Assessment of Medication Adherence in Patients: Development and Validation of a Machine Learning Model

This study addresses limitations of traditional medication adherence assessment tools by developing a machine learning model to evaluate post-discharge medication compliance in patients using drugs. The research was conducted at Nanjing Drum Tower Hospital from February 2024 to December 2024. We collected clinical data from 240 patients through questionnaires and developed a multi-class machine learning model. Feature selection employed manual screening and polynomial logistic regression. Six ML models were evaluated, with the Random Forest Classifier (RFC) demonstrating optimal performance (bad_AUC = 0.979, fine_AUC = 0.973, good_AUC = 0.917). SHAP analysis was used to explain the best-performing model. The RFC model showed superior predictive capability across all adherence levels. Model interpretation revealed key clinical factors influencing adherence patterns. The tool enables early identification of non-compliance and supports intervention strategies. This RFC-based model represents a significant advancement in medication adherence assessment, offering clinicians a practical tool for monitoring compliance. The approach shows particular promise for enhancing mental health management in this patient population by fostering better medication awareness and establishing scientific medication habits during early treatment stages.