Model selection for clustering of pharmacokinetic responses

• Published in: Computer methods and programs in biomedicine Vol. 162; pp. 11 - 18
• Main Authors: Guerra, Rui P., Carvalho, Alexandra M., Mateus, Paulo
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.08.2018
• Subjects: Clustering; Minimum description length; Model selection; Normalised maximum likelihood; Pharmacokinetics; Minimum description length; Normalised maximum likelihood; Clustering; Pharmacokinetics; Model selection
• ISSN: 0169-2607; 1872-7565; 1872-7565
• DOI: 10.1016/j.cmpb.2018.05.002

•We improve an Expectation-Maximisation method to stratify drug responses of patients. With this we aim to provide different drug doses for each stratum and so, maximise therapy efficacy while minimising its toxicity.•Two novel model selection criteria, based on the Minimum Description Length and the Normalized Maximum Likelihood, were derived and developed for clustering pharmacokinetic (PK) responses.•The method was evaluated over synthetic and real data and showed the ability to unveil the correct number of clusters underlying the mixture of PK curves.•A cost-efficient parallel implementation in Java is publicly and freely available in a GitHub repository, along with a user manual and data used in the experiments. Background and Objective: Pharmacokinetics comprises the study of drug absorption, distribution, metabolism and excretion over time. Clinical pharmacokinetics, focusing on therapeutic management, offers important insights towards personalised medicine through the study of efficacy and toxicity of drug therapies. This study is hampered by subject’s high variability in drug blood concentration, when starting a therapy with the same drug dosage. Clustering of pharmacokinetics responses has been addressed recently as a way to stratify subjects and provide different drug doses for each stratum. This clustering method, however, is not able to automatically determine the correct number of clusters, using an user-defined parameter for collapsing clusters that are closer than a given heuristic threshold. We aim to use information-theoretical approaches to address parameter-free model selection. Methods: We propose two model selection criteria for clustering pharmacokinetics responses, founded on the Minimum Description Length and on the Normalised Maximum Likelihood. Results: Experimental results show the ability of model selection schemes to unveil the correct number of clusters underlying the mixture of pharmacokinetics responses. Conclusions: In this work we were able to devise two model selection criteria to determine the number of clusters in a mixture of pharmacokinetics curves, advancing over previous works. A cost-efficient parallel implementation in Java of the proposed method is publicly available for the community.