Stress Detection Using Bio-Signal Processing: An Application of IoT and Machine Learning for Old Age Home Residents ^†

Abstract

Stress is a multifaceted physiological and psychological response that impacts health in diverse ways. This work introduces an IoT- and ML-based wearable stress detection prototype system for elderly care. The prototype developed utilizes Heart Rate, Skin Temperature, and GSR (Galvanic Skin Response) sensors, integrated with data input for real-time analysis. Among different prediction models, Random Forest was found to achieve the highest performance measured in terms of Accuracy (95.06%), Precision (95.22%), Recall (95.06%) and F1-Score (94.38%) and hence was employed for the stress-prediction purpose. Validated on old age home residents (SHEOWS, New Delhi), the device demonstrated satisfactory performance, enabling personalized care and improved stress management through precise, data-driven insights. This is preliminary research which needs to be extended appropriately in the future for further improvements and will work as an input for stress-reduction techniques for elderly people.

1. Introduction

Emotional states like stress and anxiety have a big impact on a person’s quality of life [1]. Problems resulting from stress affect a person’s mental and physical health [2]. Stress and anxiety are virtually omnipresent, affecting almost all aspects of our daily lives and society [3]. Early detection will reduce the costs associated with the condition and prevent it from becoming chronic [4].

Stress-related issues often manifest in numerous ways, affecting both mental and physical health [2]. From a mental health perspective, stress can contribute to disorders such as anxiety and depression, impacting emotional stability, concentration, and decision-making abilities [5,6,7]. The brain initiates the stress response in reaction to sensory information received from the nose, ears, or eyes [8,9]. The threat that the body perceives may be fictitious or genuine. Then, to protect itself, the body’s defense mechanisms initiate a rapid, innate process referred to as the “fight-or-flight” response or the stress response [8,9]. The brain alerts the hypothalamus through a distress signal [10,11]. The brain’s command centers are comparable to the hypothalamus. Through the Autonomic Nervous System (ANS), the hypothalamus regulates involuntary bodily functions [8,9]. The ANS consists of two main components, namely the Sympathetic Nervous System (SNS) and the Parasympathetic Nervous System (PNS). The SNS is like an automobile’s gas pedal. When faced with a fight-or-flight scenario, the hypothalamus triggers the SNS, which releases stress hormones like cortisol and adrenaline, and causes the body to react in an emergency. When the SNS is active, the Heart Rate goes up, the muscles tense up, the blood pressure goes up, and the frequency of breathing goes up.

Perceived stress is another sort of stress that results from an individual’s mental evaluation and perception of stressful conditions [8,9]. Two individuals’ periodic self-reports can be used to quantify their perceived stress. Stated differently, it refers to how an individual perceives stress [8,9].

Various studies conducted in the past have shown a positive correlation between stress and human diseases. For example, the work carried out in [12] highlights the relation between various infectious diseases and stress [12]. The study in [13] highlights how stress affects the biological system, while [2,14,15] demonstrates the harmful effects of prolonged stress on both physical and mental health.

Stress affects the fertility of males and females. For example, research work carried out in [16,17] demonstrates the adverse effect of stress on male semen quality and fertility [16,17]. Similarly, the work carried out in [18,19] highlights the negative effect of stress on female fertility. Various studies that have been carried out in the past have shown that stress causes Bowel Dysfunction and cancer in human beings [20,21].

In addition, various studies, including recent ones, have found various physiological and mental issues in old age adults. For example, the work carried out in [22] shows the vulnerability of this population to being stressed and also discusses its impact along with relevant discussions for tackling this issue [22]. But there are limitations in studies in terms of the quality or quantity for the elderly people, who must be given more importance than the conventional population [23,24].

Given the significance and multifaceted nature of stress, it becomes crucial to explore methods for its effective detection and management. Advancements in technology, including IoT, machine learning, wearable sensors that are appropriately integrated to make a system, have paved the way for objective, real-time detection, and monitoring of stress levels [25,26,27,28,29,30]. The obtained system holds the promise of moving beyond subjective assessments to provide data-driven insights that can help individuals and healthcare providers in better understanding stress levels, their reasons, and their management.

