KHUNG MÔ HÌNH HỌC MÁY DỰ ĐOÁN TUÂN THỦ ĐIỀU TRỊ TRONG BỆNH PARKINSON: CÂN BẰNG GIỮA HIỆU SUẤT DỰ ĐOÁN VÀ KHẢ NĂNG DIỄN GIẢI LÂM SÀNG

Abstract

Background: Medication non-adherence among patients with chronic diseases such as Parkinson’s disease (PD) is a major challenge to effective treatment management. Applying Machine Learning (ML) models is a promising approach, but their application in clinical settings is often limited by small sample sizes, imbalanced clinical data, and especially their "black-box" nature.

Objective: This study aims to develop a robust, transparent ML framework to predict treatment adherence in PD patients, providing actionable clinical explanations, especially on a small dataset.

Methods: We analyzed a clinical dataset of 100 patients, which contains demographic, socioeconomic, and clinical features. To evaluate ML models with this small dataset, we utilized a Leave-One-Out Cross Validation (LOOCV) strategy. Additionally, the class imbalance (72% adherent vs. 28% non-adherent) problem was addressed using cost-sensitive learning algorithms. Feature engineering was also applied to resolve multicollinearity. We evaluated performance of Logistic Regression, Random Forest, and Support Vector Machine (SVM) models. Furthermore, Explainable AI (xAI) techniques were used to investigate feature contributions.

Results: The balanced Support Vector Machine (SVM) model achieved the highest overall performance of 85.0% accuracy and a F1-score of 0.746. Feature analysis indicated that medication regimen complexity, age, and disease duration are main indicators in predicting adherence behaviors.

Conclusion: Our proposed framework demonstrates that ML, when coupled with xAI, can reliably forecast treatment adherence even with limited clinical data. This transparent approach provides a foundation for developing robust Clinical Decision Support Systems (CDSS), and helps healthcare professionals to identify high-risk patients early and tailor personalized intervention strategies.