Personalization of convolutional neural networks within the stress detection task using heart rate variability data Текст научной статьи по специальности «Медицинские технологии»

ИНФОРМАЦИОННЫХ ТЕХНОЛОГИЙ, МЕХАНИКИ И ОПТИКИ

ИСКУССТВЕННЫЙ ИНТЕЛЛЕКТ И КОГНИТИВНЫЕ ИНФОРМАЦИОННЫЕ ТЕХНОЛОГИИ ARTIFICIAL INTELLIGENCE AND COGNITIVE INFORMATION TECHNOLOGIES

doi: 10.17586/2226-1494-2023-23-6-1178-1186

Personalization of convolutional neural networks within the stress detection task

using heart rate variability data Maksim O. Dobrokhvalov1®, Anton Yu. Filatov2

Saint Petersburg Electrotechnical University "LETI", Saint Petersburg, 197022, Russian Federation

1 [email protected]®, https://orcid.org/0000-0002-0571-5836

2 [email protected], https://orcid.org/0000-0003-4298-8523

Abstract

Stress detection is an active area of research with important implications for personal, occupational, and social health. Most modern approaches use features computed from multiple sensor modalities, i.e., grouping different types of data from multiple sources for processing. These include electrocardiogram, electrodermal activity, electromyogram, skin temperature, respiration, accelerometer data, etc. Also, traditional machine learning algorithms (decision tree, discriminant analysis, support vector machine, etc.) or fully-connected neural networks are mostly used. Using these methods requires large amounts of data. Researchers are considering different approaches to personalization or generalization of models relative to subjects, namely subject-independent and subject-dependent (initially personal or adapted) models. The aim of the presented work is to develop a method for detecting stress based on heart rate variability data, taking into account the process of personalization of neural networks. The use of a convolutional neural network is proposed. The dependence of accuracy on the length of the input signal is studied. The dependence of accuracy on the data dimensionality reduction layer (one-dimensional convolutional layer, maximizing and averaging pooling) used in the network is also considered. The importance of personalizing models is demonstrated to significantly increase the accuracy of models of specific subjects. It is shown that the proposed method, based on 60 intervals between heartbeats, makes it possible to binary determine whether a person is under stress. Personalization allowed increasing the accuracy from 91.8 % to 98.9 ± 2.6 %. The F1-score value increased from 0.907 to 0.983 ± 0.038. The proposed personalized networks can be used in systems for monitoring the functional state of a person. They can also be used as part of a system that grants or restricts access to private resources based on whether a person is currently at rest. Keywords

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

The article was prepared within the project "Methods of hybrid intelligence for building heterogeneous multi-agent systems with self-learning and self-organization" of the development program of St. Petersburg Electrotechnical University "LETI".

For citation: Dobrokhvalov M.O., Filatov A.Yu. Personalization of convolutional neural networks within the stress detection task using heart rate variability data. Scientific and Technical Journal of Information Technologies, Mechanics and Optics, 2023, vol. 23, no. 6, pp. 1178-1186. doi: 10.17586/2226-1494-2023-23-6-1178-1186

l/ITMO

НАУЧНО-ТЕХНИЧЕСКИМ ВЕСТНИК ИНФОРМАЦИОННЫХ ТЕХНОЛОГИИ, МЕХАНИКИ И ОПТИКИ ноябрь-декабрь 2023 Том 23 № 6 http://ntv.ifmo.ru/

SCIENTIFIC AND TECHNICAL JOURNAL OF INFORMATION TECHNOLOGIES, MECHANICS AND OPTICS November-December 2023 Vol. 23 No 6 http://ntv.ifmo.ru/en/

ISSN 2226-1494 (print) ISSN 2500-0373 (online)

© Dobrokhvalov M.O., Filatov A.Yu., 2023

УДК 004.032.26

Персонализация сверточных нейронных сетей в задаче обнаружения стресса с использованием данных вариабельности сердечного ритма

Максим Олегович Доброхвалов1Н, Антон Юрьевич Филатов2

!>2 Санкт-Петербургский государственный электротехнический университет «ЛЭТИ» имени В.И. Ульянова (Ленина), Санкт-Петербург, 197022, Российская Федерация

1 mght1337bot@gmaiLcomи, https://orcid.org/0000-0002-0571-5836

2 [email protected], https://orcid.org/0000-0003-4298-8523

Аннотация

Введение. Обнаружение стресса является активной областью исследований с важными последствиями для личного, профессионального и социального здоровья человека. Большинство современных подходов используют признаки, вычисленные на основе нескольких сенсорных модальностей, т. е. группируют для обработки различные типы данных, полученные из нескольких источников. К ним относятся электрокардиограмма, кожно-гальваническая реакция, электромиограмма, температура кожи, дыхание, данные акселерометров и др. При этом чаще используются традиционные алгоритмы машинного обучения, такие как решающие деревья, дискриминантный анализ, метод опорных векторов и другие, а также полносвязные нейронные сети. Использование этих методов требует больших объемов данных. Исследователи рассматривают отличающиеся подходы к персонализации или общности моделей относительно субъектов, а именно субъекто-независимые и субъекто-зависимые (изначально персональные или адаптированные) модели. Целью представленной работы является разработка метода детектирования стресса на основе данных вариабельности сердечного ритма с учетом процесса персонализации нейронных сетей. Метод. Для решения поставленной задачи предложено применение сверточной нейронной сети. Исследована зависимость точности детектирования от длины входного сигнала. Рассмотрена зависимость точности от используемого в сети слоя уменьшения размерности данных (одномерный сверточный слой, максимизирующий и усредняющий пуллинги). Продемонстрирована важность персонализации моделей, для значительного увеличения точности детектирования для конкретных субъектов. Основные результаты. Показано, что предлагаемый метод на основании 60 интервалов между ударами сердца позволяет бинарно определить, находится ли человек в состоянии стресса. Персонализация сверточных нейронных сетей позволила повысить точность с 91,8 до 98,9 ± 2,6 %. Значение F1-меры повысилось с 0,907 до 0,983 ± 0,038. Обсуждение. Предложенные персонализированные сети могут применяться в системах мониторинга функционального состояния человека. Также могут быть использованы как часть системы, предоставляющей или ограничивающей доступ к приватным ресурсам на основании того, находится ли человек в состоянии покоя в данный момент. Ключевые слова