2. Proposed System

The proposed system is an extension of the work carried out in [30], where GSR, Skin Temperature, and Heart Rate (HR) sensor data are being used to form a stress classification system using fuzzy logic. The proposed system (refer to Figure 1) is an emerging approach using machine learning models for stress monitoring in elderly individuals using a smart glove device integrated with multiple sensors, including Skin Temperature, Galvanic Skin Response (GSR), Heart Rate, and Infrared (IR) sensors. The primary goal is to capture comprehensive physiological data and facilitate stress classification and prediction, enhancing well-being and enabling proactive care for elderly people.

System Overview

• Smart glove: The system’s hardware centers on a cost-effective, user-friendly smart glove (refer to Figure 2a) which is equipped with sensors which continuously monitor essential physiological parameters related to stress. A microcontroller (ESP32) enables seamless data transmission from the glove to a local host interface (mobile phone in our case).
• Local host interface: The collected sensor data are displayed in real-time on an intuitive local host interface (refer to Figure 2b), allowing healthcare professionals and caregivers to monitor stress levels effectively. This interface also facilitates the manual recording of data in an Excel spreadsheet to ensure data completeness, accuracy, and analysis.

3. Methodology Adopted

The methodology adopted for the development of the system is inspired by [30], where data obtained from the sensors are used to classify the stress utilizing the fuzzy logic algorithm. The methodology used in this work is explained below.

3.1. Data Acquisition

Data were collected from SHEOWS (Saint Hardyal Educational and Orphans’ Welfare Society), residents of an old age home, New Delhi (refer to Figure 3). Sensors were used to measure physiological parameters (Heart Rate, Galvanic Skin Response—GSR, and Skin Temperature) which are relevant to stress detection. These parameters served as input features for the stress classification system.

3.2. Stress Classification Using Fuzzy Logic

The Fuzzy Inference System (FIS), as shown in Figure 4, was developed and used to classify the stress levels based on the input signals received from GSR, HR (Heart Rate), and H &T (Body Temperature). This is same as was used in [30]. In this work, the Mamdani FIS algorithm was used since it offers various advantages over Sugeno FIS, including suitability for complex decision-making [30]. Another justification for using Mandami FIS is its usage in the previous study on which this work is based [30].

To represent the input variables (GSR, Body Temperature, and Heart Rate) and the output variable (Four Stress Levels), a triangular membership function, as shown in Figure 5, is used in the fuzzy logic system. This membership function is the same as was used in [30]. As can be seen from Figure 5, the input and output variables are categorized as ‘Low’, ‘Medium’, ‘High’, and ‘Very High’ (refer to Figure 5). The numerical representation for these functions is shown in

Table 1. Classification of stress based on membership function values [30].

Stress Level	GSR (μS)	Heart Rate (bpm)	Temperature (°C)
Relax	<2	60–70	36–37
Calm	2–4	70–90	35–36
Anxious	4–6	90–100	33–35
Stress	>6	>100	<33

[30], which depicts the various ranges of sensor output that would lead to different stress levels.

The stress classification was governed by a comprehensive set of rules (refer to

Table 2. Fuzzy rules for stress classification based on input signals membership function [30].

GSR	Body Temperature	Heart Rate
		Low	Medium	High	Very High
Low	Low	Relax	Anxious	Anxious	Low
	Medium	Relax	Calm	Anxious	Anxious
	High	Relax	Calm	Calm	Anxious
	Very High	Relax	Relax	Calm	Anxious
Medium	Low	Anxious	Anxious	Anxious	Medium
	Medium	Calm	Calm	Anxious	Anxious
	High	Calm	Calm	Calm	Anxious
	Very high	Relax	Calm	Calm	Relax
High	Low	Anxious	Calm	Anxious	High
	Medium	Anxious	Anxious	Anxious	Anxious
	High	Calm	Calm	Anxious	Anxious
	Very high	Calm	Calm	Anxious	Stress
Very High	Low	Anxious	Anxious	Stress	Anxious
	Medium	Anxious	Anxious	Anxious	Stress
	High	Anxious	Calm	Anxious	Stress
	Very high	Calm	Calm	Anxious	Stress

) derived from expert knowledge and previous work [30]. These rules map the combinations of GSR, Body Temperature, and Heart Rate to the corresponding stress levels. The defuzzification process uses the centroid method to output a crisp value to represent the final stress classification [30]. The centroid method is used as it is best for precision and smooth decision-making with respect to other prevalent methods of defuzzification [30].

