Using cluster ensemble and validation to identify subtypes of pervasive developmental disorders

Pervasive Developmental Disorders (PDD) are neurodevelopmental disorders characterized by impairments in social interaction, communication and behavior. Given the diversity and varying severity of PDD, diagnostic tools attempt to identify homogeneous subtypes within PDD. Identifying subtypes can lead to targeted etiology studies and to effective type-specific intervention. Cluster analysis can suggest coherent subsets in data; however, different methods and assumptions lead to different results. Several previous studies applied clustering to PDD data, varying in number and characteristics of the produced subtypes. Most studies used a relatively small dataset (fewer than 150 subjects), and all applied only a single clustering method. Here we study a relatively large dataset (358 PDD patients), using an ensemble of three clustering methods. The results are evaluated using several validation methods, and consolidated through an integration step. Four clusters are identified, analyzed and compared to subtypes previously defined by the widely used diagnostic tool DSM-IV.