детектирование стресса, сверточные нейронные сети, машинное обучение, вариабельность сердечного ритма,

субъекто-зависимые модели

Благодарности

Работа подготовлена в рамках проекта «Методы гибридного интеллекта для построения гетерогенных многоагентных систем с самообучением и самоорганизацией» программы развития СПбГЭТУ «ЛЭТИ». Ссылка для цитирования: Доброхвалов М.О., Филатов А.Ю. Персонализация сверточных нейронных сетей в задаче обнаружения стресса с использованием данных вариабельности сердечного ритма // Научно-технический вестник информационных технологий, механики и оптики. 2023. Т. 23, № 6. С. 1178-1186 (на англ. яз.). doi: 10.17586/2226-1494-2023-23-6-1178-1186

Introduction

Stress is the body's response to perceived physical or psychological threats [1] and it is defined as the transition from a calm state to an excited state, triggering a set of physiological responses [2]. Moreover, stress detection is important for many health problems, such as depression, anxiety, heart attacks and strokes [3]. Stress also affects a person's decision-making ability, attention span, learning and problem-solving ability [4]. Therefore, stress detection is an important task.

Various classical machine learning methods as well as neural networks are used in various studies to solve this problem. Also, various input data for stress detection are used from various data sensors, such as electrocardiogram (ECG), electrodermal activity (EDA), etc. This study proposes to use a convolutional neural network which receives a set of RR intervals (Heart Rate Variability (HRV)) as input data. In [5], the authors successfully applied personalization to

EDA data. Based on that research, this paper examines the process of personalization of convolutional neural networks with HRV input data. Thus, the aim of the work is to develop a method for stress detection based on HRV data, taking into account the process of personalization of neural networks, as well as the implementation of this method. The proposed approach is competitive with other modern methods. The code used in the work is available1.

Related works

Summary of review related works is presented in Table 1. Wearable Stress and Affect Detection (WESAD) is a commonly used dataset in related works (14 of 20) but also some studies use their own data. Among the works reviewed, ECG [2, 3, 6-13], EDA [2-6, 9, 11, 14-19] and

1 Available at: https://github.com/Nightbot1448/human

stress_detection (accessed: 10.01.2023).

Table 1. Summary of reviewed related works

Paper Year Dataset Subjects Data Model Accuracy, % Window, s

[2] 2018 WESAD 15 ECG, EDA, BVP, Temp, Resp, EMG, ACC kNN, DT, RF, LDA, AB 92.83 0.25, 5, 60

[3] 2020 WESAD 15 ECG, EDA, BVP, Temp, Resp, EMG, ACC kNN, SVM, AB, FCN 95.21 1

[4] 2020 WESAD 15 EDA kNN, SVM, RF 91.6 —

[5] 2021 WESAD 15 EDA CNN 92.85 60

[6] 2022 WESAD 15 ECG, EDA, BVP, Temp, Resp, EMG, ACC CNN using GAF 94.8 —

[7] 2016 Other 42 ECG C4.5 tree 79 180

[8] 2021 Other 20 ECG CNN 83.5 10

[9] 2021 WESAD 15 ECG, EDA, BVP, Temp, Resp, EMG, ACC CNN 97.75 ± 2.55 60

[9] 2021 WESAD 15 ECG CNN 91.75 ± 9.73 60

[10] 2019 AffectiveROAD, Other 9, 17 ECG FCN 90.19 10, 60

[11] 2021 WESAD 15 ECG, EDA, BVP, Temp, Resp, EMG, ACC LR 85.71 60

[12] 2021 Other 27 ECG kNN, SVM, FCN, RF, GB 83 30

[13] 2019 Other 20 ECG CNN 82.7 10

[14] 2020 Other 20, 3 HR, EDA CNN 82.5, 93.8 —

[15] 2021 WESAD 15 EDA sTree 95.8 4

[16] 2018 Other 58 HR, EDA, Resp FCN 89.7 90

[17] 2020 WESAD 15 EDA, BVP, ACC, Temp RF, DT, LR 96.68 ± 3.2 0.25

[18] 2020 Other 41 BVP kNN, LDA, FCN 82 60

[19] 2021 WESAD 15 EDA kNN, SVM, FCN, RF 87.5 60

[20] 2022 WESAD 15 BVP FCN 99.04 300

[21] 2019 WESAD 15 Temp, BVP, HR LDA, QDA, RF 87.4 ± 10.4 15, 30, 60, 90, 120

[2, 3]. K-nearest neighbor's classifier (kNN) was used in [2-4, 12, 18, 19]. Tree-based classifiers (like Decision Tree (DT)) were utilized in [2, 3, 7 (C4.5 tree [22]), 15, 17]. Also Logistic Regression (LR) was used in [11, 17]. Indikawati and Winiarti [17] are the only ones who directly used the signal without feature extractions with the classical machine learning methods. Work [23] used convolutional and long short-term memory [24] neural networks for encoding signal with sequent passing to clustering algorithms.

