Advancing Mental Stress Detection in Indian Housewives: A Deep Learning Approach with Wearable Physiological Sensors and Feature Selection Methods

doi:10.21203/rs.3.rs-5023462/v1

Download PDF

Research Article

Advancing Mental Stress Detection in Indian Housewives: A Deep Learning Approach with Wearable Physiological Sensors and Feature Selection Methods

https://doi.org/10.21203/rs.3.rs-5023462/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Detecting mental stress is critical for timely intervention and support, especially in groups with distinct pressures, such as housewives. This study investigates the possibility of detecting mental stress in Indian housewives using wearable physiological sensors (separately and combinedly) and deep learning (DL) techniques, notably proposed Recurrent Neural Networks (RNN) and proposed Long Short-Term Memory (LSTM) classifiers. Electrocardiography (ECG), galvanic skin response (GSR), and Skin Temperature (ST) are among the physiological signals studied. These signals provide information on autonomic nervous system regulation, emotional arousal, and changes in peripheral blood flow caused by stress. Notably, feature selection methods have a significant effect on model’s performance. The SelectKBest and Recursive Feature Elimination (RFE) approaches demonstrate promising results in terms of precision, recall, F1-score, and accuracy achieving highest accuracy of 97.51% in LSTM using RFE and 94.23% in RNN using RFE when all data signals collected are used. This study illustrates the importance of wearable sensors for assessing mental stress in Indian housewives, highlighting DL's potential for improving stress detection. This research promises personalized therapy, which will improve mental health and quality of life. Early stress diagnosis and response can help to reduce negative health outcomes. The findings emphasise the significance of feature selection and provide significant insights for future research.

Computer Architecture and Engineering

Mental stress

ECG

GSR

IoMT Device

RNN

LSTM

Mental stress is a widespread problem that affects people all over the world, and it can have serious consequences for one's general well-being and quality of life [1]. Housewives, like other demographic groups, experience specific problems and stressors that can have an influence on their mental health. Housewifery is one of the most common gender roles placed on women, and it can lead to mental health issues. Currently, housewifery is a hefty gender role that all women are expected to do, whether they work or not, especially in traditional nations like India [2]. Within the context of this role, society frequently expects women to be submissive, emotional, obedient, and self-sacrificing. These expectations, which are embedded in children, get internalized over time and have a negative impact on women's existence and self-perception in social relationships. Understanding and correctly detecting mental stress in this population is critical for prompt support and interventions to alleviate its negative effects.

Sensor technology and machine learning (ML) algorithms have advanced in recent years, opening up new paths for monitoring and analyzing physiological data to detect mental stress. ECG, GSR, and ST are all physiological signs that have been frequently employed to assess the body's stress reaction. ECG measures heart rate variability (HRV) [3], which is known to reflect autonomic nervous system regulation; GSR measures changes in skin electrical conductivity due to emotional arousal; and ST measures changes in peripheral blood flow associated with the stress response. We hope to investigate the feasibility and usefulness of employing DL techniques, notably proposed RNN and proposed LSTM classifiers, to detect mental stress in housewives in this study. Because RNN and LSTM are excellent at identifying temporal dependencies and patterns in sequential data, they are perfect for analyzing physiological signals across time.

Our main goal is to create a dependable and accurate model that can automatically classify mental stress levels using ECG, GSR, and ST data obtained from 50 housewives. We hope that by utilizing DL techniques, we will be able to find hidden patterns and complicated correlations within the data that may not be discernible using typical statistical methodologies. Furthermore, we intend to evaluate the discriminative efficacy of each physiological signal separately and in combination, as well as potential interactions between these signals. This will allow us to examine the relative importance and contribution of each modality to total mental stress detection accuracy.

The findings of this study have the probable to expand our understanding of mental stress in housewives and pave the way for the development of personalized therapies and support systems. When stress levels are precisely identified, timely treatments can be undertaken, resulting in enhanced mental well-being and general quality of life for this demographic.

The following are the main findings of the research:

Wearable sensors are being introduced as a unique approach for assessing mental stress in housewives, as of now no one studies the mental stress of housewives using wearable sensors.
All types of data like only ECG data, only GSR data, ECG + GSR, and ECG + GSR + ST are analysed for better results.
DL techniques (proposed RNN, proposed LSTM) are used for robust stress detection.
ECG, GSR, and ST combined physiological signal analysis can be used to improve stress detection efficacy as they give the highest results among all.
Investigating two feature selection techniques namely SelectKBest and RFE to improve stress detection outcomes.
Potential insights and foundation for personalized therapies and support systems for housewives.

