Evaluating clustering methods for psychiatric diagnosis

This report represents an empirical evaluation of the major clustering approaches on psychiatric diagnostic data. Experienced psychiatrists, using 17 psychopathological variables, developed 88 archetypal psychiatric patients to represent four diagnostic categories (manic-depressive depressed, manic-depressive manic, simple schizophrenic, and paranoid schizophrenic). Ten computerized methods representative of the major clustering approaches and using various measures of similarity between patients were applied to this data set to develop de novo patient groupings. Evaluative criteria included the concordance of clustering output to the structure of the original data, and clustering replicability. Considerable differences were obtained among clustering methods. The best-ranked procedures were nearest centroid sorting methods and complete and centroid linkage hierarchical methods. The overall poorest ranking were obtained for multivariate normal mixture analysis and facial representation of multidimensional points. Further evaluation of cluster analytic methods on real biological and psychosocial data sets yielded similar rankings.