Patient Medication Adherence: Using machine learning to predict vulnerable individuals in risk of non-compliance
Medication non-adherence remains a persistent and significant challenge for healthcare systems worldwide, contributing to poorer health outcomes and increased healthcare costs. Although numerous factors influencing adherence have been identified, large-scale, population-based studies that offer comprehensive insights into its multifaceted nature remain limited. With recent advancements in machine learning (ML), new opportunities have emerged to analyse medication non-adherence more effectively using large datasets. This study aims to identify the most important predictors of medication non-adherence. We leverage a comprehensive, population-wide dataset from Denmark to predict non-adherence using ML techniques. The focus is on adults prescribed medications for chronic conditions – specifically ACE inhibitors, beta blockers, oral anti-diabetics, and statins – between 2014 and 2018. Using complete pharmacy refill records, adherence is measured via the medication possession ratio. A wide range of features encompassing patient, provider, and structural characteristics is incorporated. We integrate extreme gradient boosting (XGBoost) with the least absolute shrinkage and selection operator (LASSO) method to achieve optimal predictive performance (AUC = 0.94). Shapley additive explanation (SHAP) values reveal that prescription days’ supply, medication strength, treatment duration, and price are among the strongest predictors. Our findings demonstrate the advantage of ML in accurately identifying features related to non-adherence and offer new insights into identifying patient groups at elevated risk.
Speaker profile
Maiken Skovrider Aaskoven is an Assistant Professor at the Danish Centre for Health Economics at University of Southern Denmark. Her research interests include investigating determinants of individual health (behaviours) and responses to detriments in health; examining how health impairments spill over to family members and why they have heterogenous responses; and identifying factors contributing to individual wellbeing. She employs a variety of methods, focusing on linking survey data with individual register data and using quasi-experimental approaches based on register data.
Weekly seminar series
As part of our Centre's vibrant research culture, we host a weekly seminar series. Visiting and invited researchers present current research relating to the economics of health and wellbeing, and the healthcare sector. Visitors are welcome to join these sessions where discussion and debate is encouraged.
For further information on our seminar series, please contact Trong-Anh.Trinh@monash.edu.
Event Details
- Date:
- 21 May 2025 at 12:00 pm – 1:00 pm
- Venue:
- Caulfield campus, Building H, level 8, room H813
- Categories:
- CHE Seminar; General
Description
Medication non-adherence remains a persistent and significant challenge for healthcare systems worldwide, contributing to poorer health outcomes and increased healthcare costs. Although numerous factors influencing adherence have been identified, large-scale, population-based studies that offer comprehensive insights into its multifaceted nature remain limited. With recent advancements in machine learning (ML), new opportunities have emerged to analyse medication non-adherence more effectively using large datasets. This study aims to identify the most important predictors of medication non-adherence. We leverage a comprehensive, population-wide dataset from Denmark to predict non-adherence using ML techniques. The focus is on adults prescribed medications for chronic conditions – specifically ACE inhibitors, beta blockers, oral anti-diabetics, and statins – between 2014 and 2018. Using complete pharmacy refill records, adherence is measured via the medication possession ratio. A wide range of features encompassing patient, provider, and structural characteristics is incorporated. We integrate extreme gradient boosting (XGBoost) with the least absolute shrinkage and selection operator (LASSO) method to achieve optimal predictive performance (AUC = 0.94). Shapley additive explanation (SHAP) values reveal that prescription days’ supply, medication strength, treatment duration, and price are among the strongest predictors. Our findings demonstrate the advantage of ML in accurately identifying features related to non-adherence and offer new insights into identifying patient groups at elevated risk.
Speaker profile
Maiken Skovrider Aaskoven is an Assistant Professor at the Danish Centre for Health Economics at University of Southern Denmark. Her research interests include investigating determinants of individual health (behaviours) and responses to detriments in health; examining how health impairments spill over to family members and why they have heterogenous responses; and identifying factors contributing to individual wellbeing. She employs a variety of methods, focusing on linking survey data with individual register data and using quasi-experimental approaches based on register data.
Weekly seminar series
As part of our Centre's vibrant research culture, we host a weekly seminar series. Visiting and invited researchers present current research relating to the economics of health and wellbeing, and the healthcare sector. Visitors are welcome to join these sessions where discussion and debate is encouraged.
For further information on our seminar series, please contact Trong-Anh.Trinh@monash.edu.