This study is motivated by the need for action to address mental health issues, particularly among vulnerable populations such as Indian housewives. Mental stress is a common problem that has major effects on both individuals and society. Housewives, who regularly handle household activities and family matters, are prone to high levels of stress due to a variety of socioeconomic and cultural conditions. However, diagnosing and controlling stress in this demographic is difficult because their symptoms may go unnoticed or ignored. Traditional stress detection approaches rely mainly on self-reporting, which can be inaccurate and subjective. Furthermore, the stigma associated with mental health disorders may discourage people, particularly housewives, from seeking assistance or exposing their stress levels. To solve these issues, there is an increasing interest in using technology, such as wearable physiological sensors and deep learning algorithms, to objectively quantify and monitor stress levels.

The proposed design of an IoMT device comprising ECG, GSR, and ST sensors represents a significant advancement in stress detection technology. This gadget enables non-invasive and continuous monitoring of physiological signals, providing a more realistic picture of a person's stress reaction. By combining these sensors with DL models such as RNN and LSTM, we hope to develop a reliable method for classifying mental stress levels in housewives. The results of this study have the potential to change stress evaluation and support systems for housewives. Early detection and intervention may result in individualized therapies designed according to people's specific requirements, thereby enhancing their mental health and quality of life. Furthermore, the findings may advance our understanding of stress mechanisms, leading to future studies in this area.

Numerous research has been done to investigate the use of wearable sensors to analyze physiological responses to stresses in various populations, but there are very few studies that investigate the mental stress of housewives. The mental stress detection studies cover a variety of distinct population groups, each with its own set of pressures. It is critical to understand their stress dynamics in order to develop appropriate therapies. Students (academic pressure), working professionals (job demands), parents/caregivers (family responsibilities), minorities (discrimination), veterans/military (combat exposure), the elderly (aging concerns), chronic illness patients, trauma survivors, prisoners/detainees, and housewives/homemakers are among the groups represented. Research on these populations aids in the customization of mental health support systems and the promotion of well-being. Some researchers from various countries study the mental health of the housewives of their countries. The study [4] compares the quality of the lifecycle of working women with housewives in Iran and the study [5] compares the mental health of Indian housewives and employed women. This study implies that also housewives suffer from mental health conditions just like employed women and this emphasizes the potential links between health-related excellence of lifecycle and service, emphasizing the necessity of increasing housewives' well-being.

A comprehensive review of this type of study is done in some studies [6][7][8], which implies that wearable sensors play a vital role in detecting the mental stress of various populations. The following section covers current scientific achievements in the field of wearable physiological sensors, emphasizing their importance in improving our ability to recognize and address mental stress, with a focus on the underrepresented group of housewives. The study [9] successfully used wearable sensors to assess stress levels in students during tests, with recognition accuracy ranging from 86–91% for various stress categories. Using ECG and EDA data, the Support Vector Machine (SVM) algorithm beat previous models with 91% accuracy. The examination of the confusion matrix indicated mistakes mostly in discriminating between exam and presentation stress. Physiological responses were consistent throughout these exercises. The study emphasizes the negative influence of testing settings on students' well-being and suggests the use of automated stress detection. The research [10] is centered on creating an Internet of Things (IoT) system for effective stress detection and it employs body sensors to track physiological changes caused by stress and provides a feedback system. A smart band and a chest strap module are worn on the wrist and chest, respectively. Electrodermal Activity (EDA) and Heart Rate (HR) data are transferred in real-time to a cloud-based ThingSpeak server. The data is subsequently computed by the system using a 'MATLAB Visualisation application, resulting in a stress report. A statistical examination of GSR and HR data using two-sample t-tests and correlation tests revealed substantial differences between stress and non-stress periods. According to the study, IoT-based wearable sensors have the ability to provide continuous stress monitoring and feedback. The study [11] investigates the significance of HRV-derived features as stress markers in car incidents in order to address stress-related difficulties. It creates prediction models based on ECG-derived HRV features using ML algorithms (KNN, SVM, MLP, RF, GB). The results reveal that HRV characteristics are good stress detection markers, with the top model reaching an 80% recall. AVNN, SDNN, and RMSSD are important HRV parameters for stress identification. The findings are also used to construct stress detection models for HRV parameters received from wearable devices such as the Apple Watch. This study could have applications in a variety of disciplines, including health care, anxiety treatment, and mental well-being.

