ECG based stress detection using covolutional neural network

Abstract

Stress refers to a man’s ability to respond to any external or internal threat or pressure and has a direct or indirect impact on one’s health. Hence, there is a growing need to detect stress at an early stage as one might not know whether he/she is stressed or not, and stress left undetected for a long time can become chronic and life endangering. Many health-related disorders are linked to stress and thus monitoring, measuring, and managing stress is simply a lifesaving remedy. There are so many physiological methods to measure stress, however, the issue with most of those physiological methods is the complexity in measuring the signals and the methods are not convenient for day-to-day use. This study implements the use of Electrocardiograph (ECG) signals which are recorded using a RAQIB smartwatch and does not need any second person intervention. Features extracted from the ECG signals have been the key to detect stress for many years, however, recent advancements in neural networks have prompted us to apply the convolutional models as well apart from the traditional machine learning approaches. Researchers are rapidly moving towards the neural networks approach due to the automatic ability of these neural networks to learn features and due to higher accuracy classification models of Convolutional Neural Networks (CNNs). Based on the advantages of CNNs over the traditional machine learning approaches, this study based on the two-dimensional CNN model is proposed for the detection and classification of ECG signals into two distinct classes: namely, stress and no-stress. It is the first such study for stress detection where the one- dimensional ECG signals are converted into 2-D scalogram images by virtue of a Continuous Wavelet Transform (CWT). The proposed CNN model consists of an input layer to feed a scalogram image followed by 4 back-to-back layers of convolution, rectified linear unit (RELU) and max pooling. The accuracy of the proposed stress detection model is compared with both the handcrafted features approach and other relevant models. The average accuracy of 99% is certainly better than the traditional approaches for detecting stress and the usage of the smartwatch makes the model more robust and easier to use while performing day to day activities. The model can be easily and conveniently used at workplaces and offices to determine stress without any assistance from a second person and the method is quick.