Machine Learning in Pharmaceutical Research: Data Clustering, Why so and How so

Background: In clinical data subgroups can sometimes be identified using regression analysis of subgroup characteristics against some outcome variable, but in data samples without an available outcome variable cluster analysis is a suitable alternative. It is based on the concept that patients with closely related characteristics may also be more related in other fields like prognoses and treatment efficacies. Objective: To compare the performance of three different cluster methodologies, hierarchical , k-means, and density-based clustering. Methods: A simulated data example of fifty patients with mental depression was used. Results: Each cluster methodology identified three clusters. However, the cluster patterns were very different. The hierarchical method showed round patterns different in size, the k-means method round patterns equal in size, and the density-based method non-circular patterns also different in size. The patterns from the hierarchical method were better in agreement with the patterns as clinically expected, than those from the other methods. Conclusions: 1. Cluster analysis is little used in clinical research. 2. Hierarchical cluster is adequate if subgroups in the data are expected to be different in size but, otherwise, Gaussian-like. It is available in the module Classify of SPSS. 3. K-means cluster analysis is adequate if subgroups are expected to be approximately similar in size. It is also available in the module Classify of SPSS. 4. Density-based cluster analysis is adequate if small outlier groups between an, otherwise, homogeneous population is expected. It is not available in SPSS, but an interactive JAVA Applet is freely obtainable at the Internet.