Machine Learning Supporting Virtual Reality and Brain–Computer Interface to Assess Work–Life Balance Conditions for Employees

Abstract

The widespread adoption of the Industry 5.0 paradigm puts people and their applications at the center of attention and, with the increasing automation and robotization of work, the need for workers to acquire new, more advanced skills increases. The development of artificial intelligence (AI) means that expectations for workers are further raised. This leads to the need for multiple career changes from life and throughout life. Belonging to a previous generation of workers makes this retraining even more difficult. The authors propose the use of machine learning (ML), virtual reality (VR) and brain–computer interface (BCI) to assess the conditions of work–life balance for employees. They use machine learning for prediction, identifying users based on their subjective experience of work–life balance. This tool supports intelligent systems in optimizing comfort and quality of work. The potential effects could lead to the development of commercial industrial systems that could prevent work–life imbalance in smart factories for Industry 5.0, bringing direct economic benefits and, as a preventive medicine system, indirectly improving access to healthcare for those most in need, while improving quality of life. The novelty is the use of a hybrid solution combining traditional tests with automated tests using VR and BCI. This is a significant contribution to the health-promoting technologies of Industry 5.0.

1. Introduction

New times call for new solutions. The threat of depression and professional burnout is a real threat, leading to the temporary or permanent loss of qualified professionals who are difficult to replace and impossible to recreate quickly [1]. Hence, the development of preventive medicine systems, including within the Industry 5.0 paradigm, puts the human being and his or her environment at the center [2]. Quality of life (QoL) signifies a difficult-to-define subjective assessment of the positive and negative aspects of life: health, education, environment, safety, engagement/activity and work–life balance. Health-related quality of life (HRQoL) poses similar definitional problems [3]. It is worth objectifying these areas and giving them specific, generally understandable numerical meaning, using artificial intelligence (AI) [4]. Subjective well-being (SWB) should be a comprehensive measure of life satisfaction in many areas, often equated with happiness, satisfaction with life, quality of life, or mental health in a positive sense.In scientific terms, being happy means liking life, having a life close to perfect, being satisfied, being convinced that life has delivered what was desired and that nothing needs changing within in it [5]. Work-related stress and maintaining work–life balance are increasingly important problems both for work efficiency and physical and mental health. Continuous, accurate monitoring of employee stress and well-being and early detection of harmful symptoms can help prevent absenteeism through adaptive health interventions [6,7,8,9]. The widespread implementation of the Industry 5.0 paradigm puts humans and their environment in the spotlight. The increasing automation and robotization of work is creating demand for new, often higher-skilled workers. The implementation of artificial intelligence (AI) is raising expectations in this area of employee preparation. This involves changing jobs several times in a lifetime. Meanwhile, younger generations of workers expect a more comfortable life than their parents. This paper presents a predictive tool based on machine learning for finding factors affecting the subjective sense of work–life balance, supporting intelligent factory systems to improve the quality of work by providing workers with optimal mental conditions. In this case, artificial intelligence (primarily machine learning) is used in reasoning, imaging and prediction to improve quality of life, including improving employees’ work–life balance and reducing their stress at work, caused by a multitude of tasks. Possible results could help develop commercial, widely used industrial preventive medicine systems to predict and prevent the future occurrence of work–life balance disorders in different groups of workers. This shifts the center of gravity for controlling the depression epidemic from treatment to prevention, providing direct benefits to the economy and indirect benefits to the health care system, improving the quality of life of the population. ML can also improve processes and relieve employees of some of their responsibilities by finding more effective ways of performing tasks. In the long run, this should improve employees’ work–life balance and overall quality of life. ML can also be used directly to look for factors that influence a target’s subjective sense of work–life balance, predicting and preventing irregularities in this area.ML can be used to predict work-related stress and work–life balance in a multi-stage manner, at each stage supporting existing tools and procedures differently, as follows:

