Impact of Dimensionality on Medication Adherence Prediction: Assessing Accuracy and Computational Efficiency

Non-adherence to medication represents a significant concern in managing chronic conditions, resulting in poor health outcomes and increased mortality. Artificial Intelligence (AI) and Machine Learning (ML) models offer promising solutions by analyzing large datasets to improve the understanding and prediction of non-adherence. However, the dimensionality of data and associated computational costs, including environmental impact, remain a challenge in the application of AI systems. This study explores the influence of data dimensionality on the evaluation metrics and computational costs of ML models in predicting medication adherence. Utilizing adherence behaviour data collected in the BEAMER project, dimensionality reduction and Logistic Regression (LR) methods are performed. The results provide a comprehensive understanding of the models' performance after the application of Recursive Feature Elimination with Cross Validation (RFECV) and Principal Component Analysis (PCA), while also emphasizing the importance of evaluating the carbon footprint in ML models. The analysis indicates that pipelines including all variables and those incorporating RFECV present better evaluation metrics compared to those utilizing PCA, although this approach requires the least computational resources. These findings highlight the importance of tailoring project designs to their specific use cases and requirements.Clinical Relevance- This study highlights the importance of balancing accuracy, computational efficiency, interpretability and environmental impact in AI-based adherence to medication models, aiding clinicians in selecting optimal approaches for healthcare applications.