Stress is a pervasive issue in modern society, significantly impacting both physical and mental health. Effective real-time classification of stress levels from biosignals is critical for timely interventions and improving overall well-being. However, real-time stress classification is computationally demanding and often impractical with traditional methods. This study aims to address these challenges by developing parallel computing techniques to enhance the efficiency and accuracy of stress classification using ultra-short biosignals. In this research, we utilized CUDA technology and OpenMP to accelerate the preprocessing and classification stages of biosignals. The primary goal was to reduce computation time while maintaining high classification accuracy. Using CUDA for parallel processing, we achieved a speedup of 13.21 times compared to sequential processing. OpenMP also provided a significant speedup of 4.65 times using 8 threads. These results highlight the efficiency gains from leveraging parallel computing architectures. We also explored the use of ultra-short biosignals for stress classification, achieving an accuracy of 87.98%. This represents the highest accuracy reported in studies utilizing such short time intervals, demonstrating the feasibility of using brief biosignals for effective stress detection. The findings indicate that the proposed parallel computing methods not only reduce processing time significantly but also maintain high accuracy in stress classification.
The use of parallel programming technologies to classify stress levels based on biosignal
analysis is an urgent problem, as it not only contributes to the development of innovative health
monitoring methods [
For the above reasons, it is important to develop a fast automated system that can measure
a person's stress level throughout the day, detect threatening conditions, and notify the user
about them. Given that stress is regulated by the autonomic nervous system, it can be detected
based on physiological indicators such as electrocardiogram (ECG), galvanic-skin response
(GSR), electromyogram (EMG), heart rate variability (HRV), heart rate (HR), blood pressure,
respiratory rate, and temperature [
In addition, bio-signal processing is a computationally complex process that requires a lot of time and resources, which is also a problem when the goal is to create a near real-time stress level classification system. To solve this problem, parallel programming technologies such as OpenMP and CUDA can be used.
The aim of this study is to develop a parallel biosignal processing algorithm for classifying
human stress level and the possibility of implementing it in an automated system that would
work in real time. Additionally, this study introduces the use of ultra-short biosignals of two
seconds in length. These ultra-short signals, under specific conditions, can significantly speed
up the biosignal processing process while maintaining accuracy. Studies have shown that
ultrashort HRV analysis (2 to 5 seconds) can reliably estimate stress levels under controlled
conditions [
As a result, the main contributions of this work are as follows: 1) A parallel algorithm for stress level classification based on human biosignal processing has been developed. 2) Performance indicators for the use of parallel computing technologies OpenMP and
CUDA were found and analyzed. 3) The possibility of using ultra-short biosignals of two seconds in length was proposed, which made it possible to significantly speed up the process of biosignal processing.
The developed stress level classification algorithm, optimized by parallel computing and the use of ultrashort biosignals, can be integrated into automated medical monitoring systems, which will allow faster detection of patients' stressful conditions and provide appropriate support. In particular, the analysis of performance indicators and data processing speed of such an algorithm is also critical to ensuring the security and confidentiality of patients' biometric data, which are key aspects in the field of cybersecurity in medical systems.
Stress is a growing problem in modern society that has a direct impact on human health and functioning. Therefore, there are currently many studies that aim to determine the level of human stress during the day, which allows healthcare professionals to control and influence this condition.
Mohammad Naim Rastgoo and other authors [
In their study [12], the authors Dun Hu and Lifu Gao used heart rate variability as a characterizing feature. As a result, they were able to achieve an accuracy of 93.7% using a KNN classifier.
Kun Liu and other authors in their study [
Ali I. Siam and others [
Neural networks are also often used to classify stress levels. Mahtab Vaezi and other authors in their study [13] were able to achieve an accuracy of 93.6% using an ECG signal to predict stress levels, which is still the highest among similar studies.
All of the above studies use the DriveDB [14] dataset to train the model and evaluate the classification accuracy. There are also other studies on determining the level of stress, in particular [15] proposes a method that uses a finger tapping test to analyze psychophysical states, and develops a mobile application and a machine learning model to detect and predict anomalies, which provided information for future experiments and improved model accuracy. The developed model is a multilayer recurrent neural network that demonstrates 1.5% error rate on synthetic data and 5% on real data with a similar distribution. The phenomenon of professional stress was studied in [16]. The authors presented its sources, symptoms, and development models, developing a stress classification and an information system for determining the level of stress exposure of an operator during professional activities.
Recent studies have also explored the use of ultra-short biosignals to classify stress levels,
showing promising results. For example, Shaffer and Ginsberg [
Based on the above-mentioned studies, we can say that the analysis of several biosignals to determine the level of stress is rarely used, since processing such a large amount of data takes a relatively long time. At the same time, studies that use more than one characterizing signal show a better level of accuracy compared to those that use a single signal. Therefore, our research will focus on speedup signal preprocessing by parallelization. Also, this approach will allow us to study ultra-short signals without compromising the speed of execution, which is crucial in real-time systems.