• At the data collection stage: ML algorithms help collect and aggregate data from a variety of sources: from employee surveys (processed from paper and electronic), Human Resources (HR) records, assessment and motivation tools (measuring and improving job performance) and wearables (health monitors, fitness monitors and others) and Internet of Things (IoR) and Industrial Internet of Things (IIoT) devices. The aforementioned data may include information on working hours, tasks, deadlines, social interactions, workload, well-being, physical activity levels and stress-related symptoms. Data can be automatically checked for accuracy (e.g., completeness, value levels, anomalies), normalized, and preprocessed using Ai/ML
• At the trait selection stage (trait vectors), identifying the mechanisms, relevant features or variables that contribute to work-related stress and work–life balance is crucial. Feature selection algorithms can help to identify the most relevant factors from the collected data, shorten the data vectors to those necessary for classification/prediction, and speed up further processing, generation of alerts and warnings, etc. In addition, processing feature vectors instead of processing entire datasets allows for data anonymization and compliance with RODO requirements.
• At the model training selection phase, once relevant features are identified, machine learning models can be trained to predict work-related stress levels or assess work–life balance. Various models, such as decision trees, random forests, support vector machines and neural networks, can be used, depending on the complexity of the problem and the nature of the data. The learned models can be periodically retrained to adapt better to the personal characteristics of a specific user.
• At the stage of anomaly detection and predictive analytics, detecting anomalies or unusual patterns in work-related stress levels or work–life balance indicators can help identify individuals or teams who may be at risk of excessive stress or work–life imbalance, and in private. Predictive models can recognize trends (unchanged, decreased, increased) in work-related stress or work–life balance levels, or predict past levels of stress and well-being based on historical data. These forecasts can help employers and employees take proactive actions to prevent burnout, improve work–life balance and improve overall well-being as part of social, branch (e.g., for medical staff) and local strategies.
• At the stage of generating personalized recommendations, by analyzing individual behaviors and preferences, ML algorithms can provide employees with personalized recommendations for better managing workload, prioritizing tasks, allocating time for relaxation and improving work–life balance, maximizing the effect of improving work–life balance in individual people as part of preventive medicine (healthy people’s medicine) and eHealth.
• At the stage of continuous improvement, ML models learn and adapt to changing work environments and individual needs. Regular updates (and replacing ML algorithms with newer, more effective ones), along with retraining models with new data, will ensure that predictions remain accurate and relevant over time. Moreover, it can be assumed that, as social awareness increases and this group of ML systems develops, the situation will improve, so systems will have to be more sensitive or aimed at a different goal: maintaining a constant level of well-being instead of restoring it.

In the article, we propose a hybrid approach to assessing the condition of an employee, i.e., a combination of traditional tests and tests using virtual reality (VR) and brain–computer interfaces (BCI). The aim of the article is to test the proposed hybrid approach and identify directions for further related research.

The hypothesis tested in this study is that the combination of machine learning (ML), virtual reality (VR), and brain–computer interface (BCI) technologies can effectively assess and potentially improve employees’ work–life balance by identifying and optimizing conditions conducive to well-being in smart workplaces. The authors suggest that the joint use of these tools can create a comprehensive model that not only predicts work–life balance based on subjective experiences but also informs systems designed to enhance comfort and quality of work. This hypothesis is based on the ability of ML models to interpret complex datasets related to employee experiences, while VR and BCI add immersive, interactive dimensions to data collection and analysis. The study contributes to Industry 5.0 by proposing these tools as potential solutions to prevent work–life imbalance, thereby bringing economic benefits and indirectly supporting preventive healthcare by reducing stress-related health problems. Statistical analysis is crucial here, as it confirms the reliability and accuracy of the ML models used. The authors test the performance of the ML models using rigorous statistical methods to ensure that they can meaningfully capture and predict work–life balance trends from the collected data. Statistical validation confirms the performance of the models by showing their ability to generalize findings, thus increasing confidence that the conclusions obtained from this approach are robust and repeatable. Additionally, statistical techniques help identify important characteristics and correlations in the data, which influences the selection of model features and increases its interpretability. This structured and statistically validated approach to ML model development and feature selection allows for better comparability and reliability of the findings.

2. Review of Literature

2.1. Data Set and Devices

Gaining a comprehensive understanding of current research and trends is essential. To conduct this study, we applied bibliometric methods to the analysis of scientific publications in this field. This included formulating research questions to identify key areas, such as the evolution of research topics over time, geographical patterns of publications, and the most influential authors and articles. We also examined collaboration networks between researchers and institutions and explored emerging topics that may have an impact on future research. By interpreting bibliometric data, this study aims to enrich ongoing discussions and establish a solid foundation for future research.

