Risk Stratification and Prediction for Tailored Population Health Management Interventions

Shan Xie, Purdue University

Abstract

The need to shift from volume to value-based care model has encouraged healthcare providers to move towards an approach known as “population health management (PHM).” The main concept of PHM is to maintain or improve the well-being of individuals by better identifying and monitoring individuals to address health needs at all points along the care continuum through cost-effective and tailored health solutions. Care coordination and chronic disease management are the two important aspects of PHM as the targeted population is usually people with high health risk and high healthcare spending. Risk stratification and prediction techniques can help to identify patients who would be most likely to benefit from care coordination and disease management programs and provide insights to inform better decisions regarding intervention design. Under this scope, specific population such as rural residents or certain care concerns such as hospital readmission and medication nonadherence are understudied.This research evaluates and identifies risk stratification and prediction strategies that could help target at-risk population more efficiently with regard to three areas under care coordination and chronic disease management. The three areas include 1) emergency department (ED) utilization in critical access hospital (CAH); 2) thirty-day all-cause hospital readmissions; and 3) medication adherence behavior for type 2 diabetes patients. The application is demonstrated using administrative and medical claims data. Results showed that, for CAHs, over 50% of the ED visits could be treated at primary care settings which indicates a shortage of primary care physicians in rural areas and opportunities for improved patients’ continuity of care. We also found a significant proportion of patients with ED transfers returned to their local community within a relatively short time frame, which suggests the need to improve care coordination between the local CAH, the transferring hospital as well as other community resources to ensure proper follow-up care. For larger hospitals with readmission concerns, additional evaluation metrics and data imbalance issues might be worth considering for boosting the predictive performance of prediction models that estimate 30-day readmission risk. The recent machine learning techniques were found to perform better than the conventional logistic regression. Finally, group-based trajectory models provided value in identifying different medication adherence behavior for patients with type 2 diabetes and helping providers make better decisions. Overall, this work illustrates some limitations of the current methods and identifies opportunities for more accurate approaches to help design targeted interventions and fulfill the goal of PHM.

Degree

Ph.D.

Advisors

Yih, Purdue University.

Subject Area

Engineering|Information science|Health care management

COinS