Classification of urinary stones by cluster analysis of ionic composition data

The cluster analysis technique is considered for classifying kidney stones based on data for nine chemical analysis parameters. A set of 214 stones is used, which has been previously classified using empirical classification rules into three stone types using the percentage concentrations of the urate, oxalate, and phosphate radicals. We investigate whether cluster analysis utilising data on all parameters leads to different classifications and explore the possibility of other effective classifiers. We also compare the performance of various clustering techniques, distance and similarity measures and data standardisation methods. Results indicate that inclusion of the additional six parameters does not improve the classification accuracy. Best matching with the empirical classification (6% error) is achieved using the average linkage (between groups) clustering method and the squared Eculidean distance measure without data standardisation. Excluding these three main radicals causes a 63% matching error. Cluster analysis results suggest that carbon ions alone provide a single classifier for the three stone types, giving a matching error of ≃10% with the empirical classification.