2.2. Methods

This study utilized the bibliographic databases Medline, PubMed, Scopus, Web of Science, dblp, and Ebsco, chosen for their extensive research coverage and rich citation data, which support thorough bibliometric analysis. To focus on relevant literature, we applied specific filters, narrowing the scope to original English-language articles. After filtering, we conducted a manual review of each article to ensure alignment with our study’s criteria, which determined our final sample size (Figure 1). Descriptive statistics were then applied to analyze the dataset’s main features, including notable authors, research groups, topic clusters, and emerging trends. This allowed us to map the evolution of key terminology and major research developments in the field. We tracked temporal trends to monitor shifts in research focus over time and to group publications into topic clusters, revealing relationships among various research areas. This process highlights significant themes and subfields within the study. Additionally, we analyzed metrics, such as publication years, citation counts, and authorship patterns, for a quantitative overview, while citation analysis provided insights into the impact and scientific relevance of each publication.

The study followed some of the points in the PRISMA 2020 guidelines (Figure 2): namely point #3—rationale; point #4—objectives, point #5—eligibility criteria; point #6—sources of information; point #7—search strategy; point #8—selection process; point #9—data collection process; point #13a—synthesis methods; point #20b—synthesis results; and point #23a—discussion.

For bibliometric analysis, we used tools embedded in the databases, as well as the Biblioshiny tool from the Bibliometrix Rv.4.1.3 package (GNU GPL). This methodology supports bibliometric and scientometric studies, often allowing for refined categorization by conceptual structures, research areas, authors, documents, and sources.

To refine our search to suit our research objectives, we used advanced filtered queries, limiting results to English-language articles: e.g., in WoS, searches were performed using the “Subject” field (consisting of title, abstract, keyword plus and other keywords); in Scopus, using article title, abstract and keywords; and in PubMed and the other data bases, using manual keyword sets. The databases were searched for articles using keywords. The selected publications were then further refined (see Figure 4), by manually reviewing articles, removing irrelevant items and duplicates, which resulted in our final sample size.

3. Results

A literature review of six leading databases (Medline, PubMed, Scopus, Web of Science, dblp, Ebsco) using the keywords ‘work-related stress’ and related words showed 2228 publications (189 reviews) (Figure 3), including in the last 10 years: 1518 (68.13%) and in the last 5 years: 958 (43.00%), indicating a high rate of increase in the number of publications in recent years).

Among the searched publications, only three were observed with the keywords ‘work-related stress’ + ‘artificial intelligence’ or ‘AI’ [7,8,9]. The most commonly studied group is health care workers, particularly their strategies for coping with stress and their perceptions of the effectiveness of the above strategies. It is interesting to note that at least six such strategies have so far been developed that may be applicable to other professional groups, including those with lower stress burdens: seeking social support, problem solving, adopting a healthy lifestyle, developing self-compassion, using mindfulness-based stress reduction methods, and avoidance and escape. Regardless of the strategy used, it is important to increase institutional support and develop training programs to improve the above skills [7]. Presenteeism, i.e., loss of productivity due to individual, occupational and psychological factors, is becoming an increasing problem in workplaces. The main factors include, in addition to subjectively perceived levels of pain (e.g., back pain), insufficient sleep duration [8]. Stress levels, motivation, and job satisfaction are linked to solving the health care worker (HCW) shortage and are important in addressing the HCW shortage. To evaluate the above-mentioned problems, models based on machine learning (ML) are increasingly being used, including to study the relationship between the characteristics of medical staff and their facility with reported satisfaction, stress, motivation, and the ability to meet the needs of clients/patients. First-line employees usually have lower motivation than management staff, as do employees of smaller facilities [8]. Furthermore, among the searched publications, only three were observed with the keywords “work-related stress” + machine learning” or “ML” [10,11,12]. Traditional methods of detecting work-related stress are characterized by difficulties in reflecting real conditions in laboratory conditions. A methodology was developed using ML to detect stress based on multimodal data (mouse movement variability, keyboard usage and heart rate) collected from discrete sources in an experiment simulating a realistic group office environment. Three levels of stress, valence, and arousal assessed using support vector machines, random forests, and gradient boosting models. were compared using cross-validation. However, the highest F1 scores in multi-class prediction of stress, arousal, and valence were 0.625, 0.631, and 0.775, respectively. A combination of mouse and keyboard functions may be better at detecting stress in an office environment than heart rate variability [10]. From a systemic perspective, optimization of productivity and quality requires taking into account many factors: machine performance, working environment and safety conditions, organization of production processes, and human factors, and their optimization and elimination of risks are not quick, simple and easy. Detecting employee stress and fatigue using wearables and ML enables the integration of data from monitoring production processes and the work environment within one platform.Stress detection based on electrocardiographic (ECG) signals using a 1D convolutional neural network gave an accuracy of 88.4%, F1 result = 0.90 [11]. An EMFi sensor integrated into the office chair can also be used to measure pressure changes caused by the user’s body movements and heartbeats, along with an ML-based model for classifying various work behaviors (moving, typing, speaking, browsing) with an accuracy of 91%.Behavioral and physiological measurements, integrated with the ML model, allow for early detection of stress within smart offices [12]. Publications with the keyword ‘work–life balance’ and related words were observed 3380 (Figure 4, including 233 reviews), including in the last 10 years: 2973 (87.96%), and in the last 5 years: 2002 (59.23%).

