AI and HEOR: A Joining of Forces

There is no shortage of applications when it comes to artificial intelligence (AI) in health care. The use of AI to predict medical events is an application in which we have seen tremendous growth, solidifying the value. What are some of the medical events that AI can help to identify? Where does health economics and outcomes research (HEOR) come into play?

• AI in Disease Detection: Identifying patients at risk of having an undiagnosed or under-reported disease is a frequent challenge with rare diseases, but it can also be applied in more common conditions.

  • HEOR’s role: Once underdiagnosed patient cohorts are identified, the burden of illness can be quantified to understand the cost and patient journey for treating the rare disease. Such results can aid in physician targeting and educating the public, investors, etc.

• AI in Patient Adherence: Identifying key drivers of non-adherence to guide targeted adherence strategies by understanding reasons why patients discontinue therapy (eg, whether they do so due to medication load, complexity of the treatment, etc.)

  • HEOR’s role: Researchers can leverage those patients who are adherent/non-adherent and determine if adherence leads to a reduction in health-care costs (and other such economic outcomes), which could then provide evidence for subsequent adherence improvement strategies, patient support programs, and physician education.

• AI in Disease Progression: Predicting patients whose disease will progress faster, especially for diseases with varied progression timelines, can help identify those patients who could benefit from more advanced treatment or closer monitoring.

  • HEOR’s role: The cost of delayed treatment and identifying patients who might benefit from earlier treatment are key areas of research that can add to the peer-reviewed literature.

• AI in Adverse Events (AE) Management: Identifying patients who are at risk of experiencing an AE can help guide proactive interventions to minimize the risk of AEs.

  • HEOR’s role: Once those patients are identified, the cost (and incidence) of the potential AE can be captured and reported.

• AI in Hospital Readmissions: Predicting 30-day readmission can lead to improving quality of care for patients, thus minimizing the risk of readmission.

  • HEOR’s role: Hospitalizations often comprise a large segment of health-care expenditures. By evaluating the cost of such a readmission, as well as potential post-discharge costs and the hazard/risk of incurring such costs, integrated delivery networks and other health-care systems can plan for such encounters within the health-care system.

A Bird’s Eye View: How AI Algorithms Work

At a high level, the foundations of how an AI algorithm works are the same regardless of the disease or the medical event. However, let’s describe how the algorithm would work for a specific example of finding undiagnosed patients with a rare disease.

The first, and frequently overlooked, step is to understand how the algorithm would be used once built and leverage this information to select a health-care dataset that most resembles the deployment environment. The next step is to identify patients diagnosed with a rare disease in the chosen dataset. This can range from a simple, clean diagnosis code, to a combination of confirmatory test results, to a need to link in data from a rare disease patient registry. Identifying a control cohort to include patients who do not have the rare disease of interest is just as important.

Once the cohorts are built, the next step includes pulling all medical history of the selected patients, including diagnoses, procedures, treatments, tests, demographics, etc. This can be done in two complementary ways. A researcher can compile a list of hypotheses, or medical events that are expected to be strong predictors. Alternatively, we can allow the algorithm to analyze all possible medical events – often tens of thousands of unique events and their interactions – and surface to the top the features that are indeed important predictors of undiagnosed patients.

When the data is assembled, the algorithm learns the many different patterns of medical histories while trying to identify the ones most predictive of under-diagnosis. While there are many software packages that can build an out-of-the-box algorithm in minutes, there are countless study design considerations not to be overlooked. These include, but are not limited to, careful subsampling of data with class imbalance, projection of model precision based on expected incidence of the disease, identification and correction of algorithmic bias, clinical interpretation, and validation of the algorithm.

Once the algorithm is built, we return to where we started: the planned deployment. An AI algorithm can be used to score medical data of hundreds of millions of patients on a routine basis. This can be done in secondary data to get better insight into the size of the undiagnosed patient population and key misdiagnosis patterns, or at the point of care to identify patients who should undergo diagnostic screening. Recent changes in interoperability standards provide hope that AI can soon become an essential part of informing real-time clinical care.

“Finding undiagnosed patients with hepatitis C infection: an application of artificial intelligence to patient claims data,” is just one detailed publication-backed example that showcases real application of these concepts.

Parting Thoughts

The use of statistical foundations to underpin any AI algorithm has been in practice for many decades. What is ‘new’ is the excitement around AI in health-care driven by the availability of diverse data (and the ability to link them to one another), technological advancements, and improved precision driven by methodological AI innovation.

While this is an exciting time, we should exercise caution in understanding where AI is truly adding value. We should also be following key quality measures to assure that algorithms are properly validated, as well as adopting interpretability techniques to make both researchers and clinicians feel comfortable about the frequent “black box” nature of algorithms.

Blog Post Authors:

• Nadea Leavitt, Senior Director, AI for Healthcare & MedTech, IQVIA
• Mitchell DeKoven, MHSA, Senior Principal – Health Economics/Outcomes Research, IQVIA