Materials and Methods

Data and preprocessing. Many studies conducted in the field of stress detection use data collected by researchers independently. This study uses the WESAD dataset [2], which has also been used in many studies in recent years [3, 5, 8, 14, 17, 20, 21]. It is a public dataset containing ECG. RR intervals were calculated from the ECG using the heartpy python library1. Data with stress

1 Available at: https://python-heart-rate-analysis-toolkit. readthedocs.io/en/latest/ (accessed: 10.01.2023).

Blood Volume Pulse (BVP) [2, 3, 6, 9, 11, 17, 18, 20, 21] were most often used as data sources. Also other sources, such as respiration info (Resp) [2, 3, 6, 9, 11, 16], skin temperature (Temp) [2, 3, 6, 9, 11, 17, 21], electromyogram (EMG) [2, 3, 6, 9, 11], accelerometers info (ACC) [2, 3, 6, 9, 11] used in some research. It should be noted that in most cases many data sources are used when applying feature engineering [2, 3, 6, 18, 21]. Time and frequency domain of ECG, BVP and EDA are widely used in studies [2-4, 6, 7, 9, 10, 15, 16]. Some studies [2, 4, 5, 7, 8, 10, 12, 18-20] extract features from only one data source. And also there are few studies that use raw data (sometimes with applying filters but without feature extraction) [5, 9, 13, 17].

Accuracy metrics reported ranged between 79 % [7] and 99.04 % [20]. Half of the studies used neural networks. Convolutional Neural Networks (CNN) were used in 6 papers [5, 6, 8, 9, 13, 14], Fully Connected Networks (FCN) also were used in 7 studies [3, 10, 12, 16, 18-20]. Also different studies used machine learning methods.

Random Forest (RF) was used in [2, 4, 17, 21]. Support Vector Machines (SVM) were utilized in experiments [3, 4, 12, 19]. Linear Discriminant analysis (LDA) was used in [2, 18, 21]. AdaBoost classifier (AB) was utilized in

and resting state labels were taken from the dataset. The amusement state was omitted. Next, the RR interval is the interval between neighboring heart beats. The interval is the set of RR intervals used as input data.

Model. The convolutional neural network [14] has shown sufficiently high accuracy. Therefore, a 1D convolutional network architecture was chosen. The network architecture is a sequential use of the ConvX block (Fig. 1, a) and the dimension reduction layer. The ConvX block consists of a one-dimensional convolutional layer with kernel size 3, a batch normalization layer, and a ReLU activation layer. The network architecture for the interval length (input data) equal to 60 is shown in Fig. 1, b. One-dimensional convolution (kernel = 2, stride = 2), max pooling (kernel = 2) and averaging pooling (kernel = 2) were considered as dimensionality reduction blocks. The number of input layers for convolution layers or ConvX blocks is given in parentheses. The first ConvX block parameter (in) means that there may or may not be a layer in the input data containing the difference between consecutive RR intervals (numerical derivative). The architecture depends in part on the maximum interval length. The goal was to form an architecture where after each ConvX block a dimensionality reduction layer could be added (except the first and last). Thus, this architecture made it possible to obtain the required data dimension due to convolutional layers and dimensionality reduction layers (without using fully connected layers). The results of various modifications are presented in the following sections.

Results

This section presents a comparison of different modifications:

— using different interval (input data) lengths,

— choosing layers to reduce dimensionality,

— using numerical derivation.

This section also presents the impact of model personalization for subjects.

Modifications. All modifications of the convolutional neural network proposed in the research process were implemented within this study using the PyTorch framework1. In all experiments, the CrossEntropy loss function was used, the ASGD optimizer (with default parameters) was used, the number of epochs was 50, and the batch size was 8.

The first study was the choice of interval size. Normal resting heart rates range from 60 to 100 bpm [28]. Therefore, the number of RR intervals is not equivalent to the number of seconds, but they can be mapped. 60 seconds is a widely used interval in review studies. It was decided to consider no more than 60 RR intervals with a step of 15. More than 60 RR intervals were not considered due to too long initialization. The slide of intervals was 5 RR intervals. Fig. 2 shows accuracy for different modifications of models depending on input interval length. It can be seen that for the lengths 15, 30, 45 there is a direct dependence of the accuracy. In the case of input interval length equal to

ConvX (in) ConvX (16) Reduce

1 1 1

ConvX (4) Reduce ConvX (8)

1 1 1

Reduce ConvX (32) ConvX (4)

1 1 1

ConvX (8) Reduce ConvlD (2)

1 1 1

Reduce ConvX (16) Softmax

Fig. 1. Neural Network: ConvX block (a); architecture (b)

60, the accuracy for models without numerical derivative is increased. For models with numerical derivation, for the same input interval length, the accuracy decreases regardless of the method of dimensionality reduction. However, the accuracy for the interval length 60 with a max pooling layer is greater (92.16 %) than the accuracy of the other modifications.