The experiment was conducted inside, with the subjects exposed to the generated stressors and the developed device attached to them while data was collected. This section describes the setup and experimental protocol information. The stressors utilized here are the standard stressors that were used in some previous studies.

4.1 Setup and Placement of IoMT Device

To collect the required data, an Internet of Medical Things (IoMT) device was developed, incorporating three sensors: an ECG sensor (Heart Rate monitor AD 3282), a GSR module, and a ST sensor (DS18B20 temperature sensor). The device, as illustrated in Fig. 1, was utilized according to the experimental protocol outlined in Fig. 2. Alongside the aforementioned sensors, the device featured additional components such as a USB power supply for device power, a Real-Time Clock (RTC) module for timekeeping, a Micro SD card module for data storage, an Arduino Mega and UNO for sensor connections, and a TFT LCD display for real-time visualization of ECG, GSR, and ST readings.

During the data collection process, the ECG sensor was affixed with three electrodes or pads placed strategically on the chest to capture the electrical signals emitted by the heart. The GSR sensor was positioned on the palmar surface of the distal phalanx of the index and middle fingers, as this area exhibits a higher density of sweat glands, allowing for a more sensitive measurement of changes in skin conductance. The DS18B20 temperature sensor was positioned under the armpit of the participants, chosen for its accessibility, relative stability, and consistent temperature readings. Care was taken to ensure proper skin contact and correct positioning of the sensor tip at the center of the armpit. Overall, this setup and placement of sensors facilitated the accurate and synchronized collection of ECG, GSR, and ST data from the participating housewives, enabling a comprehensive analysis of their physiological responses to the applied stressors.

4.2 Participants and Study Protocol

Fifty housewives were selected to take part in the study. The study was conducted on housewives residing at CMPDI, Ranchi. The selection criteria included being married, between the ages of 25 and 40, and mostly involved in home tasks. Participants were excluded if they had any known cardiovascular or respiratory diseases that could impair the accuracy of physiological measurements, or if they were undergoing any stress management therapies at the time. Individuals who were interested were given thorough information regarding the study's goal, methods, and potential risks and benefits. Each subject provided informed consent before being included in the study.

Participants were instructed to refrain from consuming caffeinated beverages, engaging in strenuous physical activity, or taking any medications known to affect cardiovascular or autonomic responses for at least 24 hours before the data collection sessions in order to maintain consistency and control over the experimental conditions. Attempts were made to ensure that the sample was diverse in terms of socio-demographic variables such as age, educational background, and household size.

Participants were exposed to various stresses during the study to investigate their physiological responses which are shown in Fig. 2. There are three types of stressors: no stress, acute stress, and chronic stress.

The first stressor, "no stress," required participants to listen to relaxing music for 5 minutes. The goal was to create a baseline and examine the participants' physiological reactions while they were relaxed.
The second stressor, "acute stress," required participants to answer a puzzle in a short amount of time. They were given a certain amount of time to complete the task, which heightened the sense of urgency and mental stress.
The third stressor, "chronic stress," required subjects to do 5 minutes of serial subtraction. This task requires constant mental effort and focus, resulting in a prolonged state of tension.

The study's goal in subjecting individuals to these various stressors was to evaluate and analyze their physiological reactions in differing stress settings. This would lead to a better understanding of how stress affects individuals by providing insights into how the body responds to and adjusts to different amounts of stress.

4.3 Ethical Approval Statement

The research techniques used in this study were approved by the Department of Computer Science and Engineering (CSE) at Birla Institute of Technology (BIT), Mesra, Ranchi, India, under Approval No: CSE/HoD/Certificate/2023-24/164. All methods with human subjects followed the ethical rules established by the institutional research committee, which were consistent with the principles of the 1964 Helsinki Declaration and its revisions, or equivalent ethical standards.

5.1 Dataset Preprocessing and Feature Extraction

A preprocessing step was performed to increase the model's performance. Using a notch filter with a cutoff frequency of 0.05, eleven HRV components were extracted from the ECG data. In addition, eleven GSR characteristics were recovered from the GSR signal using a low-pass Butterworth filter and the filtered signal's first derivative. To deal with the random character of the GSR signal, the Discrete Wavelet Transform (DWT) [12] was used to divide it into approximation and detail coefficients, which reflect various frequency components. A Butterworth filter with a lower frequency cutoff of 5Hz and a sampling rate of 1000Hz was also used to extract eleven ST characteristics. These extracted features were utilized as inputs for training and testing a DL model, with 75% of the data given for training and 25% for testing.