It is important to investigate the relationship between physiological signs, such as electrocardiogram, electromyogram, galvanic-skin response, heart rate and respiratory rate, and the stress state of the person being measured. To achieve this, next steps are necessary: 1) Pre-processing: the signal is extracted, divided into segments according to the window size under study, and filtered. 2) Feature extraction and selection: extracting the main characteristics from the preprocessed and normalized signal. 3) Classification: directly determining the patient's condition based on the obtained characteristics using a machine learning model.
For the experiment, we used the same dataset DriveDB as in the works analyzed above, obtained by J. A. Healey and R. W. Picard as a result of a study of the stress level of drivers while driving an unknown route in Boston. During the experiment, biosignals such as electrocardiogram, electromyogram, skin conductivity, respiration, and heart rate were continuously recorded for about 50 minutes for twenty-four drivers.
The overall route is divided into several periods with different levels of stress, respectively. The rest periods are 15 minutes each at the beginning and end of the recording. They were created to measure the driver's baseline performance and to create a situation with no stress indicators. Afterwards, the volunteers went through phases of driving in a city with heavy traffic and correspondingly high stress levels, as well as periods with medium stress levels while traveling on the highway. Since we are using electrocardiogram, skin-galvanic response, electromyogram, heart rate, and respiratory rate for our study, we first need to analyze the availability of the data we need among the drivers represented in the dataset. As a result, it turned out that records 2-4 do not contain EMG, record 13 has no GSR biosignal, and record 14 has no heart rate. Therefore, we will use only records 5-12 and 15-16 for the experiment.
Baseline offset is a type of noise that most often occurs in the ECG signal, resulting from the
subject's movements or breathing during the recording of the indicator [17, 18]. This type of
noise can significantly affect the results of the study, so it is important to get rid of it before
selection of characterizing features. It can be eliminated by discarding low-frequency
components in the signal (less than 0.5 Hz) [
= 0, … , − 1 , = 1 −1 ∑ =0 ( ) =
0, 2 / ( ), > 0.5
, = 0, … , − 1 , { 0, 1, … , −1} – frequency values of the ECG signal. where = { 0, 1, … , −1} –
numerical values of the ECG signal and = for the seventh driver. used in our study, such as EMG, heart rate, and respiratory rate, do not require any additional preprocessing.
The next step is to divide the data into segments according to the driving period, each of which corresponds to a certain level of stress. In our case, the end and beginning of each phase of the experiment are marked with a special marking signal, which can be used to separate the data and assign them an appropriate stress level score (Table 1), which will determine the correctness of the classification. characterizing features.
The resulting segments are then divided into smaller non-overlapping fragments of 2 seconds each. After that, you can move on to the next stage, the extraction and selection of
For each fragment into which we divided the recordings in the previous step, we calculate 35 statistics that we will use to train our model. As a result, we extracted 8 features of the ECG signal (SDNN, SDANN, AVNN, RMSSD, pNN50, TP, LF, HF), 6 features of the electromyogram (mean, RMS, waveform length, zero crossings, slope sign changes, Willison Amplitude), 6 features for the GSR (mean, std, frequency, magnitude, duration, area), duration, area) collected from the sensors on the arm and leg, 3 characteristics of heart rate (mean, std, HRV ratio) and 6 features corresponding to respiratory rate (mean, std, ULF, VLF, LF, HF).
At this stage, we used the scikit-learn library to train and evaluate the model. As a model, we
used Random Forest, because according to the study by Ali I. Siam [
Since during the experiment we noticed that the stage of characterizing features extraction and selection takes the most time, we decided to speed up this process using modern parallel and distributed computing technologies.
OpenMP [22] is a specification for parallelizing programs in programming languages such as C, C++, and Fortran. It makes it easy to use multithreading to speed up computing on multicore architectures.
The proposed parallel method based on OpenMP technology: 1. Load the biosignal dataset into RAM. 2. Set the number of threads to be used for parallel processing. 3. Distribute the processing of each biosignal among the available threads. 4. Perform data cleaning and normalization in parallel for each biosignal. 5. Divide each biosignal into two-second segments in parallel. 6. Calculate the characteristics for each segment of the biosignal in parallel. 7. Save the calculated characteristics in a shared matrix.
Parallelization on graphics processing units (GPUs) using CUDA [23] is an effective way to speed up computing by using the parallel computing capabilities of modern GPUs.
Parallel algorithm using CUDA: 1. Load the dataset into the GPU memory, making it available for further processing. 2. Create six CUDA cores, each of which will be responsible for processing one of the biosignals. 3. In each core, allocate the number of blocks equal to the number of characteristics calculated for the corresponding biosignal. 4. Clean and normalize the data for each biosignal to prepare it for further analysis. 5. Within each block, divide the biosignal into 2-second segments for more detailed analysis. 6. Calculate the characteristics for each segment in parallel on six cores and save the results in a matrix in the device's memory using the capabilities of parallel computing. 7. Transfer the calculated characteristics from the GPU memory to the local memory and save them in a .csv file for further use in the scikit-learn library.