As mentioned earlier, one-dimensional convolution layers, max and averaging poolings were considered as dimension reduction methods. Using convolution as a layer for dimension reduction shows the lowest accuracy (Fig. 2). If a numerical derivative was present in the input data, the network with the averaging pooling determined stress with higher accuracy in all cases except when the interval length was 15. If the numerical derivative was not used, then modifications with averaging and max pooling showed greater accuracy depending on the interval length. However the modification with the max pooling showed the highest accuracy (92.16 %) with interval length equal to 60.

The inclusion of an additional layer containing the difference of two consecutive RR intervals to the input data was considered. This value can be treated as a numerical derivative. This difference shows the dynamics of changes in RR intervals, which can be perceived as the rate of change in heart rate. For intervals of length 15, 30, and 45, the accuracy of the networks, whose input was additionally fed by the derivative, is higher than for the corresponding one but without this addition.

In the case of an interval length of 60, the accuracy of modifications without the numerical derivative is slightly higher than that with it. The largest difference between modifications with and without the numerical derivative

92

cx Ï

90

о <1 88

86

/'••j.......

— Conv

■■»■• AvgPool

MaxPool

Conv (der)

■■-■■ AvgPool (der)

. MaxPool (der)

20

40

Signal len

60

1 Available at: https://pytorch.org/ (accessed: 13.01.2023).

Fig. 2. Modifications accuracy

Table 2. The metrics values of the modifications. The first three columns describe the modification type. The following ones are

metrics values for this modification

Signal length Numerical derivative Reduce type Accuracy, % Balanced accuracy, % Precision Recall F1-score ROC AUC

45 - avg 91.1 91.2 0.881 0.917 0.898 0.912

45 - conv 90.3 90.2 0.883 0.892 0.888 0.902

45 - max 90.5 90.0 0.915 0.859 0.886 0.900

45 + avg 92.0 91.7 0.915 0.896 0.905 0.917

45 + conv 91.5 91.0 0.925 0.874 0.899 0.910

45 + max 91.8 91.4 0.919 0.886 0.902 0.914

60 - avg 91.8 91.2 0.932 0.872 0.901 0.912

60 - conv 91.4 90.8 0.93 0.866 0.897 0.908

60 - max 92.2 91.8 0.922 0.893 0.907 0.918

60 + avg 91.7 91.2 0.924 0.878 0.900 0.912

60 + conv 91.2 90.8 0.908 0.883 0.895 0.908

60 + max 91.4 91.2 0.905 0.894 0.900 0.912

is 1.2 % (modifications with the max pooling or with the convolutional layer, interval length 45).

Comparison of proposed modifications. This section presents a comparison of the metrics of the various proposed modifications (Table 2) when tested using all data (without skipping subjects).

The modification with interval length of 60, max pooling, and without numerical derivative has the highest accuracy, balanced accuracy, Fl-score and ROC AUC (Area Under Receiver Operating Characteristic Curve) score. The modification with interval length 60, averaging pooling, and without numerical derivative has the highest precision score. The modification with interval length of 45, without

the numerical derivative, and with the averaging pooling has the highest recall. Almost all modifications with signal length 60 differ from the best one by no more than 0.01 on F1-score. Thus, all modifications with signal length 60 were considered for personalization. For all modifications, the maximum accuracy was achieved after the 30th epoch. But 95 % of maximum accuracy had been achieved in the first 10 epochs because accuracy of some subjects reached near 100 %. And in the process of further training the accuracy of the rest of the subjects increased.

Models personalization. As stated earlier, each type of these models may have advantages and disadvantages. Subject-dependent models require a large amount of data.

Table 3. The accuracy of each subject's personalization, %. The first value in column header is the dimensionality reduction method used. The second is the use (Derivative) or omission (Default) of the numerical derivative

Subject Avg, Derivative Max, Derivative Conv, Derivative Avg, Default Max, Default Conv, Default

2 95.4 95.4 100.0 95.4 98.5 100.0

3 100.0 100.0 100.0 100.0 100.0 100.0

4 100.0 98.6 100.0 100.0 100.0 100.0

5 100.0 100.0 100.0 100.0 100.0 100.0

6 100.0 95.2 100.0 98.8 98.8 100.0

7 100.0 100.0 100.0 100.0 100.0 98.8

8 100.0 100.0 100.0 100.0 100.0 100.0

9 90.9 89.8 76.1 84.1 83.0 85.2

10 96.7 92.6 96.7 98.4 94.2 100.0

11 100.0 100.0 100.0 100.0 100.0 100.0

13 100.0 99.1 100.0 100.0 100.0 100.0

14 100.0 100.0 100.0 100.0 100.0 100.0

15 100.0 100.0 100.0 100.0 100.0 98.0

16 100.0 100.0 100.0 100.0 100.0 100.0

17 100.0 100.0 99.0 100.0 98.0 100.0

mean 98.87 98.05 98.12 98.45 98.16 98.8

std 2.61 3.27 6.14 4.16 4.48 3.8

Table 4. The accuracy of the best model before and after personalization. Other metrics of best model after personalization

Subject Accuracy before, % Accuracy after, % Accuracy delta, % Balanced accuracy, % Precision Recall F1-score ROC AUC

