Machine Learning to Predict Noncompliance and Program Dropout in Patients Treated for Chronic Diseases: Unsupervised and Supervised Analyses From a Large Multinational Drug Access Program Database
Author(s)
Etienne Audureau, MD, PhD1, Joel Ladner, MD, PhD2, Ben Davis, .3, Joseph Saba, MD3.
1Clinical Epidemiology & Ageing, Paris Est Université, Hôpital Henri Mondor Hôpital, Public Heath, Assistance Publique Hôpitaux de Paris, Créteil, France, 2Department of Epidemiology & Health Promotion, Rouen University Hospital, Rouen, France, 3Axios International, Paris, France.
1Clinical Epidemiology & Ageing, Paris Est Université, Hôpital Henri Mondor Hôpital, Public Heath, Assistance Publique Hôpitaux de Paris, Créteil, France, 2Department of Epidemiology & Health Promotion, Rouen University Hospital, Rouen, France, 3Axios International, Paris, France.
OBJECTIVES: Patient non-compliance with treatment remains a global challenge, contributing to increased morbidity, mortality, and healthcare costs. Limited insight into underlying causes hampers targeted intervention design. This study applies machine learning to a large multi-country dataset to identify predictors of treatment non-compliance and program dropout in patients with chronic diseases.
METHODS: We analyzed data from 45,676 patients enrolled between 2016-2024 in pharmaceutical access programs across 18 low- and middle-income countries, covering 62 drugs in 11 therapeutic areas. Descriptive and unsupervised clustering analyses were performed to evaluate correlations between individual- (demographics, clinical characteristics), program- (kind and extent of interactions between patients and program stakeholders) and country-level (GDP) characteristics, and their associations with compliance (defined as the proportion of prescribed medication cycles obtained on time) and program dropout rates. Supervised predictive modeling was then performed using multivariate regression and machine learning random forest algorithms to identify predictors of compliance and program dropout, ranking factors by their relative variable importance (VIMP).
RESULTS: Preliminary analysis found that the main disease areas were oncology (32.3%), dermatology (22.5%) and rheumatology (22.3%). Mean age was 46.4 years (±standard deviation 18.7; sex ratio M:F=0.96; median follow-up 9.5 months). The overall mean compliance was 60.4%±22.5 with a global dropout rate of 40.9%. Unsupervised analyses revealed contrasted patterns of compliance and dropout rates across countries and disease areas. Random forest analyses confirmed disease area and country as major predictors of compliance and dropout, and identified the number of interactions between patients and access program stakeholders at key time points (pre-enrolment, enrolment, follow-up) as influential predictors of treatment compliance. Final results displayed at ISPOR Europe 2025.
CONCLUSIONS: These findings underscore the value of machine learning in identifying factors contributing to non-compliance and program dropout. The insights generated can support the design of tailored interventions to improve treatment compliance and retention across diverse patient populations.
METHODS: We analyzed data from 45,676 patients enrolled between 2016-2024 in pharmaceutical access programs across 18 low- and middle-income countries, covering 62 drugs in 11 therapeutic areas. Descriptive and unsupervised clustering analyses were performed to evaluate correlations between individual- (demographics, clinical characteristics), program- (kind and extent of interactions between patients and program stakeholders) and country-level (GDP) characteristics, and their associations with compliance (defined as the proportion of prescribed medication cycles obtained on time) and program dropout rates. Supervised predictive modeling was then performed using multivariate regression and machine learning random forest algorithms to identify predictors of compliance and program dropout, ranking factors by their relative variable importance (VIMP).
RESULTS: Preliminary analysis found that the main disease areas were oncology (32.3%), dermatology (22.5%) and rheumatology (22.3%). Mean age was 46.4 years (±standard deviation 18.7; sex ratio M:F=0.96; median follow-up 9.5 months). The overall mean compliance was 60.4%±22.5 with a global dropout rate of 40.9%. Unsupervised analyses revealed contrasted patterns of compliance and dropout rates across countries and disease areas. Random forest analyses confirmed disease area and country as major predictors of compliance and dropout, and identified the number of interactions between patients and access program stakeholders at key time points (pre-enrolment, enrolment, follow-up) as influential predictors of treatment compliance. Final results displayed at ISPOR Europe 2025.
CONCLUSIONS: These findings underscore the value of machine learning in identifying factors contributing to non-compliance and program dropout. The insights generated can support the design of tailored interventions to improve treatment compliance and retention across diverse patient populations.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR142
Topic
Methodological & Statistical Research, Patient-Centered Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics