Authors:
Antonio Lopez-Martinez-Carrasco
1
;
Jose Juarez
1
;
Manuel Campos
2
;
1
and
Bernardo Canovas-Segura
1
Affiliations:
1
Med AI Lab, University of Murcia, Spain
;
2
Murcian Bio-Health Institute (IMIB-Arrixaca), Spain
Keyword(s):
Methodology, Patient Phenotyping, Subgroup Discovery, Reduced Subgroup Set.
Abstract:
Subgroup Discovery (SD) is a supervised machine learning technique that mines a set of easily readable features of patients with a medical condition in the form of a subgroup set (called patient phenotype). However, using only the output obtained by a single execution of an SD algorithm hinders the discovery of the best phenotypes since it is difficult for clinicians to choose the most suitable algorithm, its best hyperparameters and the quality measure. Therefore, we propose a new phenotyping approach based on SD that evaluates the outcomes of different SD algorithms to obtain a final patient phenotype with a reduced dependency on the initial conditions of these executions and to ensure diversity in terms of coverage of the subgroups from this phenotype. For that, we first define the problem of mining a patient phenotype in the form of a reduced subgroup set and, after that, we propose a new 6-step methodology to tackle this problem. Moreover, we carry out experiments driven by this
methodology and focused on the antibiotic resistance problem by using the MIMIC-III public database and the patients infected by an Enteroccous Sp. bacterium resistant to Vancomycin as a target. Finally, we obtain a phenotype formed of 7 subgroups.
(More)