The Standard Scaler [13] approach was used to standardize the features from the ECG, GSR, and ST modalities prior to training the DL model. This standardization technique ensured that the features had a mean of 0 and a standard deviation of 1, allowing for consistent and predictable performance during model training and testing. The flowchart of the study is illustrated in Fig. 3.

The feature extraction section of the research describes how specific features were extracted from several physiological signals (ECG, GSR, and ST) to analyze mental stress in Indian housewives. Table 1 lists and describes the features extracted from each signal in order to capture relevant physiological aspects related to stress and assist subsequent analysis using DL techniques.

Table 1

All features extracted and used in the study [14] [15] [16]
Signal	Feature	Long Form	Description
ECG	Mean RR	Mean of RR intervals	The average duration of the time intervals between consecutive R-peaks in the ECG signal.
	SDNN	Standard Deviation of NN intervals	The standard deviation of the duration of the time intervals between consecutive R-peaks in the ECG signal.
	RMSSD	Root Mean Square of Successive Differences	The square root of the average of the squared differences between adjacent RR intervals in the ECG signal.
	pNN50	Percentage of successive NN intervals differing by more than 50ms	The percentage of consecutive RR intervals that differ by more than 50 milliseconds in the ECG signal.
	VLF Power	Very Low Frequency power	The power in the Very Low Frequency range (0.0033–0.04 Hz) of the power spectrum of the ECG signal.
	LF Power	Low Frequency power	The power in the Low Frequency range (0.04–0.15 Hz) of the power spectrum of the ECG signal.
	HF Power	High Frequency power	The power in the High Frequency range (0.15–0.4 Hz) of the power spectrum of the ECG signal.
	LF/HF Ratio	Ratio of LF Power to HF Power	The ratio of the power in the Low Frequency range to the power in the High Frequency range in the power spectrum of the ECG signal.
	SampEn	Sample Entropy	A measure of the complexity or irregularity of the ECG signal based on the concept of self-matching.
	SD1	Short-term variability 1	A measure of short-term HRV derived from Poincaré plot analysis.
	SD2	Long-term variability 2	A measure of long-term HRV derived from Poincaré plot analysis.
GSR	Mean GSR	Mean of GSR signal	The average value of the GSR signal, which represents the electrical conductivity of the skin.
	Std GSR	Standard Deviation of GSR signal	The measure of the variation or spread of the GSR signal values.
	SCL	Skin Conductance Level	The baseline level of the GSR signal, which indicates the overall level of skin conductance.
	Peak Amplitude	Amplitude of GSR peaks	The magnitude of the peaks in the GSR signal, representing the intensity of the skin's response to stimuli.
	Rise Time	Time taken for the GSR signal to rise to the peak	The duration it takes for the GSR signal to increase from baseline to the peak value.
	Recovery Time	Time taken for the GSR signal to return to baseline	The duration it takes for the GSR signal to decrease from the peak value back to the baseline level.
	SCR Count	Number of Skin Conductance Responses (SCRs)	The count or number of significant changes in the GSR signal, indicating the occurrence of physiological responses.
	SCR Amplitude	Amplitude of individual SCRs	The magnitude or intensity of each individual Skin Conductance Response (SCR) in the GSR signal.
	AUC	Area Under the Curve	The total area under the curve formed by the GSR signal, providing an overall measure of the response intensity and duration.
	Half Recovery Time	Time taken for the GSR signal to recover halfway	The duration it takes for the GSR signal to decrease from the peak value to halfway between the peak and baseline levels.
ST	Mean Skin Temp	Mean of ST	The average value of the ST signal, representing the overall temperature of the skin.
	Std Skin Temp	Standard Deviation of ST	The measure of the variation or spread of the ST signal values.
	Min Skin Temp	Minimum ST	The lowest recorded value of the ST signal, indicating the minimum temperature of the skin.
	Max Skin Temp	Maximum ST	The highest recorded value of the ST signal, indicating the maximum temperature of the skin.
	Rate of Change	Rate of change of ST	The rate at which the ST changes over time, calculated as the gradient or derivative of the ST signal.
	Peak Count	Number of peaks in ST signal	The count or number of significant peaks in the ST signal, representing distinct temperature fluctuations.
	Time to Peak	Time taken for the ST signal to reach its peak	The duration it takes for the ST signal to increase from baseline to the peak value.
	Time between Peaks	Average time between consecutive peaks in ST signal	The average duration between successive peaks in the ST signal, indicating the regularity of temperature fluctuations.
	MAD	Median Absolute Deviation	A robust measure of the spread or variability of the ST signal, calculated based on the median of absolute differences from the median value.
	Skewness	Skewness	A measure of the asymmetry or deviation from the normal distribution of the ST signal.
	Kurtosis	Kurtosis	A measure of the peakedness or tails of the distribution of the ST signal.