Publications with the keywords ‘work–life balance’ and ‘artificial intelligence/AI’ and related words were observed only 14 times (Figure 5).

Maintaining work–life balance within a fair work environment is also one of the six key challenges facing women in AI. This is because the demanding nature of tech roles often clashes with personal and family responsibilities. The AI industry is notorious for long hours and high pressure, which can disproportionately affect women, especially those juggling caring responsibilities. In addition, a lack of equitable workplace policies, such as flexible working hours or parental leave can further exacerbate this challenge. Women in AI also face bias and unequal opportunities, which make it harder to achieve both professional success and personal well-being. Addressing these issues requires moving towards more inclusive workplace policies and cultures that prioritize gender equality and work–life balance. It is important to ensure the lack of bias of AI/ML-based tools, promoting opportunities, institutional support and advanced reporting and analysis using ML [13]. AI may alleviate burnout among oncology workers. Interventions may be targeted at individuals (oncologists) or at the organizations in which they work. Combinatorial strategies combining other interventions may prove effective in alleviating burnout in oncology, especially through AI support [14]. Difficulties in achieving work–life balance often result from high job demands and a lack of control over work. They are also one of the reasons for the decline in the retention rate in the medical professions. Therefore, reducing the level of professional requirements (including psychological requirements) may bring beneficial changes both for family development and employees’ career development. To maintain job quality, prevention strategies should include flexible work arrangements, team management, varying levels of work engagement and the promotion of better work–life balance [15]. Interestingly, similar mechanisms operate in the academic environment: it turns out that excessive involvement can hinder career development. It is worth considering rejecting tasks that are inconsistent with the scientist’s skills/interests despite the fear of missing out on career advancement, pressure from superiors/colleagues, and hidden biases. Scientists can focus on what is truly important to them, supporting their careers and the academic community, carefully assessing commitments and effectively eliminating unaligned opportunities [16]. The professional community often shares similar problems: even among pediatric surgeons, the most common sources of anxiety were administrative issues (45.2%), work–life balance (42.3%) and, less frequently, personal problems (18.8%) or relationships with colleagues(17.9%).There was much less concern about poor leadership, possible loss of autonomy or lack of support/mentoring. Prevention strategies perceived as effective included reducing burdens (mental and physical, including those concerning wellness), inclusion in administrative decision-making processes, support (especially after adverse events), remuneration and career development opportunities. Despite the many sources of stress identified, stress alleviation interventions are relatively simple to implement to promote employee well-being [17]. In addition to work-related factors, organizational factors (positive environment), relationship factors (role model, cooperation, recognition and respect) and individual factors (spiritual drive) are important. This confirms that creating professional passion is a key element of the strategy of using professional passion and its impact on work results [18]. The exact characteristics of work environments in the area of stress factors have not been clearly assessed, but ML based on the analysis and aggregation of results from many sources can help. One such source is the Nurses’ Occupational Stress Scale (NOSS). Test–retest correlation coefficients were 0.71–0.83 [19]. Visual pollution may also prove to be a significant problem for the subjectively perceived comfort of work and life [20]. So far, links have been observed leading from work–life balance and life satisfaction to hopelessness. Improving job satisfaction, career development and regular working hours should improve employees’ subjective well-being [21]. Training programs should take into account the current and previous educational challenges of working people, especially advanced degrees and double degrees, and support their professional development and goals even after graduation [22]. Staff development also includes recruiting a unique and diverse range of graduates and trainees, especially through mentoring. The decision to pursue a career is motivated primarily by work–life balance, practical aspects of a given field and mentoring [23]. Deep learning-based predictive models were used to examine the impact of lifestyle characteristics and factors on nurses’ research activities during the COVID-19 pandemic. This demonstrated the power of secondary analyses of earlier survey data using AI [24,25]. Even ChatGPT 4.0 (OpenAI, San Francisco, CA, USA) can reduce the workload of employees by generating initial answers to customer/patient questions and creating, for example, educational resources for patients. However, the effects of ChatGPT still require editing and checking of facts/figures/analysis/predictions to ensure accuracy and appropriately high quality responses. This can increase efficiency, save time, enrich relationships or improve interaction results [26].