3.3. Machine Learning-Based Stress Prediction

The stress classifications generated by the fuzzy system are used to create a labeled dataset containing 989 samples, which are then split into training and testing sets. Various performance metrics, including Accuracy, Precision, Recall, and F1 Score, are calculated to evaluate the performance of each model (refer to the result section for more details). Based on these assessments, Random Forest is selected as the final prediction model due to its superior performance in accurately classifying stress levels (refer to the result section for more details).

3.4. Model Testing and Validation

The trained ML model is validated using 35 old age home residents, SHEOWS, New Delhi. Considering the previous similar studies [29] which have used sample sizes varying from 5 to 255 for model validation, the sample size taken in the present study appears to be sufficient. The upcoming section presents the results that are obtained considering the above-mentioned methodology.

4. Results and Discussion

The system for stress monitoring in elderly individuals showed strong potential by integrating sensor data, fuzzy logic, and machine learning. Here is a summary of the key findings.

4.1. Model Performance

For all the considered ML models, a confusion matrix is created, as shown in

Table 3. Confusion matrix for Random Forest classification model.

	Anxious	Calm	Relax	Stress
Relax	0	5	1	0
Calm	2	156	0	0
Anxious	91	5	0	0
Stress	1	0	0	2

(created here for a specific case of Random Forest model), and then used to evaluate the important parameters, which are Accuracy, Precision, Recall, and F1 Score for different models.

Table 4. Accuracy of different ML models evaluated after model testing.

ML Models	Accuracy	Precision	Recall	F1 Score
Logistic Regression	75.67%	74.17%	75.67%	74.63%
SVM	90.87%	89.44%	90.87%	89.76%
Random Forest	95.06%	95.22%	95.06%	94.38%

shows the confusion matrix for the Random Forest Model as a sample case. A similar matrix is formed for other ML models, SVM, and Logistic Regression. Table 4 shows the model-wise parameters that are derived from the corresponding confusion matrices. As can be seen from Table 4, among other considered models, Random Forest is demonstrated to have the highest performance (across all the parameters), making it a preferred choice for stress prediction for this work. Thus, the Random Forest model is used for the validation study carried out on 35 old age home residents.

Table 5. Stress classification for old age home residents using Random Forest model.

Sl No	GSR (μS)	Temperature (°C)	Heart Rate (bpm)	Predicted Stress Level (From ML Model)	Validated Stress Level (From Old Age Home Residents)
1	3.11	32.32	77	Calm	Calm
2	5.67	34.18	71	Anxious	Anxious
3	3.81	37.13	67	Relax	Relax
4	4.85	35.44	75	Calm	Calm
5	3.48	35.67	76	Calm	Calm
6	3.14	32.45	70	Calm	Anxious
7	4.79	35.53	69	Calm	Calm
8	4.42	37.25	85	Stress	Stress
9	3.31	32.58	69	Calm	Calm
10	3.09	35.44	78	Calm	Calm
11	4.9	35.59	70	Calm	Calm
12	3.44	33.11	67	Calm	Calm
13	4.66	35.55	73	Calm	Calm
14	2.96	33.38	85	Anxious	Anxious
15	3.64	37.34	69	Calm	Calm
16	2.4	33.17	77	Calm	Calm
17	4.55	36.97	76	Calm	Calm
18	2.2	35.4	78	Calm	Calm
19	4.31	34.56	67	Anxious	Stress
20	5.93	36.36	70	Calm	Calm
21	2.28	36.32	70	Calm	Calm
22	2.1	32.38	80	Anxious	Anxious
23	5	37.1	71	Calm	Calm
24	5.75	31.99	74	Calm	Calm
25	3.7	34.66	77	Calm	Calm
26	2.25	34.7	83	Anxious	Anxious
27	2.82	35.9	78	Calm	Calm
28	2.34	33.22	78	Calm	Calm
29	5.8	34.2	80	Anxious	Anxious
30	3.08	33.89	65	Calm	Calm
31	4.9	34.85	69	Anxious	Anxious
32	2.78	34	71	Calm	Calm
33	4.52	31.33	76	Calm	Calm
34	5.72	34.28	73	Anxious	Calm
35	4.12	31.91	85	Anxious	Anxious