The experiment was performed on a computer with an AMD Ryzen 7 5800H processor (8 cores/16 threads) and a discrete Nvidia GeForce RTX 3060 graphics processor.
Before starting to work on the parallel algorithm, it was decided to first estimate the time spent on each stage of the sequential algorithm. The results, namely the time spent, can be seen in Table 3. As we can see, the Feature Extraction stage takes three quarters of the total time spent on classifying the driver's stress level, so there is a need for parallelization to speed up this stage of the program.
In order to reduce the execution time, we used OpenMP and CUDA technologies that run on CPU and GPU, respectively. The calculation results are shown in Table 4. We can observe that as the number of threads used for computation increases, the execution time decreases, which indicates the feasibility of using parallel algorithms to solve the problem. It is also important to note that even though we used an 8-core processor with 16 threads, the execution time does Sequential not decrease when the number of worker threads is increased to 16. The reason is that there is not enough RAM to store all the necessary variables.
Now let's calculate the speedup and efficiency to evaluate our proposed algorithm. It should be noted that theoretical speedup and efficiency are not calculated for GPUs.
As you can see in Figure 2, the actual speedup value is lower than the theoretical one. This is due to the presence of critical sections in the parallel part of the code, as well as the need to synchronize, allocate and close threads. For 8 threads, the theoretical speedup is already close to the actual one.
By classifying the stress level using the Random Forest model, we managed to achieve an accuracy of 87.98%. After parallelization, the accuracy did not change significantly, which allows us to say that the parallel algorithm was designed correctly. It is also worth noting that the achieved accuracy is the highest for studies using ultra-short signals of 2 seconds.
Stress is not only an unpleasant condition, but also a serious problem in modern life. Constant stress can negatively affect a person's physical and mental health. It can lead to problems such as insomnia, decreased immunity, apathy, and an increased risk of developing serious illnesses such as heart disease and depression. In addition, stress can affect productivity and the ability to concentrate, which can affect the quality of life in general.
It is for this reason that creating a method to classify stress states is a necessity that can greatly improve the life and productivity of each individual. However, as mentioned earlier, there are a number of problems associated with achieving this goal. One of these problems is the complexity of computation, which makes this process time-consuming and makes real-time classification impossible. We tried to solve this problem in this study.
In general, the use of CUDA technology gave the greatest speedup, which amounted to 13.21 at the stage of biosignal processing. This is because CUDA uses the GPU for parallel calculations, which is particularly effective for computing large amounts of data. Thus, the use of CUDA has made it possible to significantly reduce the calculation time compared to other technologies. Using OpenMP, we also managed to achieve a certain speedup, namely 4.65 times for 8 threads. However, this speedup was less than that of CUDA. It is also important to note that as the number of threads to perform calculations increases, the speedup increases. This may indicate that with the increase in CPU power, we can expect a further increase in the speed of calculations and acceleration in general.
In addition, the use of parallel algorithms has a number of other advantages: 1. Efficient use of resources: Parallel algorithms allow you to use the full potential of computing resources, such as multi-core processors or server clusters. 2. Scalability: Parallel algorithms can be easily scaled to handle large amounts of data or perform complex calculations. 3. Ability to use the latest technologies: Parallel algorithms allow the use of the latest technologies and computing system architectures to maximize performance and efficiency.
We also tried to use ultra-short biosignals for classification. We were able to achieve an accuracy of 87.98%, which is the highest result among studies that have used such short time intervals. However, it is important to note the limitations of the proposed methods. One significant limitation is the dependency on hardware capabilities, which can vary significantly among users and may affect the performance and feasibility of the parallel algorithms in less advanced systems. Additionally, the use of ultra-short biosignals, while offering speed advantages, can introduce noise and reduce the accuracy of stress classification under less controlled conditions. Further research is needed to determine the optimal conditions and preprocessing techniques to mitigate these issues.
Future work will focus on improving accuracy by exploring other machine learning models, selecting more classification features, and identifying more relevant features that have a greater impact on stress level classification results. Additionally, expanding the dataset to include a more diverse population and different stress-inducing scenarios will help generalize the findings. Investigating real-world deployment scenarios, including mobile and wearable devices, will also be a key area of future research to ensure the practicality and robustness of the proposed system. In addition, the proposed parallel computing methods can be used in other stress level studies to optimize and accelerate the signal preprocessing stage.
To summarize, the parallel method we have proposed can be used to develop automated systems and devices [24, 25] aimed at tracking and analyzing a person's stress level throughout the day, taking into account not only medical and psychological aspects but also the importance of cybersecurity to maintain the confidentiality and integrity of personal information. This can become an important tool for healthcare professionals and researchers in the fields of psychology and medicine, allowing them to provide timely assistance and interventions to support people's physical and emotional health.
The authors extend their heartfelt gratitude to the Armed Forces of Ukraine for providing the security essential for completing this work. The determination and bravery demonstrated by the Ukrainian Army were the sole factors that made this accomplishment possible.