Preliminary study: Data analytics for predicting medication adherence in Malaysian arthritis patients.

Objective: In multi-ethnic Malaysian populations, understanding and improving medication adherence in arthritis patients is crucial for enhancing treatment outcomes. Non-adherence, whether intentional or due to complex factors, can lead to severe long-term consequences such as increased disability and disease progression. This study analysed and predicted Malaysian arthritis medication adherence using 13 machine learning models.

Methods: A majority of 151 responders (82.1%) were female and 58.3% had comorbid illnesses. Notably, 90.07% of respondents were non-adherence to their prescription, with significant differences by occupation and aids in medication. This study's machine learning models perform better with recursive feature elimination for feature selection. Key variables included occupation, presence of other diseases, religion, income, medication aid, marital status, and number of medications taken per day. These variables were used to build predictive models for medication adherence.

Results: Results from machine learning algorithms showed varied performance. Support vector machine, gradient boosting, and random forest models performed best with AUC values of 0.907, 0.775, and 0.632 utilizing all variables. When using selected variables, random forest (AUC = 0.883), gradient boosting (AUC = 0.872), and Bagging (AUC = 0.860) performed best. Model interpretation using SHapley Additive exPlanations analysis identified occupation as the most important variable affecting medication adherence. The study also found that unemployment, concomitant disease, income, medication aid type, marital status, and daily medication count are connected with non-adherence.

Conclusion: The findings underscore the multifaceted nature of medication adherence in arthritis, highlighting the need for personalized approaches to improve adherence rates.