Publications with the keyword ‘work–life balance’, ‘machine learning/ML’ and related words were observed only six times (Figure 6).

In a study using ML on a balanced set of data, with features selected using the select K best algorithm, the relationship between the sense of balance and actual working time turned out to be the most important. Less important factors include the amount of free time during the week, working only on weekends, running a business as an individual and a subjective assessment of one’s financial situation.ANN, composed of two hidden layers with 50 and 25 neurons (ReLU and ADAM) predicts work–life balance with an accuracy of 81% [27]. We particularly noted that 35% of employees had more working hours than contracted, which can negatively affect employees’ work–life balance [27]. Internal corporate policies promote work–life balance, gender equality and a harassment-free work environment. A deep learning (DL) platform was developed that provides machine learning (ML)-based visualizations of employee survey data [28]. ML-based Digital Case Manager software determines the relationship between caregivers’ situation and subsequent stages of care based on a survey, and then helps caregivers balance work and private life in order to continue caring for family members without sacrificing their own work and personal ambitions [29]. Collecting, analyzing and visualizing the work–life balance of employees with irregular sleep and work habits (shift workers) is already possible using wearable sensors and ML-based analysis of results. This may also improve the effects of online cognitive behavioral therapy. The ability to predict well-being will reduce sleep time, thereby improving overall employee well-being, but the full potential of this system to improve sleep disorders remains to be determined in further research [30]. Qualitative and quantitative data collected in the form of semi-structured interviews are also amenable to ML analysis. On-site health promotion keeps employees healthy, productive and builds an employee brand, and easy access to fitness facilities is a key enabler [31].

4. Brain–Computer Interfaces

Brain–computer interfaces (BCIs) represent a fascinating combination of neurotechnology and computer science, enabling communication between the human brain and computer systems. While the promise of these interfaces is immense, there are a number of cyber security issues that need to be considered in the context of their development and application. Brain–computer Interfaces (BCIs) are advanced technologies that enable direct communication between the human brain and external devices (Figure 7). Different BCI paradigms use various brain signals to achieve this communication. Three prominent BCI paradigms are P300, SSVEP, and ERD/ER [31,32,33]. The P300 paradigm is a popular BCI approach that leverages a brainwave called the P300 event-related potential. The P300 is a positive voltage deflection in the electroencephalogram (EEG) signal that occurs approximately 300 milliseconds after the presentation of a stimulus. The P300 is elicited when a user recognizes a rare or target stimulus within a series of non-target stimuli. This paradigm is often used for spelling or selection tasks. In P300-based BCIs, users focus their attention on a specific target among multiple choices displayed on a screen. The BCI system detects the P300 brainwave when the desired target is presented, allowing users to make selections by directing their attention. P300-based BCIs have potential applications as communication aids, in assistive technology, and gaming [31,32,33,34].

