Identifying and Monitoring Risk Factors in Chronic Disease for Personalized eHealth

Non-communicable diseases (NCDs) have surged to the forefront of global health concerns, necessitating a profound understanding of their intricate nature. This study introduces an innovative approach that leverages Causal Bayesian Networks (CBNs) to untangle the intricate web of causal relationships among variables linked to NCDs, with a specific emphasis on diabetes. NCDs encompass a broad spectrum of health conditions, including cardiovascular diseases, diabetes, cancer, and respiratory disorders. These conditions are characterized by their chronic, non-communicable nature, deeply entrenched in lifestyle choices, genetics, and environmental influences. Diabetes, in particular, assumes a central role in this investigation.

CBNs offer vital insights to inform evidence-based health policies and resource allocation for policymakers. It is, however, crucial to acknowledge the inherent challenges in constructing CBNs, which require substantial data resources and expertise in statistical modeling. The validation of identified causal relationships through rigorous experimental research remains an integral component of this methodology.

This study introduces a framework to address the challenges posed by the proliferation of variables in CBN structure learning. As the number of variables increases, the complexity of assessing possible graph structures grows exponentially. To mitigate this, we propose a feature selection approach aimed at reducing the number of variables. Our research combines feature selection methods with ensemble classification techniques to achieve highly accurate early-risk prediction. We initially filter high-dimensional data using filter methods and then employ wrapper methods in conjunction with ensemble classifiers. Wrapper methods involve two critical steps: searching for the optimal feature subset and evaluating the selected subset. While various search methods have been proposed, our research advocates the use of meta-heuristic algorithms or hybrids of existing meta-heuristic algorithms for this purpose.