Data-driven versus a domain-led approach to k-means clustering on an open heart failure dataset