Two-Stage Approach to Cluster Categorical Medical Data

Medical datasets, apart from interval data, usually contain variables of categorical kind, either nominal or ordinal, such as symptoms and stage of disease, diagnosis, medications used, gender, etc. The attempts to cluster such data brings specific challenges that are not feasible with sets consisting of interval variables only, because assumptions behind most clustering algorithms, such as Gaussianity of noise, are violated, and the idea of a prototype vector (centroid) to represent a cluster becomes unclear. A reliable clustering technique could be extremely useful e.g. in diagnostics of cardiological illnesses. Here, we proposed overcoming the issues mentioned above using a two-stage clustering procedure. The first stage was to project multidimensional categorical vectors into a 2D continuous space. Then, in the second stage, clustering was performed on continuous-value vectors. The additional advantage of this procedure was visual inspection of the data distribution, facilitating the selection of clusters.