AI in HEOR: Reflections on Early Innovations

My involvement in AI and predictive modeling within Health Economics and Outcomes Research (HEOR) began during a pivotal time in the 2000s when traditional statistical modeling was increasingly complemented by machine learning techniques. Initially, predictive modeling focused on straightforward statistical tools to forecast healthcare utilization and costs using large healthcare datasets, primarily insurance claims data. Although these early methods weren't called "AI," they laid essential groundwork by enabling data-driven patient risk stratification.

Early Risk Stratification Models

In the 1990s and early 2000s, foundational tools such as the Chronic Disease Score and Johns Hopkins' Adjusted Clinical Groups (ACGs) were developed to classify patients by risk: the Chronic Disease Score from automated pharmacy data, and ACGs by diagnosis-based morbidity patterns. A 2005 Medical Care study I co-authored with Powers and colleagues used the Pharmacy Health Dimensions (PHD) index, a pharmacy claims-based risk index, highlighting the effectiveness of pharmacy claims data in predicting total healthcare expenditures. Remarkably, a simple ordinary least squares (OLS) model provided predictive accuracy similar to more complex econometric methods, underscoring an important lesson: detailed data could outperform sophisticated algorithms.

Innovations at Caremark

During my tenure at Caremark (later CVS Caremark) in the mid-2000s, we took predictive modeling a step further by incorporating medication adherence into pharmacy-based predictive models. Our team employed boosted regression trees, an advanced machine learning technique, to predict healthcare costs more accurately. Our 2006 study, which we presented at a conference, showed that integrating adherence metrics yielded incremental gains in predictive accuracy, showcasing AI's potential to capture complex interactions among medication use patterns. However, it also taught us crucial lessons about the risks of overfitting, highlighting the need for rigorous validation.

These experiences illustrated that machine learning could enhance traditional actuarial models by detecting nonlinear relationships and nuanced patient behaviors. My work, alongside contemporaries exploring neural networks and decision trees, signaled a pivotal shift in HEOR analytics toward more sophisticated, data-driven methods.

Patient-Reported Outcomes Integration

Another notable trend during this period was the integration of patient-reported outcomes (PROs). Researchers began applying clustering methods to patient-reported symptom and quality-of-life data, expanding predictive modeling beyond claims and utilization data. While computational limitations and data availability constrained early efforts, these pioneering attempts hinted at today's rapidly evolving use of AI in personalized medicine and outcome prediction.

Key Insights from Early AI in HEOR

Reflecting on these early milestones, several key insights stand out:

• ●Pharmacy claims alone predicted total healthcare costs well, and simpler models performed about as well as complex econometric ones.
• ●Machine learning methods, such as boosted regression trees, offered valuable but incremental improvements over traditional statistical models.
• ●The complexity of AI models required careful handling to ensure generalizability and prevent overfitting.

Ultimately, my early work at Caremark, along with foundational contributions by other researchers, set the stage for the sophisticated predictive analytics and AI-driven approaches that now characterize modern HEOR. Understanding this journey highlights the evolution and ongoing potential of AI in healthcare research.