Characterizing Cardiovascular Risk Through Unsupervised and Interpretable Techniques

Abstract

In the last decades, the prevalence of noncommunicable diseases has significantly increased, causing millions of deaths worldwide. Among them, cardiovascular diseases (CVDs) have become a global health problem due to the high rates of mortality and morbidity. This issue is exacerbated in patients with type 1 diabetes (T1D) since they have a higher risk of developing CVD and death. Machine Learning (ML) methods have reached remarkable results in both industry and academia. In the clinical setting, these methods are promising to extract clinical knowledge and help to prevent acute events caused by CVDs. This paper aims to identify clusters of patients with different CVD risks and interpret the clinical variables that play a significant role in CVD risk. Towards that end, unsupervised and interpretable ML techniques were considered, in particular k-prototypes and the agglomerative hierarchical clustering technique with Gower’s distance. We used a dataset that collects information from 677 adult Danish patients with T1D. Experimental results showed that albuminuria, smoking, and gender are crucial for identifying T1D patients at different risks of suffering CVD. Our paper contributes to the identification of CVD risk factors for T1D patients, paving the way for clinical decision-making. This identification of individuals at higher CVD risk would allow physicians to make early interventions and adequate treatments to prevent the onset of diseases.