By Gurjeet Singh

Artificial intelligence (AI) and machine learning have become catchphrases for the latest generation of vaporware. Almost any intelligent application or analytic software can now be classified as “AI”, especially if it can be applied to large data sets which, by definition, are what health care runs on. But in health care, the term AI refers not only to technology, but also to a specific approach that must follow certain rules to have any real impact on clinical care.

Take population health management, which is being increasingly adopted by health care organizations so that they can succeed under value-based reimbursement. To maintain and improve the health of their patients, health care organizations need a way to understand and harness the huge amounts of data that could potentially be applied to achieving these goals. Artificial intelligence systems can derive actionable insights from large, complex data sets at the scale required by health care enterprises. AI can also uncover subtle predictive trends that traditional analytics platforms may miss. But it can do this only if it is deployed in the right way.

For starters, an AI solution must aggregate and normalize the financial and clinical data from health care information systems, along with claims data from payers, in some kind of cloud-based infrastructure such as the Hadoop framework. The AI software must be live in an organization’s clinical and business processes, and must be able to process data in near-real time to be of value in clinical and financial decisions.

Beyond that, there are five components that an AI platform needs in order to deliver the results that health care organizations seek. These can be summarized under the headings of Discover, Predict, Justify, Act and Learn, as follows.

Discover
An AI platform must be capable of performing unsupervised learning. Unsupervised learning is critical because, in large and complex data sets such as those in health care, the odds of asking the right question of your data is effectively zero. AI needs to discover all of the patterns or relationships that exist in the data, independent of human input.

For example, a health care organization might employ AI to automatically discover groups of patients who share certain kinds of characteristics. These groups, e.g.: low-income, opioid-addicted, obese patients who live alone and have two or more chronic diseases, might be targeted with personalized interventions and care paths. AI can identify these kinds of subgroups without being told what to look for. This can dramatically increase a doctor’s ability to craft care plans for people in this subpopulation. It can also help health care organizations design customized campaigns to address the medical and social needs of these patients.

Predict
By applying supervised learning methods to large datasets, AI can predict what is likely to happen in the future with an exceptionally high degree of accuracy. In health care, predictive analytics can be used to stratify patient populations by their health risks; to predict which individuals will get sick or sicker, or have catastrophic health events; to predict the probable effects of various interventions on particular patients; and to analyze the financial impacts on the organization of the disease burdens within its population and subpopulations.

Clinicians remain the key decision-makers. By providing this superior level of predictive ability, however, AI can give them and their organizations valuable insights into the future needs, costs, disease burdens, and risks of patients.

Justify
Doctors and nurses are rightfully skeptical about the recommendations of computer algorithms, particularly when they diverge from a provider’s experience and existing standards of care. Therefore, the results of AI modeling must include explanations of how the model arrived at its conclusions in terms that are familiar to clinicians. For its predictions to have value, AI must be able to justify and explain its assertions, and be able to diagnose its failures.

For instance, AI solutions must always provide reasons to support decisions on classifying patients as high or low risk. Without a thorough explanation of the variables that AI used in its risk stratification, there is no reason why clinicians should trust it. AI must always justify its predictions, discoveries, and actions so clinicians can feel confident enough about its recommendations to act on them.

Act
AI recommendations must also be actionable to be of any use in clinical decision support. For example, if AI states that a particular care path is best for a certain patient, the steps required to implement that care path must be provided, along with an explanation of why it is likely to produce an optimal outcome.

AI must provide these insights quickly, so that clinicians can act on them when they make medical decisions and devise care plans. Consequently, intelligent applications based on AI must be embedded into the clinical workflow. They must ingest new data and must automatically execute the loop of Discover, Predict, and Justify, frequently enough to provide value in the clinical decision process.

Learn
Intelligent systems are designed to detect and react as the data evolves. That means they are able to learn from changes in data patterns as well as the results of their recommendations. An intelligent system is always learning and constantly improving.

AI learns from the data and from feedback about its errors. As it discovers trends, predicts events, and recommends actions, it must constantly improve to be of value to health care organizations. AI must also be designed to create a positive user experience, which facilitates both decision support and the collection of feedback to facilitate continuous learning.

Conclusion
This framework represents a starting point for any health care organization looking to deploy artificial intelligence. Whether it is evaluating a point solution for radiology or a platform on which to build specific applications, these five components should be present. By deploying AI along the lines described here, health care organizations can accelerate their journeys to success in value-based care.



About the Author: Gurjeet Singh is executive chairman and co-founder of Ayasdi, an advanced analytics company.