2 68.2 95.4 27.3 93.2 1.0 0.864 0.927 0.932

3 74.3 100.0 25.7 100 1.0 1.0 1.0 1.0

4 100.0 100.0 0.0 100 1.0 1.0 1.0 1.0

5 56.6 100.0 43.4 100 1.0 1.0 1.0 1.0

6 100.0 100.0 0.0 100 1.0 1.0 1.0 1.0

7 60.7 100.0 39.3 100 1.0 1.0 1.0 1.0

8 90.0 100.0 10.0 100 1.0 1.0 1.0 1.0

9 54.6 90.9 36.4 88.6 0.926 0.806 0.862 0.886

10 43.0 96.7 53.7 95.8 1.0 0.917 0.957 0.958

11 97.3 100.0 2.7 100 1.0 1.0 1.0 1.0

13 100.0 100.0 0.0 100 1.0 1.0 1.0 1.0

14 97.2 100.0 2.7 100 1.0 1.0 1.0 1.0

15 91.8 100.0 8.2 1.0 1.0 1.0 1.0 1.0

16 100.0 100.0 0.0 100 1.0 1.0 1.0 1.0

17 76.0 100.0 24.0 100 1.0 1.0 1.0 1.0

General models are not taking into account the uniqueness of the subjects. So personalization may be a good solution. Another solution may be subjects grouping by similar patterns of intervals or, in a simpler version, with similar biological traits — gender, age, ethnicity, etc.

As stated previously, a personalization process was performed for all modifications with an interval length of 60. Table 3 presents accuracy after personalization for the modifications. The leave-one-subject-out (LOSO) approach [29] was used for personalization. For each subject, the following actions were performed:

— exclusion of the subject's training data from the total training dataset;

— CNN training;

— testing with the test data of the excluded subject;

— personalization of NN on the subject's training data;

— testing on subject test data.

In the process of personalization, the weights were adjusted not only for the predictor (the last convolutional layer in the network), but also for all other layers. Loss function, optimizer, count of epochs and size of batch was same with model modifications experiments (subsection

Table 5. Accuracy of the proposed convolutional neural network and analogs on the WESAD dataset

Paper Method Data Accuracy, %

[2] LDA All Chest 92.83

[2] LDA ECG 85.44

[3] MLP All 95.21

[4] kNN EDA 91.60

[5] CNN EDA 90.00

[6] CNN using GAF All 94.80

[9] CNN All 97.75 ± 2.55

[9] CNN ECG 91.75 ± 9.73

[11] Logistic regression-based classifier HR 76.38

[15] sTree EDA 95.80

[17] (Subjects only) RF Wrist 96.68 ± 3.2

[19] SVM EDA 87.50

[20] MLP BVP 99.04

[21] LDA Skin Temp, BVP, HR 87.4 ± 10.4

Ours CNN ECG 91.80

Ours (personalized) CNN ECG 98.87 ± 2.61

"Modifications"). However, the highest accuracy for most subjects was obtained within 10 epochs. Sizes of subject datasets (amount of intervals) are in range [264, 487] (mean: 373.13, std: 63.20) for training sets and in range [66, 121] (mean: 92.67, std: 15.80) for test sets.

Based on Table 3, it can be concluded that the personalized models of all modifications, on average, give approximately the same result. The modification with numerical derivative and averaging pooling shows the highest accuracy (98.87 %) averaged over users. The difference in accuracy before and after model personalization is presented in Table 4. The table also presents other metrics values of the models after personalization.

Table 5 compares the accuracy of the proposed model with analogues. Based on the table, it can be concluded that the proposed network is competitive with analogs. It can be seen that the accuracy in [20] is higher. However, in that paper, various additional features were calculated as preprocessing, which is additional resource consumption. The proposed method avoids this action.

Conclusion

This research paper proposes a convolutional neural network for human mental stress detection. The input data

were sets of consecutive RR intervals of different lengths. These intervals were calculated from ECG data of the WESAD dataset. Different modifications of convolutional neural networks for the task of stress detection were also considered. The following modifications were considered: input signal length, layer used for dimension reduction (convolutional layer, averaging pooling and max pooling were considered), and use of numerical derivative calculated as difference between consequence RR intervals. This paper proposes initial training of the network on common data, followed by personalization of the model for each individual subject. The best model, which showed the highest average accuracy after personalization, is the network using averaging pooling, numerical derivation, and with an input length of 60. The accuracy of the model at the total given is as 91.7 %. The accuracy obtained after personalization is 98.87 % ± 2.61 %. It should be noted that when learning using LOSO, personalization made it possible to significantly increase the accuracy of the model for some subjects. This is because people, both at rest and under stress, can have different HRV values.

As further plans, the training of the proposed models in the problem of determining 3 classes is considered. Validation using the SWELL dataset is also considered. The use of other input data, such as EDA in combination with HRV, is considered as an area of research.

References

1. Selye H. The Stress of Life. 1956. New York, McGraw-Hill, 324 p.

2. Schmidt P., Reiss A., Duerichen R., Marberger C., Van Laerhoven K. Introducing WESAD, a multimodal dataset for wearable stress and affect detection. Proc. of the 20th ACM International Conference on Multimodal Interaction, 2018, pp. 400-408. https://doi. org/10.1145/3242969.3242985

3. Bobade P., Vani M. Stress detection with machine learning and deep learning using multimodal physiological data. Proc. of the Second International Conference on Inventive Research in Computing Applications (ICIRCA), 2020, pp. 51-57. https://doi.org/10.1109/ icirca48905.2020.9183244

4. Aqajari S.A.H., Naeini E.K., Mehrabadi M.A., Labbaf S., Rahmani A.M., Dutt N. GSR analysis for stress: Development and validation of an open source tool for noisy naturalistic GSR data. arXiv, 2020, arXiv:2005.01834. https://doi.org/10.48550/ arXiv.2005.01834

