Medication adherence prediction through temporal modelling in cardiovascular disease management

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Cardiovascular Disease Management

Background:

• Chronic conditions like cardiovascular disease (CVD) pose a significant healthcare burden globally.
• Medication non-adherence is a persistent challenge in managing long-term conditions, despite effective therapies.
• Predicting adherence is crucial for optimizing CVD risk reduction strategies.

Purpose of the Study:

• To evaluate the benefits of integrating patient history using temporal models for medication adherence prediction.
• To compare the predictive performance of deep learning models against traditional methods.
• To assess the impact of observation window length on adherence prediction accuracy.

Main Methods:

• Utilized a large cohort (564,180 patients) linked to national health datasets.
• Defined medication adherence using Proportion of Days Covered (PDC) ≥ 80%.
• Compared temporal models (LSTM, Simple RNN) with non-temporal models (MLP, RC, LR) using two years of historical data to predict five-year adherence.

Main Results:

• Temporal models significantly outperformed non-temporal models in predicting medication adherence.
• Long short-term memory (LSTM) achieved the highest predictive performance (ROC AUC 0.805).
• Increasing the observation window length further enhanced the predictive advantage of temporal models.

Conclusions:

• Deep temporal models effectively integrate patient history for improved medication adherence prediction.
• Recurrent Neural Network (RNN) architecture, specifically LSTM, offers superior predictive capabilities over other models.
• Leveraging sequential patient data with advanced models is key to addressing non-adherence in chronic disease management.