How Data-Driven Decisions Can Improve Health Outcomes

Dr. Ilbin Lee, an assistant professor at the Alberta School of Business in the Department of Accounting and Business Analytics, is pioneering research on how modeling different subpopulations can enhance decision-making processes, especially in healthcare. His recent paper, Is Separately Modeling Subpopulations Beneficial for Sequential Decision-Making?, investigates how tailoring decision strategies to subpopulations can improve patient outcomes.

Sequential decision-making is a process where decisions are made over time, with each choice influencing future outcomes. “A good example is treatment planning for chronic disease,” explains Lee. “Suppose that a patient with a chronic disease sees a doctor and gets a blood test regularly, say every three months, to monitor the disease. At each appointment, the doctor assesses the patient’s status and determines the course of action for the next three months, such as drug dosage. This action will affect the patient’s future status at the next appointment.”

Lee emphasizes that considering subpopulations within this framework is vital. “Your optimal decision rule will be different depending on how your disease status changes,” he notes. Patients with varying rates of disease progression need tailored treatment plans. For instance, those whose conditions worsen rapidly might need aggressive treatments, while patients with slower progression could benefit from a more moderate approach to avoid side effects.

Lee’s motivation for exploring this area stems from the growing availability of large-scale health data, which presents opportunities to improve decision-making.

“The main question of this paper (‘Should we model subpopulations separately or not’) has been considered in the literature, but the decision has relied mostly on the medical domain knowledge. I wanted to develop theory and methods to make this decision based on data.”

The research identifies key factors influencing the benefits of stratifying subpopulations. One crucial factor is the similarity in transition patterns between subpopulations. “If two groups behave very similarly, then we cannot expect much benefit from stratifying, because optimal decisions for those two groups are not likely to differ,” says Lee. However, when subpopulations behave differently, stratification can lead to more effective treatment plans.

Another factor is the data size for different subpopulations. “Even when two groups behave differently in observed data, if their data sets are too small, then we do not have much confidence in our estimated models, which makes it unclear whether stratifying can be beneficial,” Lee points out.

To determine the benefits of modeling subpopulations separately, Lee’s study introduces an algorithm that estimates the performance measure difference when stratifying subpopulations versus not stratifying them. This involves evaluating outcomes such as average patient health improvements.

A practical application of this research is in deciding the optimal time to initiate a therapy for chronic disease patients. “We should begin the treatment earlier for those patients who tend to get worse faster, whereas we might want to wait more for those who deteriorate more slowly,” Lee explains. His findings show that depending on the transition patterns of subpopulations, optimal decision rules can vary significantly.

Beyond healthcare, Lee’s research holds potential applications in any field where sequential decision-making is critical. By applying data-driven methods, researchers and practitioners can refine and optimize decision-making processes, leading to more personalized and effective outcomes.

Lee's work highlights the importance of using large-scale data to inform strategic decisions in healthcare and beyond. "There is a great potential to improve our decisions on the main question of the paper by using large-scale health data, which are becoming increasingly more available," he concludes, paving the way for future exploration into data-driven decision-making.