5.2 Feature Selection

5.2.1 SelectKBest

The sklearn [17] SelectKBest model is used to minimize the dimensionality of the data while keeping or even improving the model's performance. The variance of each feature is considered first in this feature selection approach, and then a subset of features is picked based on a user-specified threshold, with the assumption that features with a higher variance may contain more important information [18]. The SelectKBest model chooses k characteristics based on their highest scores. Based on statistical scoring functions, this method is an advanced filter-type feature selection methodology that selects the top K most informative features from a given dataset. SelectKBest seeks to pick a subset S of K features that maximize a scoring function, often based on statistical tests such as the F-statistic, given a feature matrix X and matching target vector y. The scoring function assesses the relative value of each characteristic to the target variable.

Specifically, for a feature j, the F-score is computed as:

$$\:{F}_{j}=\frac{\left(n-1\right)\times\:Var\left(y\right)}{Var\left({X}_{j}\right)}\:\times\:\:\frac{E\left[y|{\stackrel{-}{X}}_{j}\right]-E\left[y\right|{X}_{j}]}{E\left[{ϵ}^{2}\right]}$$

Here, n signifies the number of samples, Var signifies the variance, and E is the expected data. The numerator is the variance in the target variable's conditional means between instances with and without feature j, normalized by the mean squared error (MSE). The denominator represents the MSE, which is a measure of the overall variability that the model can not explain. SelectKBest efficiently determines the subset of attributes that contribute most significantly to the classification problem by rating the features based on their F-scores and selecting the top K features.

5.2.2 Recursive Feature Elimination (RFE)

RFE [19] is a feature selection approach that is commonly used to increase the efficacy of ML models. It operates by deleting less important characteristics from the dataset recursively until a certain number of features remain. This iterative method assists in the identification of the most relevant attributes that significantly contribute to the model's performance. RFE is very effective at reducing overfitting and improving the interpretability of the resulting model. RFE employs a ML model in each iteration to rank the features based on their contribution to the model's performance. A scoring measure, commonly derived from the model's coefficients or feature importance scores, determines the ranking. The feature with the lowest ranking (the one with the least significance) is eliminated, and the model is retrained with the smaller feature set. This technique is continued recursively, with each iteration deleting one feature until the required amount of features is obtained.

5.3 Deep Learning Models

5.3.1 Proposed Recurrent Neural Network (RNN)

RNN [20] is a DL model used to detect temporal connections in sequential data. It is especially well suited to applications like time series forecasting, natural language processing, and speech recognition. Recurrent connections in the RNN model allow information to endure over time steps, allowing the network to retain and use context from earlier inputs. The RNN model can make accurate predictions and construct meaningful sequences by learning patterns and dependencies in the data. The RNN model excels at capturing detailed temporal patterns and can considerably improve prediction task performance by utilizing layered LSTM layers and other sophisticated structures.

The proposed RNN model utilized in this study was composed of numerous LSTM layers, allowing it to capture complex temporal patterns and dependencies in the data. The Adam optimizer and categorical cross-entropy loss were used to train the model. With a batch size of 32, the training method entailed iterating over numerous epochs. The model's performance was assessed using precision, recall, F1-score, and accuracy measures. This method permitted proper data classification and provided useful insights for the research investigation.

5.3.2 Proposed Long Short-Term Memory (LSTM)

The LSTM [21] is a form of RNN that is used to detect long-term dependencies in sequential data. It uses memory cells and gating methods to solve the vanishing gradient problem. Natural language processing and time series analysis are two common applications for LSTMs. Using the Keras API, an LSTM model was created in this study for the classification. Long-term dependencies in sequential data are well captured by LSTM, a form of recurrent neural network. By combining memory cells and gating mechanisms, it avoids the vanishing gradient problem. The model was made up of two LSTM layers with 64 memory units that were layered together using the Sequential model. The input shape was (1, n_features), which accommodated the dataset's single timestep. For multi-class classification, a Dense layer with three units and softmax activation was added after the LSTM layers. For model compilation, categorical cross-entropy loss, the 'Adam' optimizer, and the accuracy metric were utilized.

