Applying Machine Learning Models to Predict Medication Nonadherence in Crohn's Disease Maintenance Therapy

Area of Science:

• Gastroenterology and Hepatology
• Medical Informatics
• Pharmacology

Background:

• Medication adherence is critical for managing Crohn's disease (CD), but rates remain low.
• Predictive models can help identify patients at risk of nonadherence to azathioprine (AZA).

Purpose of the Study:

• To develop and compare machine learning models for predicting AZA nonadherence in CD patients.
• To identify key factors associated with AZA nonadherence.

Main Methods:

• A cross-sectional study of 446 CD patients prescribed AZA.
• Development and comparison of logistic regression (LR), back-propagation neural network (BPNN), and support vector machine (SVM) models.
• Assessment using accuracy, recall, precision, F1 score, and AUC.

Main Results:

• The SVM model achieved the highest accuracy (87.7%) and AUC (0.930).
• Key predictors of nonadherence included medication concern beliefs, lower education, anxiety, and depression.
• Medication necessity beliefs and knowledge were protective factors.

Conclusions:

• Machine learning models, particularly SVM, show promise in predicting AZA nonadherence in Chinese CD patients.
• Psychological distress, medication beliefs, and knowledge are significant correlates of AZA nonadherence.