Authors:
María Jurado-Camino
1
;
David Chushig-Muzo
1
;
Cristina Soguero-Ruiz
1
;
Pablo Bohoyo
2
and
Inmaculada Mora-Jiménez
1
Affiliations:
1
Dep. Signal Theory and Communications, Rey Juan Carlos University, Camino del Molino 5, Madrid, Spain
;
2
University Hospital of Fuenlabrada, Madrid, Spain
Keyword(s):
Data Augmentation, Imbalance Learning, Decision Trees, Clinical Codes, Chronic Diseases.
Abstract:
Chronic diseases (CD) are the leading cause of death worldwide, presenting higher mortality rates and economic burden (both in the health and social context) as the complexity of the CD increases. The use of Electronic Health Records (EHRs) and Machine Learning (ML) contribute to significant progress in health domain research, supporting identifying the patient's health status for early interventions. Despite these achievements, the class imbalance can limit the generalization capability of many ML models and data augmentation techniques are proposed to face this limitation. In this work, a Generative Adversarial Network named medWGAN is used to generate synthetic patients considering clinical data collected from EHRs linked to the University Hospital of Fuenlabrada. Data are associated with patients diagnosed with both simple CD (diabetes, hypertension, congestive heart failure, chronic obstructive pulmonary disease) and multiple CD. Experimental work using decision trees as predi
ctors to determine the patient's health status showed the ability of medWGAN for preserving the underlying (high-dimensional and sparse) clinical patterns. Our results indicate that the identification of patients with multiple CD may benefit from the use of medWGAN as long as the data used for its training is diverse enough, contributing to supporting clinical decision-making in complex scenarios with many features.
(More)