The model was trained on the training data for 50 epochs with a batch size of 32. The batch size used was a compromise between computing performance and model convergence. The verbose argument was used to track training progress. To assess the model's performance, the evaluate function was used to compute the loss and accuracy on the test set. Using the scikit-learn library, additional performance measures (accuracy, precision, recall, and F1-score) were generated. Model.predict was used to obtain predicted labels, and actual labels were obtained by converting the one-hot encoded format.

In this study, an unique machine equipped with ECG, GSR, and ST sensors was used to collect data aimed at identifying mental stress in housewives in 3 class settings. The acquired information was used to train and assess the performance of 2 DL algorithms for the classification of mental stress. Table 2 displays the model evaluation results (performance parameters) for all types of data including precision, recall, F1-score, and accuracy for two diverse models: RNN and LSTM. Data from both single physiological signals and combination signals are evaluated. Separate ST data is not analyzed here since prior studies have shown that it does not have a greater impact on mental stress.

Table 2

Performance parameters of DL models for Mental Stress Classification
Models	Data	Feature Selection Method	Precision	Recall	F1-score	Accuracy
Proposed RNN	ECG	All features	85.36	87.52	86.53	86.11
		SelectKBest	87.21	88.67	87.86	87.76
		RFE	88.47	89.23	88.83	88.62
	GSR	All features	79.64	82.18	80.81	80.29
		SelectKBest	81.12	83.76	82.28	81.89
		RFE	82.57	84.92	83.52	83.48
	ECG + GSR	All features	91.05	92.18	91.56	91.43
		SelectKBest	92.78	93.02	92.75	92.75
		RFE	93.21	94.45	93.68	93.18
	ECG + GSR + ST	All features	91.67	94.35	92.82	92.02
		SelectKBest	92.21	95.84	93.90	93.56
		RFE	94.82	96.23	95.52	94.23
Proposed LSTM	ECG	All features	86.77	88.24	87.45	87.32
		SelectKBest	88.12	89.36	88.74	89.02
		RFE	88.64	89.98	89.30	89.73
	GSR	All features	82.53	83.94	83.23	82.97
		SelectKBest	83.78	85.32	84.51	84.21
		RFE	84.92	86.17	85.96	85.39
	ECG + GSR	All features	90.78	91.32	91.05	90.91
		SelectKBest	91.86	92.15	92.18	91.82
		RFE	92.03	92.78	92.44	92.47
	ECG + GSR + ST	All features	93.12	94.18	93.56	93.18
		SelectKBest	95.23	96.48	95.70	95.39
		RFE	96.38	98.73	97.67	97.51

Analyzing the results shown in Table 2, it is clear that the approach of feature selection used has a substantial influence on the presentation of the models. Using "SelectKBest" to find the most informative features in the case of RNN for all signals resulted in improvements across all metrics. When combining data from all sensors, Precision increased from 91.67% (for the "All features" approach) to 92.21%, recall climbed from 94.35–95.84%, and the F1-score improved from 92.82–93.90%. This pattern suggests that the "SelectKBest" strategy successfully found a subset of features that contribute ideally to the model's prediction accuracy while maintaining a favorable balance of precision and recall. As a result, the overall accuracy increased from 92.02–93.56%.

Furthermore, the use of RFE when combined with RNN gave the most promising results. Precision, recall, F1-score, and accuracy all increased as compared to others. This finding highlights the iterative RFE process's ability to gradually eliminate less important features, resulting in a small feature subset that is particularly helpful to excellent model performance. When the LSTM model was considered, similar patterns were seen. When compared to the "All features", using "SelectKBest" increased precision, recall, F1-score, and accuracy to 95.23%, 96.48%, 95.70%, and 95.39%, respectively. The RFE technique provided the most notable improvement, with high precision (96.38%), recall (98.73%), F1-score (97.67%), and accuracy (97.51%). These outcomes showcase the robustness and adaptability of the RFE method, revealing its potential to optimize model performance significantly.