shows the physiological output obtained from sensors in the II, III, and IV columns, predicted stress level, as obtained from the Random Forest ML model, in the V column, and validated result from old age home residents in the VI column. It is to be noted that the stress level validation (refer to column VI of Table 5) of the predicted stress levels (obtained using ML model) has been verified manually by interacting with old age home residents at SHEOWS, New Delhi. As can be seen from Table 5, except at three instances (Serial numbers 6, 19, and 34 of the Table 5), the predicted results match with the validated results, signifying the accuracy of the trained ML model, which is Random Forest.

4.2. Strengths, Challenges, and Future Scope of the System

The main strength of the system is its high performance (Accuracy, Precision, Recall, and F1 Score). The use of fuzzy logic and machine learning led to reliable stress classification and prediction in real-time with the Random Forest and SVM models standing out among others (with more than 90% accuracy).

The current system consists of multiple components and sensors mounted on a hand glove, making it cumbersome for elderly residents in old age homes, especially for prolonged use. The glove-based design restricts mobility and may cause discomfort, reducing user convenience. Additionally, the absence of an integrated power supply limits continuous monitoring, posing a challenge for long-term usage. Furthermore, the dataset is limited in size and diversity, which may impact the accuracy and generalizability of the system’s predictive capabilities. Another concern is data security and privacy, as the system collects and transmits sensitive physiological information.

The identified challenges can be effectively addressed by integrating the following improvements into the existing design. The system should be developed into a more compact and user-friendly wearable form, such as wristbands or rings. This can be achieved using advanced integrated sensors capable of simultaneously measuring multiple physiological parameters (e.g., Heart Rate, temperature, and electrodermal activity), reducing the need for separate sensors, and enhancing user comfort. Incorporating an integrated power supply, along with low-power sensors and energy-efficient microcontrollers, will enable continuous monitoring and improve portability. Additionally, expanding the dataset with diverse samples and leveraging advanced machine learning techniques will enhance data processing and prediction accuracy. To address security concerns, implementing end-to-end encryption and secure data storage mechanisms will safeguard user information from unauthorized access and breaches. A redesigned system incorporating these improvements will provide a seamless, non-intrusive, and 24 × 7 monitoring experience, particularly for elderly residents in old age homes. This enhanced design will significantly improve usability, comfort, mobility, and functionality while ensuring data security and addressing existing limitations. These advancements define the future scope of this work.

5. Conclusions

The proposed system demonstrates substantial promise for enhancing the stress prediction and monitoring of elderly individuals through advanced technologies of IoT, ML, and data analytics. Different ML models were used to classify the sensor data as ‘Stress’, ‘Anxious’, ‘Calm, and ‘Relax’. The Random Forest model showed the highest performance (in terms of Accuracy—95.06%, Precision—95.22%, Recall—95.06%, and F1-Score—94.38%), outperforming SVM and Logistic Regression models, indicating its potential as the primary predictive model for stress prediction. The integration of IoT and fuzzy logic supported comprehensive, interpretable stress level classification along with transfer and display of the data on user interface conveniently.

Real-world deployment requires addressing issues like data security; sensor degradation and accuracy; the need for diversified data; the miniaturization of the system into more convenient wearable form, etc. Optimizing these aspects will enhance the system’s effectiveness, reliability, and ease of use in clinical and home care settings, ultimately contributing to better stress management and improved well-being for the elderly population.