5. Sah R.K., Ghasemzadeh H. Stress classification and personalization: Getting the most out of the least. arXiv, 2021, arXiv:2107.05666. https://doi.org/10.48550/arXiv.2107.05666

6. Ghosh S., Kim S., Ijaz M.F., Singh P.K., Mahmud M. Classification of mental stress from wearable physiological sensors using image-encoding-based deep neural network. Biosensors, 2022, vol. 12, no. 12, pp. 1153. https://doi.org/10.3390/bios12121153

7. Castaldo R., Xu W., Melillo P., Pecchia L., Santamaria L., James C. Detection of mental stress due to oral academic examination via ultra-short-term HRV analysis. Proc. of the 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2016, pp. 3805-3808. https://doi.org/10.1109/ embc.2016.7591557

8. Dalal S., Khalaf O.I. Prediction of occupation stress by implementing convolutional neural network techniques. Journal of Cases on Information Technology (JCIT), 2021, vol. 23, no. 3, pp. 27-42. https://doi.org/10.4018/jcit.20210701.oa3

9. Lai K., Yanushkevich S.N., Shmerko V.P. Intelligent stress monitoring assistant for first responders. IEEE Access, 2021, vol. 9, pp. 2531425329. https://doi.org/10.1109/access.2021.3057578

10. Cho H.M., Park H., Dong S.Y., Youn I. Ambulatory and laboratory stress detection based on raw electrocardiogram signals using a

Литература

1. Selye H. The Stress of Life. 1956. New York: McGraw-Hill, 324 p.

2. Schmidt P., Reiss A., Duerichen R., Marberger C., Van Laerhoven K. Introducing WESAD, a multimodal dataset for wearable stress and affect detection // Proc. of the 20th ACM International Conference on Multimodal Interaction. 2018. P. 400-408. https://doi. org/10.1145/3242969.3242985

3. Bobade P., Vani M. Stress detection with machine learning and deep learning using multimodal physiological data // Proc. of the Second International Conference on Inventive Research in Computing Applications (ICIRCA). 2020. P. 51-57. https://doi.org/10.1109/ icirca48905.2020.9183244

4. Aqajari S.A.H., Naeini E.K., Mehrabadi M.A., Labbaf S., Rahmani A.M., Dutt N. GSR analysis for stress: Development and validation of an open source tool for noisy naturalistic GSR data // arXiv. 2020. arXiv:2005.01834. https://doi.org/10.48550/ arXiv.2005.01834

5. Sah R.K., Ghasemzadeh H. Stress classification and personalization: Getting the most out of the least // arXiv. 2021. arXiv:2107.05666. https://doi.org/10.48550/arXiv.2107.05666

6. Ghosh S., Kim S., Ijaz M.F., Singh P.K., Mahmud M. Classification of mental stress from wearable physiological sensors using image-encoding-based deep neural network // Biosensors. 2022. V. 12. N 12. P. 1153. https://doi.org/10.3390/bios12121153

7. Castaldo R., Xu W., Melillo P., Pecchia L., Santamaria L., James C. Detection of mental stress due to oral academic examination via ultra-short-term HRV analysis // Proc. of the 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). 2016. P. 3805-3808. https://doi.org/10.1109/ embc.2016.7591557

8. Dalal S., Khalaf O.I. Prediction of occupation stress by implementing convolutional neural network techniques // Journal of Cases on Information Technology (JCIT). 2021. V. 23. N 3. P. 27-42. https:// doi.org/10.4018/jcit.20210701.oa3

9. Lai K., Yanushkevich S.N., Shmerko V.P. Intelligent stress monitoring assistant for first responders // IEEE Access. 2021. V. 9. P. 2531425329. https://doi.org/10.1109/access.2021.3057578

10. Cho H.M., Park H., Dong S.Y., Youn I. Ambulatory and laboratory stress detection based on raw electrocardiogram signals using a

convolutional neural network. Sensors, 2019, vol. 19, no. 20, pp. 4408. https://doi.org/10.3390/s19204408

11. Iqbal T., Redon-Lurbe P., Simpkin A.J., Elahi A., Ganly S., Wijns W., Shahzad A. A sensitivity analysis of biophysiological responses of stress for wearable sensors in connected health. IEEE Access, 2021, vol. 9, pp. 93567-93579. https://doi.org/10.1109/access.2021.3082423

12. Dalmeida K.M., Masala G.L. HRV features as viable physiological markers for stress detection using wearable devices. Sensors, 2021, vol. 21, no. 8, pp. 2873. https://doi.org/10.3390/s21082873

13. He J., Li K., Liao X., Zhang P., Jiang N. Real-time detection of acute cognitive stress using a convolutional neural network from electrocardiographic signal. IEEE Access, 2019, vol. 7, pp. 4271042717. https://doi.org/10.1109/access.2019.2907076

14. Woodward K., Kanjo E., Brown D.J., McGinnity T.M. On-device transfer learning for personalising psychological stress modelling using a convolutional neural network. arXiv, 2020, arXiv:2004.01603. https://doi.org/10.48550/arXiv.2004.01603

15. Liapis A., Faliagka E., Katsanos C., Antonopoulos C., Voros N. Detection of subtle stress episodes during UX evaluation: Assessing the performance of the WESAD bio-signals dataset. Lecture Notes in Computer Science, 2021, vol. 12934, pp. 238-247. https://doi. org/10.1007/978-3-030-85613-7_17

