a pathway to automate tasks that could improve patient monitoring, particularly to spur earlier intervention for patients that may otherwise end up in the emergency department for avoidable reasons.

The application of AI tools for chronic conditions is an area of keen interest for Geneia’s Data Intelligence Lab. The group is focused on AI and data science techniques to create models that reduce costs for health plans, physicians and employers through earlier identification and engagement of those at risk of high-cost conditions such as heart failure and diabetes.

MedCity News recently spoke with Fred Rahmanian, Chief Analytics and Technology Officer of Geneia about how the Data Intelligence Lab is serving as a test bed for these models, particularly for diabetes.

One of the biggest stumbling blocks for the adoption of AI in healthcare seems to be the quality of data and transparency about sources of data that institutions use to “teach” their algorithms. Do you agree or disagree?

At Geneia we are using both publicly available and private sources of data. It is a huge volume of data and is increasingly varied. The quality of wearable data is becoming better. Even the heart rate data we get from Apple Watch is more reliable than it used to be, as demonstrated by the recent FDA clearance the company received. Quantified self data from Fitbits, Garmins and other wearables, and even cell phone data – can provide data on the users’ level of activity, GPS, ambient temperature, elevation, etc. These can then be used as a proxy for other types of information. For instance, GPS can point to a location from which we can infer the weather and allergy conditions. In general, the variety of data is improving data quality and the models we’re creating with that data.

More data sources are becoming available. I do agree that the quality of data is quite important and evaluating data quality is part of the process we go through when exploring new areas. We have developed robust processes in-house to help us improve data quality in a manageable amount of time. As for best practices for sourcing data, you need to use trusted sources and validate the data as we do at Geneia. As a part of the validation process, you need to understand that biases in data can result in biases of algorithms and outcomes. That is something we take very seriously.

What kinds of data/data sources does the Geneia Data Intelligence Lab use to develop its algorithms?

Claims data, clinical data, data from lab results, doctor visits, pharmacy data. Data derived from social determinants of health is another important source; it’s publicly-available at the zip code, county and state levels. Air quality, for example, is a commonly cited source of SDoH data, but it’s just as important to know whether a shaggy rug in an elderly person’s home increases the patient’s  fall risk.

Some of the other data sources that are quite rich that we have started to use are patient reported outcomes. Patient surveys are data rich and are used for analysis.

I have never seen a piece of data I didn’t like. As I often say, what is noise today may be a signal tomorrow.

Diabetes has been a particular area of interest. Can you give a specific use case for how the Geneia Data Intelligence Lab has applied its predictive analytics tech to address this issue?

Diabetes seems to be a starting point for many other comorbidities in a patient and can be very costly. We are doing a lot of work in this area.

One focus of our work is identifying people who have not developed it yet but are predisposed to becoming Type 2 diabetic — the sooner we get the diagnosis, the sooner the healthcare system can work to engage the person resulting in better outcomes and ultimately lower costs. This benefits the patient, their physician, the health system, and the health plan. Earlier identification means primary care physicians are more likely to be able to manage the people who are predisposed to diabetes.

What other models is Geneia developing?

One area we are looking into is how can we use recent advances we have seen in deep learning and apply them in population health management. Deep learning has been used in medical imaging but only recently has it been applied to structured data such as for care management and utilization management. It impacts much of what we do.

As examples of things we’re doing at the research stage, we’re working on better ways of identifying and stratifying patients according to their risk level. We are also interested in addressing inappropriate emergency department overuse. How do we prevent avoidable ED use by, for example, improving drug adherence? What happens to ED utilization if we identify people who have certain chronic conditions and prevent complications from chronic kidney disease?

High risk pregnancy is another area of interest — how do we predict pregnancies that have complications, and from there identify them early enough to prevent complications?

What are some milestones for the program in the next 12 months?

For diabetes in particular, it is not just about identifying people with diabetes. Most models have a long time horizon for prediction of an event, such as the next 12 months. It would be more helpful for organizations with a care management team to predict events sooner.

Every organization has limited resources so they need to know where best to allocate those resources for the biggest impact. We are looking into identifying people who would develop the condition in a shorter time span and group them by time span such as “within the next three months” or the “next three to six months.” Then healthcare organizations can prioritize those patients. The ability to allocate resources to the right people at the right time is just as important as is the ability to adjust resource availability. We think it is a big deal. This is an area we are concentrating more at a population level compared with more targeted groups.

Calling patients or having them come in for a doctor visit is a costly method of risk mitigation. With the Geneia Data Intelligence Lab, it is not just about looking at what is the norm. We are trying to develop new, better ways to engage patients and help our customers, but not every idea pans out. Some prove to be impractical or too costly.

Photo: Peter Howell, Getty Images