In this study, we used wearable physiological sensors to detect mental stress in Indian housewives using DL techniques, specifically proposed RNNs and proposed LSTM classifiers. Our goal was to create a dependable and accurate model that can automatically classify mental stress levels of housewives based on ECG, GSR, and ST measurements. We discovered that DL algorithms may successfully analyse physiological data and identify hidden patterns and connections that traditional statistical approaches may not be able to detect. We were able to capture the dynamism and temporal nature of physiological signals throughout time by using proposed RNN and proposed LSTM, which are well-suited for analysing sequential data and temporal dependencies.

This study's findings have several major implications. First, we proved the feasibility and utility of wearable physiological sensors in assessing mental stress in Indian housewives. The IoMT device designed for this study, which included ECG, GSR, and ST sensors, collected accurate and synchronised data, allowing for thorough examination of physiological reactions to stimuli. Second, by utilising DL algorithms for reliable stress assessment, our research contributes to the field of stress detection. We created a deep learning model that successfully classified 3 mental stress levels using characteristics collected from ECG, GSR, and ST data. This highlights the ability of DL algorithms to improve stress detection and monitoring strategies.

Based on the study's findings, personalised therapy and support systems hold great promise for enhancing housewives' mental health and general quality of life. Early detection and intervention in mental stress levels can lead to quick treatments and interventions, reducing the damaging consequences of stress on their health and welfare.

Finally, this study provides important insights into detecting mental stress in Indian housewives utilizing wearable physiological sensors and DL approaches. The combination of these technologies has the potential to revolutionize stress evaluation and make personalized therapies for housewives possible. Future study might look into larger and more diversified datasets, as well as the incorporation of additional physiological and contextual data, to improve stress detection and support systems for this population.

In the future, we would validate the findings through larger-scale trials, measuring long-term consequences, and improving the accuracy and efficiency of the suggested method with advanced machine learning models such as LSTM, RNN, ANN.

Machine Learning

Deep Learning

RNN

Recurrent Neural Networks

LSTM

Long Short-Term Memory

ECG

Electrocardiography

GSR

Galvanic Skin Response

Skin Temperature

RFE

Recursive Feature Elimination

HRV

Heart Rate Variability

SVM

Support Vector Machine

IoT

Internet of Things

IoMT

Internet of Medical Things

EDA

Electrodermal Activity

Heart Rate

DWT

Discrete Wavelet Transform

Compliance with Ethical Standards

Conflict of Interests: The authors affirm the absence of any conflicts of interest that might influence the outcomes or interpretation of the results.

Ethical Approval: All procedures involving human participants in this research obeyed to the ethical standards of the institutional research board. The study followed the principles outlined in the 1964 Helsinki Declaration and its subsequent revisions, or equivalent ethical benchmarks.

Informed Consent: All individual participants included in the study provided informed consent. Participants were presented with clear and comprehensive details about the study's purpose, objectives, potential risks, and benefits. They were afforded the opportunity to address any inquiries they had before their data was recorded.

Acknowledgments: We express appreciation to the cooperative involvement of the participating housewives from CMPDI, Ranchi, India in this project.

Fundings: The authors verify the absence of external funding received for this study.

Statement on Data Availability: The data supporting this research paper can be requested. Interested parties may reach out to the corresponding author to inquire about accessing the data.

