04-07-2026
Preventative health care — from follow-up calls after hospital discharge to lifestyle coaching programs — plays a critical role in improving patient outcomes and reducing costs. Yet one persistent challenge limits the effectiveness of many interventions: decision-makers often lack enough data to personalize care for each patient population.
New research published in Management Science by the Daniels School’s Pengyi Shi and coauthors Xinyun Chen and Shanwen Pu offers a promising solution. In “Data-Pooling Reinforcement Learning for Preventative Healthcare Intervention,” the researchers develop a data-pooling reinforcement learning framework that enables organizations to learn effective intervention strategies even when available data are limited.
Although the study focuses on health care, its insights apply broadly to organizations that must make dynamic decisions with incomplete information.
Many organizations face what researchers describe as a “small-data” problem. Most AI-enabled prediction and decision-support methods require large datasets to perform well, yet in health care, organizations like hospitals may only observe a limited number of patients within a specific program or demographic group. This makes it difficult to build reliable predictive models using local data alone.
At the same time, relying too heavily on historical data from other hospitals or populations can also lead to poor decisions if those groups behave very differently. In practice, leaders often face three imperfect options:
Each approach creates trade-offs between statistical reliability and real-world relevance and is not ideal.
To address this challenge, the researchers developed a reinforcement learning framework that allows decision systems to learn from both historical and newly observed data.
Reinforcement learning works by continuously improving policies based on observed outcomes. In this case, the system evaluates intervention decisions over time and adjusts its strategy as it learns what works best for different types of patients.
The study introduces a novel data-pooling mechanism that determines how much weight to assign to each data source. Rather than treating historical data as either fully reliable or completely irrelevant, the model adjusts its reliance on external data as new observations arrive.
Three design features make the approach especially practical:
Together, these features allow organizations to improve decision-making while addressing both statistical and privacy challenges.
The researchers demonstrate the framework using a case study focused on preventing hospital readmissions after discharge. Hospitals often use interventions such as phone calls, home visits or monitoring programs to reduce the risk of complications.
These interventions are beneficial but resource-intensive. Staff time, operational costs and patient fatigue mean hospitals must carefully determine which patients should receive what types of interventions and when.
Using reinforcement learning, the system tracks patient risk levels and learns which interventions are most effective across different stages of care. Over time, the algorithm improves its recommendations by integrating both historical evidence and new patient outcomes.
The result is a more targeted approach that helps hospitals allocate limited resources more effectively while improving patient outcomes.
Although the study focuses on health care, the insights extend to many business contexts where decisions evolve over time and data may be fragmented or limited.
The research highlights several broader lessons for leaders:
These principles apply to many operational settings, including customer engagement strategies, pricing decisions, supply chain optimization and workforce management.
The core insight from the research is that limited data does not have to limit intelligent decision-making. By combining adaptive learning with thoughtful use of external data, organizations can improve both the speed and effectiveness of their decisions.
For health care providers, this means delivering preventative interventions more efficiently and improving patient outcomes. For organizations more broadly, it suggests a powerful shift in how AI systems learn, moving from isolated datasets toward collaborative, privacy-conscious intelligence built from multiple sources of knowledge.
