Stress detection from blood volume pulse data using customized convolutional networks

Stress detection is a popular research direction due to its important implications for personal, occupational, and social health. A number of current approaches use features computed from multiple sensory modalities (such as electrocardiogram, galvanic skin response, skin temperature, respiration, accelerometer data, and others). Both classic machine learning algorithms (decision trees, discriminant analysis, support vector machines, etc.) and neural networks (fully connected, convolutional, recurrent) are widely used as methods for solving this problem. The use of classic methods, as well as fully connected neural networks, requires large amounts of data to extract features. Various studies examine subject-independent and subject-dependent (initially personal or adapted) models. This work is aimed at developing and implementing a method based on the principle of customization of convolutional neural networks for stress detection based on heart rate variability data. The proposed method is a convolutional neural network. The work explores modifications using various dimensionality reduction layers, such as a one-dimensional convolutional layer, maximizing and averaging pooling. The impact of using the numerical derivative of heart rate variability as additional information in the input data is also explored. This work demonstrates the importance of customized models, if this opportunity is available, due to their increased accuracy and reliability. The proposed method, based on 60 intervals between heartbeats, binary determines whether a person experiences stress. Prior to customization, detection accuracy is 0.853, f₁-measure value is 0.901. The accuracy of stress detection after customization reaches 0.942 ± 0.095.

Authors: M. O. Dobrokhvalov, A. Yu. Filatov

Direction: Informatics, Computer Technologies And Control

Keywords: stress detection, convolutional neural network, machine learning, heart rate variability, subject-dependent models

---

View full article