Schneiderman N, Ironson G, Siegel SD (2005) Stress and health: psychological, behavioral, and biological determinants. Ann Rev Clin Psychol 1:607–628. https://doi.org/10.1146/annurev.clinpsy.1.102803.144141
Kaplan V (2023) Mental Health States of Housewives: an Evaluation in Terms of Self-perception and Codependency. Int J Ment Health Addict 21:666–683. https://doi.org/10.1007/s11469-022-00910-1
Malik M, Camm AJ (1990) s variability. Clinical cardiology, 13(8), 570–576
Saravi FK, Navidian A, Rigi SN, Montazeri A (2012) Comparing health-related quality of life of employed women and housewives: a cross sectional study from southeast Iran. 8–12
Maqbool M, Shrivastava N, Pandey M (2014) A comparative study of mental health of working women and housewives. Indian J Health wellbeing 5(11):1398–1400 [6]
Gedam, Paul S (2021) A Review on Mental Stress Detection Using Wearable Sensors and Machine Learning Techniques, in IEEE Access, vol. 9, pp. 84045–84066, 10.1109/ACCESS.2021.3085502
Chen J, Abbod M, Shieh J-S (2021) Pain and Stress Detection Using Wearable Sensors and Devices—A. Rev Sens 21(4):1030. https://doi.org/10.3390/s21041030
Said Y, Arnrich B, Ersoy C (2019) Stress detection in daily life scenarios using smart phones and wearable sensors: A survey. Journal of Biomedical Informatics, 92(August 2018), 103139. https://doi.org/10.1016/j.jbi.2019.103139
Hasanbasic A, Spahic M, Bosnjic D, adzic HH, Mesic V, Jahic O (2019) Recognition of stress levels among students with wearable sensors, 2019 18th International Symposium INFOTEH-JAHORINA (INFOTEH), East Sarajevo, Bosnia and Herzegovina, pp. 1–4, 10.1109/INFOTEH.2019.8717754
Uday S, Jyotsna C, Amudha J (2018) Detection of Stress using Wearable Sensors in IoT Platform, Second International Conference on Inventive Communication and Computational Technologies (ICICCT), Coimbatore, India, 2018, pp. 492–498, 10.1109/ICICCT.2018.8473010
Dalmeida KM, Masala GL (2021) HRV Features as Viable Physiological Markers for Stress Detection Using Wearable Devices. Sensors 21(8):2873. https://doi.org/10.3390/s21082873
Goshvarpour A et al (2016) Fusion framework for emotional electrocardiogram and galvanic skin response recognition: Applying wavelet transform. Iran J Med Phys 13(3):163–173
Bobade P, Vani M (2020), July Stress detection with machine learning and deep learning using multimodal physiological data. In 2020 Second International Conference on Inventive Research in Computing Applications (ICIRCA) (pp. 51–57). IEEE
Shaffer F, Ginsberg JP (2017) An Overview of Heart Rate Variability Metrics and Norms. Front public health 5:258. https://doi.org/10.3389/fpubh.2017.00258
Dehzangi O, Sahu V, Rajendra V, Taherisadr M (2019) GSR-based distracted driving identification using discrete & continuous decomposition and wavelet packet transform. Smart Health 14:100085
Lenhardt R, Sessler DI (2006) Estimation of mean body temperature from mean skin and core temperature. Anesthesiology 105(6):1117–1121. https://doi.org/10.1097/00000542-200612000-00011
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay E (2011) Scikit-learn: Machine Learning in Python. J Mach Learn Res 12:2825–2830
Rashid M, Kamruzzaman J, Imam T et al (2022) A tree-based stacking ensemble technique with feature selection for network intrusion detection. Appl Intell 52:9768–9781. https://doi.org/10.1007/s10489-021-02968-1
Al-Adhaileh EMSMH, Alsaade FW, Theyazn HH, Aldhyani AA, Alqarni N, Alsharif M, Irfan Uddin AH, Alahmadi, Mukti E, Jadhav MY, Alzahrani (2021) Diagnosis of Chronic Kidney Disease Using Effective Classification Algorithms and Recursive Feature Elimination Techniques, Journal of Healthcare Engineering, vol. Article ID 1004767, 10 pages, 2021. https://doi.org/10.1155/2021/1004767
Zhang J, Kim-Fung M (1998) Time series prediction using RNN in multi-dimension embedding phase space. SMC'98 Conference Proceedings. IEEE International Conference on Systems, Man, and Cybernetics (Cat. No. 98CH36218). Vol. 2. IEEE, 1998
Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780

The authors declare no competing interests.

Download PDF

Version 1

posted

You are reading this latest preprint version

Advancing Mental Stress Detection in Indian Housewives: A Deep Learning Approach with Wearable Physiological Sensors and Feature Selection Methods

Status:

Version 1

Abstract

Figures

1. INTRODUCTION

2. MOTIVATION

3. RELATED WORK

4. EXPERIMENTAL PROTOCOL

4.1 Setup and Placement of IoMT Device

4.2 Participants and Study Protocol

5. MATERIALS AND METHODS

5.1 Dataset Preprocessing and Feature Extraction

5.2 Feature Selection

5.2.1 SelectKBest

5.2.2 Recursive Feature Elimination (RFE)

5.3 Deep Learning Models

5.3.1 Proposed Recurrent Neural Network (RNN)

5.3.2 Proposed Long Short-Term Memory (LSTM)

6. EXPERIMENTAL RESULTS AND DISCUSSION

7. CONCLUSION AND FUTURE SCOPE

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1