| Summary: | Medication nonadherence is a significant public health concern that leads to ineffective treatment, which in turn engenders complications such as increased morbidity risks, unnecessary hospitalisations, and premature mortality. Technologies of the Fourth Industrial Revolution, such as machine learning, provide breakthroughs in identifying the most important features for building predictive models of medication adherence. Due to the diversity and complexity of medication adherence, it is crucial to leverage machine learning to determine significant medication adherence factors. This study systematically reviewed articles exhibiting feature selection and feature importance in research utilising machine learning to analyse medication adherence among Non-communicable diseases (NCD) patients. The articles were retrieved using Google Scholar, Research4Life, IEEE Xplore, and PubMed. The requirements for inclusion were met by 27 papers published between 2010 and 2022. The publications reviewed incorporate machine learning while also demonstrating feature selection and the importance of predictors of medication adherence in NCD patients with hypertension (n = 6), cardiovascular diseases (n = 6), diabetes (n = 4), opioid use disorder (n = 3), and other NCDs (n = 8). The findings demonstrate that medication adherence is a multifactorial issue influenced by various features such as sociodemographic and economic characteristics, medication information, behavioural, disease-related, and healthcare system-related factors. Some of these features, such as the patient's age, gender and race, cannot be modified. Once the patients with nonmodifiable features have been identified, they must be proactively monitored for medication adherence. On the other hand, adjustable risk features, such as self-efficacy and medication knowledge, can be modified and improved through medication education or medication adherence awareness. Various techniques for selecting and ranking features have emerged. These include filter-based feature selection, mutual information measures, and wrapper-based methods. In short, feature selection involves either feature weighting, feature ranking, or the creation of a subset of the entire candidate feature set based on a subset evaluation process, as in wrapper-based methods, which entail the selection of a feature subset with the highest predictive power. An in-depth understanding of feature selection approaches results in more effective models and a deeper understanding of the underlying data structure and features. The study concluded that machine learning-based feature selection and feature importance ranking techniques are more effective alternatives to conventional statistical and non-statistical methods for identifying significant features in predicting medication adherence in NCD patients.
|
|---|