Perfecting the Knowledge Flow for Medical AI

Medical knowledge spans the history of humankind. From the first herbal treatments and natural remedies to the nanobots and genetic therapies of today, there has always been a process by which new medical knowledge is evaluated, its worth defined, and its application spread to other healthcare practitioners.

Hundreds of years ago, this process was defined more by first-hand observation and intuition then by any kind of objective analysis. Today, objective analysis increases at an exponential rate through wide-ranging data-collection capabilities, now processed through artificial intelligence (AI).

But while medicine becomes more objective, the diffusion of new information still has its challenges. Research shows that it takes up to 17 years for medical evidence to reach routine clinical practice. The fact is that there are some blind spots in our data-driven world, and there are some underlying weaknesses in healthcare delivery that can be traced to deficiencies in the adoption, distribution, and application of emerging knowledge.

How Knowledge Moves Within Systems of Care

Let’s look at how knowledge dissemination has evolved over the years. Knowledge, as it becomes more complex and more specialized, further distributes among more and more participants in the value chain. The creation of new medical knowledge results in the emergence of new types of professionals to take on the new work.

In centuries past, one care practitioner would meet with a patient, evaluate their condition, and prescribe or perform a treatment. Now, a patient may meet with general physicians, cardiologists, nurses, physiotherapists and healthcare navigators. Add to this a huge supply chain of healthcare service providers, including everything from patient-centered medical homes to hospitals.

In this specialized state, an effective care process happens because a patient’s unique needs were matched to the body of medical knowledge relating to the patient’s disease. This match only works if data about the patient’s status is easily accessible and actionable.

But this is a challenge today, as data tends to proliferate in 2 dimensions. The first dimension consists of data generated at all points of care within the system of care. This takes on broader perspectives such as that of population health. The second dimension is more introspective, with data generated by the patient’s own condition with a continuously proliferating array of lab results, EKG, blood pressure readings, genomic tests, etc.

But these two separate dimensions must align into 1 point-of-view that is easily accessible at the moment of care. Data generated anywhere in the system must be available when it is needed and in whichever workflow it is needed in.

How to Make Knowledge Flow Freely

So, what does this mean when designing the algorithms behind AI and the predictive modeling that it enables? The best answer actually comes from an examination of the success driving a different industry: transportation and logistics. A parallel can be drawn between the movement of knowledge and how appropriate packaging facilitates movement of a physical product.

A quick history lesson: Until the 1950s, the shipping industry was constrained by long ship loading/unloading times, cumbersome transfers and intensive labor at marine ports. Malcolm Mclean, an entrepreneur, believed that the freight business could be made more efficient only if goods could be moved seamlessly from the source to the destination, and had his engineers design a container that could be filled up at source, carried by rail or road and then transferred directly onto ships. It enabled ships to be unloaded and loaded within a few hours, and was eventually universally adopted to make cargo movement less about ports and more about how fast the cargo can be moved from source to destination.

This design principle can be applied to both medical knowledge and data. Medical best practices, such as clinical guidelines, decision trees and checklists can be packaged as simple or complex algorithms and kept relatively independent of the contexts within which they will be used. This separation essentially allows knowledge to be distributed based on the needs of the market. This relative independence of the algorithm could result in different implementations—as a drop down in an EMR, a decision process in a population health app, or a patient portal—but the key point is that this availability enables managers to insert knowledge into the appropriate workflow where it can yield the best results for their patients.

Think about how a decision support algorithm in chronic wound care could evolve. Management decisions are typically based on visual examination by trained doctors or APRNs at wound clinics. With the availability of automated wound analysis tools to objectively assess wound healing, this decision making task could shift to home visit nurses or skilled nursing facilities. These objective tools could be backed up by decision tree algorithms at different points in the care process. The decision trees act like treatment recipes for non-physician care practitioners to follow.

The core principle at work here is the acceleration of the adoption, distribution, and application of medical knowledge by bringing data and algorithms together in the right workflow, at the right time and to the right user. If AI is built and operated according to this principle, then it will yield exponential benefits in both outcomes and efficiencies of care.

More from Sawad Thotathil:

3 Keys to Cost-Conscious Population Health Management

What Tesla's Electric Truck Can Teach Us About Healthcare's Digital Transformation​

Personalizing Healthcare: The Netflix Approach

Dr. Sawad Thotathil, clinician and senior director, Persistent Systems, directs product strategy and management for the company’s Digital Healthcare group, helping health systems achieve their digital transformation goals.