Deep Neural Networks (DNNs) may be very effective for the classification over highly-sized data sets, especially in the medical domain, where the recognition of the occurrence of a specific event related to a disease is of high importance. Unfortunately, DNNs suffer from the drawback that a good set of values for their configuration hyper-parameters must be found. Currently, this is done through the use of either trial-and-error methods or sampling-based ones. In this paper we propose a new approach to find the most suitable structure for a DNN used for a classification problem in terms of achievement of the highest classification accuracy. This approach is based on a distributed version of Differential Evolution (DE), a variety of an Evolutionary Algorithm. To evaluate the approach, in this paper we investigate this issue with reference to Obstructive Sleep Apnea (OSA). OSA is an important medical problem consisting of episodes taking place during night in which a subject stops breathing due to a constriction of the upper airways. This deteriorates the quality of life and may have dangerous, and even lethal, consequences on both short and long term. An accurate classification is a very crucial step for the OSA treatment, because understanding automatically that a subject is experiencing such an episode may be decisive if prompt medical action is needed. In our experiments, classification takes place on a data set in which each item contains the values of 17 Heart Rate Variability parameters, extracted from ElectroCardiography signals, and the annotation of OSA events. We have extracted this data set from the real-world Sleep Heart Health Study database. The results obtained by the distributed DE are compared against those of the Grid Search as well as against those achieved by 13 well-known classification tools. The use of a distributed DE version turns out to be very effective in automatically obtaining DNN structures with higher classification accuracy with respect to Grid Search (72.95% versus 72.61%), and allows saving a high amount of time (three hours as opposed to 65 h and 40 min). Moreover, the proposed method outperforms in terms of higher accuracy all the other classifiers investigated, as it is evidenced also by statistical analysis. Numerically, the runner-up, i.e., JRip, achieves as its best value 72.01% and 71.50% on average over 25 runs, both values being lower than 72.95% and 72.74% obtained by our dDE. (C) 2019 Elsevier B.V. All rights reserved.
