Abstract
Objectives This study analyzed patient factors in medication persistence after discharge from
the first hospitalization for cardiovascular disease (CVD) with the aim of predicting
persistence to lipid-lowering therapy for 1 to 2 years.
Methods A subcohort having a first CVD hospitalization was selected from 313,207 patients
for proportional hazard model analysis. Logistic regression, support vector machine,
artificial neural networks, and boosted regression tree (BRT) models were used to
predict 1- and 2-year medication persistence.
Results Proportional hazard modeling found significant association of persistence with age,
diabetes history, complication and comorbidity level, days stayed in hospital, CVD
diagnosis type, in-patient procedures, and being new to therapy. BRT had the best
predictive performance with c-statistic of 0.811 (0.799–0.824) for 1-year and 0.793 (0.772–0.814) for 2-year prediction
using variables potentially available shortly after discharge.
Conclusion The results suggest that development of a machine learning-based clinical decision
support tool to focus improvements in secondary prevention of CVD is feasible.
Keywords
medication persistence - cardiovascular diseases - machine learning - prediction models
