Work in Progress: Bridging gaps in chronic disease research: Predictive modeling with health administrative data

Belal Hossain
Statistician, Advancing Health
PhD Candidate, School of Population and Public Health, UBC

Hanna Frank
PhD Candidate, School of Population and Public Health, UBC

Health administrative data offers a rich resource for understanding and predicting outcomes in chronic disease populations. However, its use is often limited by missing or unobserved clinical predictors and the risk of overfitting in predictive models. In this work, we present two complementary approaches that leverage health administrative data to address these challenges. First, we demonstrate how high-dimensional prediction models can compensate for the lack of known clinical predictors in tuberculosis (TB) research by incorporating additional healthcare variables from linked databases. Using Cox-LASSO with cross-validation, we improve the prediction of long-term mortality risk in TB patients, even in the absence of key clinical variables like smoking or alcohol use. Second, we develop a disease-specific comorbidity index for multiple sclerosis (MS) using machine learning techniques, such as LASSO and random survival forests, to identify interactions between comorbidities and optimize for outcomes like treatment initiation and mortality. Both approaches emphasize the importance of internal and external validation to ensure generalizability and the use of time-dependent c-statistics and calibration slopes to evaluate model performance. By addressing the limitations of health administrative data and leveraging advanced statistical methods, these studies highlight the potential of predictive modeling to improve risk stratification, clinical decision-making, and patient outcomes in chronic disease populations