16. Al Abdi R.M., Alhitary A.E., Abdul Hay E.W., Al-Bashir A.K. Objective detection of chronic stress using physiological parameters. Medical & Biological Engineering & Computing, 2018, vol. 56, no. 12, pp. 2273-2286. https://doi.org/10.1007/s11517-018-1854-8

17. Indikawati F.I., Winiarti S. Stress detection from multimodal wearable sensor data. IOP Conference Series: Materials Science and Engineering, 2020, vol. 771, no. 1, pp. 012028. https://doi. org/10.1088/1757-899x/771/1/012028

18. Hantono B.S., Nugroho L.E., Santosa P.I. Mental stress detection via heart rate variability using machine learning. International Journal on Electrical Engineering and Informatics, 2020, vol. 12, no. 3, pp. 431-444. https://doi.org/10.15676/ijeei.2020.12.3.3

19. Ninh V.T., Smyth S., Tran M.T., Gurrin C. Analysing the performance of stress detection models on consumer-grade wearable devices. Frontiers in Artificial Intelligence and Applications, 2021, vol. 337, pp. 524-537. https://doi.org/10.3233/faia210050

20. Albaladejo-González M., Ruipérez-Valiente J.A., Gómez Mármol F. Evaluating different configurations of machine learning models and their transfer learning capabilities for stress detection using heart rate. Journal of Ambient Intelligence and Humanized Computing, 2023, vol. 14, no. 8, pp. 11011-11021. https://doi.org/10.1007/s12652-022-04365-z

21. Siirtola P. Continuous stress detection using the sensors of commercial smartwatch. Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2019 ACM International Symposium on Wearable Computers, 2019, pp. 1 198-120 1. https://doi. org/10.1145/3341162.3344831

22. Quinlan J.R. C4.5: Programs for Machine Learning. Elsevier, 2014, 312 p.

23. Oskooei A., Chau S.M., Weiss J., Sridhar A., Martínez M.R., Michel B. DeStress: deep learning for unsupervised identification of mental stress in firefighters from heart-rate variability (HRV) data. Studies in Computational Intelligence, 2021, vol. 194, pp. 93-105. https://doi.org/10.1007/978-3-030-53352-6_9

24. Hochreiter S., Schmidhuber J. Long short-term memory. Neural Computation, 1997, vol. 9, no. 8, pp. 1735-1780. https://doi. org/10.1162/neco.1997.9.8.1735

25. Xu Q., Nwe T.L., Guan C. Cluster-based analysis for personalized stress evaluation using physiological signals. IEEE Journal of Biomedical and Health Informatics, 2015, vol. 19, no. 1, pp. 275-281. https://doi.org/10.1109/jbhi.2014.2311044

26. Garcia-Ceja E., Brena R. Building personalized activity recognition models with scarce labeled data based on class similarities. Lecture Notes in Computer Science, 2015, vol. 9454, pp. 265-276. https://doi. org/10.1007/978-3-319-26401-1_25

27. Lu H., Frauendorfer D., Rabbi M., Mast M.S., Chittaranjan G.T., Campbell A.T., Gatica-Perez D., Choudhury T. Stresssense: Detecting stress in unconstrained acoustic environments using smartphones. Proc. of the 2012 ACM Conference on Ubiquitous Computing, 2012, pp. 351-360. https://doi.org/10.1145/2370216.2370270

28. Aladin A.I., Whelton S.P., Al-Mallah M.H., Blaha M.J., Keteyian S.J., Juraschek S.P., Rubin J., Brawner C.A., Michos E.D. Relation of resting heart rate to risk for all-cause mortality by gender after

convolutional neural network // Sensors. 2019. V. 19. N 20. P. 4408. https://doi.org/10.3390/s19204408

11. Iqbal T., Redon-Lurbe P., Simpkin A.J., Elahi A., Ganly S., Wijns W., Shahzad A. A sensitivity analysis of biophysiological responses of stress for wearable sensors in connected health // IEEE Access. 2021. V. 9. P. 93567-93579. https://doi.org/10.1109/access.2021.3082423

12. Dalmeida K.M., Masala G.L. HRV features as viable physiological markers for stress detection using wearable devices // Sensors. 2021. V. 21. N 8. P. 2873. https://doi.org/10.3390/s21082873

13. He J., Li K., Liao X., Zhang P., Jiang N. Real-time detection of acute cognitive stress using a convolutional neural network from electrocardiographic signal // IEEE Access. 2019. V. 7. P. 4271042717. https://doi.org/10.1109/access.2019.2907076

14. Woodward K., Kanjo E., Brown D.J., McGinnity T.M. On-device transfer learning for personalising psychological stress modelling using a convolutional neural network // arXiv. 2020. arXiv:2004.01603. https://doi.org/10.48550/arXiv.2004.01603

15. Liapis A., Faliagka E., Katsanos C., Antonopoulos C., Voros N. Detection of subtle stress episodes during UX evaluation: Assessing the performance of the WESAD bio-signals dataset // Lecture Notes in Computer Science. 2021. V. 12934. P. 238-247. https://doi. org/10.1007/978-3-030-85613-7_17

16. Al Abdi R.M., Alhitary A.E., Abdul Hay E.W., Al-Bashir A.K. Objective detection of chronic stress using physiological parameters // Medical & Biological Engineering & Computing. 2018. V. 56. N 12. P. 2273-2286. https://doi.org/10.1007/s11517-018-1854-8

17. Indikawati F.I., Winiarti S. Stress detection from multimodal wearable sensor data // IOP Conference Series: Materials Science and Engineering. 2020. V. 771. N 1. P. 012028. https://doi. org/10.1088/1757-899x/771/1/012028

