Predicting medication adherence using ensemble learning and deep learning models with large scale healthcare data.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Patient Adherence Research

Background:

• Medication non-adherence affects 30-50% of patients, impacting therapeutic efficacy and population-level drug data.
• Accurate prediction of non-adherence is vital for targeted interventions, especially for patients self-administering injectable medications.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting medication adherence in patients using self-administered injectable therapies.
• To leverage event history data from connected sharps bins for accurate adherence prediction.

Main Methods:

• Applied various ensemble learning and deep learning models, including Long Short-Term Memory (LSTM), to predict medication adherence.
• Utilized a real-world dataset of 342,174 injection disposal records spanning over 5 years.
• Evaluated model performance using metrics like accuracy, specificity, sensitivity, precision, F1 score, and ROC AUC.

Main Results:

• The Long Short-Term Memory (LSTM) model demonstrated superior performance in predicting medication adherence.
• The proposed pipeline achieved 77.35% accuracy, with an ROC AUC of 0.8390.
• The best-performing model (LSTM) showed good generalization on a future testing dataset.

Conclusions:

• Machine learning models, particularly LSTM, can effectively predict medication non-adherence in patients using self-administered injectable medications.
• The developed prediction pipeline, utilizing connected sharps bin data, can significantly improve the efficiency of adherence interventions.
• Prioritizing patients at high risk of non-adherence enables timely and targeted support, potentially improving therapeutic outcomes.