Knowledge extraction based on wavelets and DNN for classification of physiological signals: Arousals case

Edwar Macias Toro, Antoni Morell, Javier Serrano, Jose Lopez Vicario
Wireless Information Networking (WIN) Group, Telecommunications and Systems Engineering Dept, UAB


Aim: The objective of this work was to use deep neural networks (DNN) as a mechanism for automatic extraction of knowledge and classification for arousal regions in physiological signals (PS).

Methods: Each record has 13 PS, with more than 5 million of samples, with each sample labeled in three possible classes, two of them are used to classify regions of the PS in arousal and no-arousal, the another one is no used. Initially, the PS and their labels are segmented with a window of constant length, N. Then, all the segments of the PS are concatenated to generate a new one with length 13*N. In addition, the segment of labels is summarized in a global one, defined by the majority of labels found in a window. Then, a DNN is trained, which is responsible for learning the best possible non-linear function that matches the inputs to the correct classification labels, in an iterative process to have the minimum possible error between the input and the desired output. Finally, to evaluate its performance, it is used on data that has not been used to train.

Results: Taking a window of length N = 1000, a total of 80 of 994 availa-ble records are segmented, obtaining a total of 53311 segments of length 13000. Then a DNN, with two hidden layers with 150 and 45 neurons and 2 for the classification output layer, is trained. Finally, this model was tested with ROC AUC metrics, obtaining a AUROC of 0.67+-0.03 and AUPRC of 0.19+-0.12 in the training data.

Conclusion: Although the model presents a considerable number of false positives, due the unbalanced classes, without knowing a priori features and with less than 10% of the training samples, the DNN is able to extract enough knowledge to carry out a correct classification.