18. Hantono B.S., Nugroho L.E., Santosa P.I. Mental stress detection via heart rate variability using machine learning // International Journal on Electrical Engineering and Informatics. 2020. V. 12. N 3. P. 431444. https://doi.org/10.15676/ijeei.2020.12.3.3

19. Ninh V.T., Smyth S., Tran M.T., Gurrin C. Analysing the performance of stress detection models on consumer-grade wearable devices // Frontiers in Artificial Intelligence and Applications. 2021. V. 337. P. 524-537. https://doi.org/10.3233/faia210050

20. Albaladejo-González M., Ruipérez-Valiente J.A., Gómez Mármol F. Evaluating different configurations of machine learning models and their transfer learning capabilities for stress detection using heart rate // Journal of Ambient Intelligence and Humanized Computing. 2023. V. 14. N 8. P. 11011-11021. https://doi.org/10.1007/s12652-022-04365-z

21. Siirtola P. Continuous stress detection using the sensors of commercial smartwatch // Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2019 ACM International Symposium on Wearable Computers. 2019. P. 1 1 98-120 1. https://doi. org/10.1145/3341162.3344831

22. Quinlan J.R. C4.5: Programs for Machine Learning. Elsevier, 2014. 312 p.

23. Oskooei A., Chau S.M., Weiss J., Sridhar A., Martínez M.R., Michel B. DeStress: deep learning for unsupervised identification of mental stress in firefighters from heart-rate variability (HRV) data // Studies in Computational Intelligence. 2021. V. 194. P. 93-105. https://doi.org/10.1007/978-3-030-53352-6_9

24. Hochreiter S., Schmidhuber J. Long short-term memory // Neural Computation. 1997. V. 9. N 8. P. 1735-1780. https://doi.org/10.1162/ neco.1997.9.8.1735

25. Xu Q., Nwe T.L., Guan C. Cluster-based analysis for personalized stress evaluation using physiological signals // IEEE Journal of Biomedical and Health Informatics. 2015. V. 19. N 1. P. 275-281. https://doi.org/10.1109/jbhi.2014.2311044

26. Garcia-Ceja E., Brena R. Building personalized activity recognition models with scarce labeled data based on class similarities // Lecture Notes in Computer Science. 2015. V. 9454. P. 265-276. https://doi. org/10.1007/978-3-319-26401-1_25

27. Lu H., Frauendorfer D., Rabbi M., Mast M.S., Chittaranjan G.T., Campbell A.T., Gatica-Perez D., Choudhury T. Stresssense: Detecting stress in unconstrained acoustic environments using smartphones // Proc. ofthe 2012 ACM Conference on Ubiquitous Computing. 2012. P. 351-360. https://doi.org/10.1145/2370216.2370270

28. Aladin A.I., Whelton S.P., Al-Mallah M.H., Blaha M.J., Keteyian S.J., Juraschek S.P., Rubin J., Brawner C.A., Michos E.D. Relation of resting heart rate to risk for all-cause mortality by gender after

considering exercise capacity (the Henry Ford exercise testing project). The American Journal of Cardiology, 2014, vol. 114, no. 11, pp. 1701-1706. https://doi.org/10.1016/j.amjcard.2014.08.042 29. Cawley G.C., Talbot N.L.C. Efficient leave-one-out cross-validation of kernel fisher discriminant classifiers. Pattern Recognition, 2003, vol. 36, no. 11, pp. 2585-2592. https://doi.org/10.1016/s0031-3203(03)00136-5

considering exercise capacity (the Henry Ford exercise testing project) // The American Journal of Cardiology. 2014. V. 114. N 11. P. 1701-1706. https://doi.org/10.1016/j.amjcard.2014.08.042 29. Cawley G.C., Talbot N.L.C. Efficient leave-one-out cross-validation of kernel fisher discriminant classifiers // Pattern Recognition. 2003. V. 36. N 11. P. 2585-2592. https://doi.org/10.1016/s0031-3203(03)00136-5

Authors

Maksim O. Dobrokhvalov — PhD Student, Saint Petersburg Electrotechnical University "LETI", Saint Petersburg, 197022, Russian Federation, https://orcid.org/0000-0002-0571-5836, night1337bot@gmail. com

Anton Yu. Filatov — PhD, Associate Professor, Saint Petersburg Electrotechnical University "LETI", Saint Petersburg, 197022, Russian Federation, sc 57194078312, https://orcid.org/0000-0003-4298-8523, [email protected]

Авторы

Доброхвалов Максим Олегович — аспирант, Санкт-Петербургский государственный электротехнический университет «ЛЭТИ» имени В.И. Ульянова (Ленина), Санкт-Петербург, 197022, Российская Федерация, https://orcid.org/0000-0002-0571-5836, night1337bot@ gmail.com

Филатов Антон Юрьевич — кандидат технических наук, доцент, Санкт-Петербургский государственный электротехнический университет «ЛЭТИ» имени В.И. Ульянова (Ленина), Санкт-Петербург, 197022, Российская Федерация, sc 57194078312, https://orcid.org/0000-0003-4298-8523, [email protected]

Received 19.05.2023

Approved after reviewing 31.10.2023

Accepted 22.11.2023

Статья поступила в редакцию 19.05.2023 Одобрена после рецензирования 31.10.2023 Принята к печати 22.11.2023

Работа доступна по лицензии Creative Commons «Attribution-NonCommercial»