MoStress: um modelo de aprendizado profundo para detecção de estresse a partir de sinais fisiológicos

Abstract

The COVID-19 pandemic showed how the preparation and fight against those diseases plays a crucial role on the modern society. However, the coronavirus is not the only pandemic diseases which afflicts the globe: mental illnesses also afflict a large number of the world population. Nowadays, stress, anxiety and depression are classified as mental illness and proximally 10.7% of the world population suffers with one of those diseases, therefore, mental illness might have a high pandemic potential and should be treated with the necessary urgency. One approach to deal with mental illness is to use machine learning algorithm which uses time series as input to detect those diseases. Considering the huge variety of physiologic measured by modern sensors, such as temperature, heart rate and others, and also considering the increase popularity of those sensors in our society, the use of those signals to monitor all kind of the diseases gains more relevance. In that sense, dealing with time series which represents physiologic signs with modern machine learning technics, may result in a substantial improvement of life quality of the population, because with those algorithms, several diseases might be classified quickly and more efficient, making more ease the health care professional diagnoses and avoiding the diseases to reach an worst scenario. This work introduce the MoStress, a deep learning model which get as input time series which represents physiologic signs and make stress classification. The MoStres is made by a pre-processing step, which consists in using Fourier Transform to clean noise, Rolling Z-Score to normalize the data, windowing by class frequency to window classification and weight calculation to deal with unbalance data. Besides that, the MoStress also have a deep neural network which make the classification using the pre-processed data, where this neural network consists on one of the following models: a recurrent neural network, a Echo State Network or a combination of the NBeats and a Multi Layer Perceptron network. The MoStress used public physiologic data collected by the Siegen University, in Germany (the dataset is named WESAD), where this dataset is constituted also by 3 different classes: baseline, stress amusement. Considering this, the MoStress using physiologic signals of respiration, temperature, electrocardiogram, electromyogram and electrodermal activity, collected via chest sensor after pre-process these data and using a recurrent neural network, achieved accuracy of 96.5% on the 3 class classification problem and also achieved recall, f1-score and precision of 96%, 93% and 94%, respectively, for the stress class, showing the good performance on classification problem with pre-processed data and a recurrent neural network.