The Steady-State Visual Evoked Potential (SSVEP) paradigm is another widely used BCI approach that relies on visual stimuli to elicit brain responses. When a user stares at a flickering visual stimulus, the brain generates electrical responses at the same frequency as the flicker, known as the SSVEP. Different stimuli flicker at distinct frequencies, enabling multiple choices or commands. In SSVEP-based BCIs, users direct their gaze toward one of several visual stimuli, each flickering at a unique frequency. The BCI system detects the frequency-specific SSVEP response and identifies the user’s selection based on the chosen stimulus. SSVEP BCIs offer high accuracy and can be employed in various applications, such as communication, control of robotic devices, and VR interactions [31,32,33,34]. The Event-Related Desynchronization/Event-Related Synchronization (ERD/ERS)paradigm focuses on modulations in the brain’s electrical activity related to motor imagery and movement. When a person imagines or executes a movement, certain frequency bands of brain activity decrease (ERD) or increase (ERS) in amplitude. ERD/ERS-based BCIs detect changes in these frequency bands during motor imagery tasks. Users are instructed to imagine performing specific movements, such as moving a hand or foot. The BCI system then identifies the patterns of ERD/ERS in the EEG signal, allowing users to control external devices or communicate through their motor imagery. BCIs can be invasive, requiring the implantation of electrodes directly into the brain tissue, or non-invasive, relying on external sensors placed on the scalp [31,32,33,34]. These BCI paradigms showcase the diversity of brain signals that can be harnessed for direct communication and control. While each paradigm has its advantages and challenges, the overall goal is to enable individuals to interact with technology seamlessly using their brain activity, offering new possibilities for communication, accessibility, and interaction [35,36,37]. The aforementioned elements are subject to neuro-security research. The search for alternative solutions for command and control has been taking place for years, ever since it was established that the mouse and keyboard are not optimal solutions. Brain–computer interfaces (BCIs) have become one of the alternatives explored. Although BCI research has focused on clinical applications for the last few decades, the last few years have shown the potential of using BCI within the Industry 4.0 paradigm to improve the performance of digital tools and cyber-physical systems, mainly thanks to electroencephalography (EEG)-based BCI applications. This is also important for research into ergonomics or occupational safety (e.g., vehicle control). It seems that a wider implementation of BCIs will allow more accurate prediction of the effects of actions and faster and more natural actions mediated by BCIs [32,33,34,35,36,37].

Brain–computer interfaces (BCIs) are technologies that enable direct communication between the brain and an external device, translating brain activity into signals that computers or other devices can interpret. BCIs work by capturing neural signals using sensors that are often placed on the scalp using electroencephalography (EEG) or sometimes implanted directly into the brain for more accurate readings. These signals are then processed and decoded to identify patterns of brain activity, allowing the BCI to provide feedback or control devices based on the user’s mental state. In healthcare, BCIs are increasingly being used in the early diagnosis and treatment of mental health conditions, such as stress and burnout, and of difficulty interacting with computers. For example, BCIs can monitor brainwave patterns associated with stress or mental fatigue, providing real-time data to detect early signs of burnout or chronic stress. By tracking changes in specific brainwaves, BCIs can alert users or healthcare professionals to increased stress levels, potentially allowing intervention before symptoms worsen. Additionally, BCIs are being used to improve a person’s ability to work effectively with a computer by providing personalized training or neurofeedback. Neurofeedback systems within BCIs can help individuals develop better focus, reduce anxiety, and improve cognitive control, which are essential for maintaining productivity and positive interactions with technology. These systems can be particularly beneficial for people who struggle with prolonged computer use, helping them adapt to digital environments with reduced mental load. In therapeutic contexts, BCIs are also being used to guide relaxation techniques by providing users with real-time feedback on their brain activity, supporting stress management and emotional resilience. As such, BCIs offer the potential not only for early diagnosis but also for ongoing support, enabling individuals to develop coping strategies and improve their digital interaction skills.

5. Materials and Methods

5.1. Dataset

The inclusion criteria for the study were age of at least 18 years and working in the profession for at least the last 12 months. A group of physiotherapists educated and practicing in the profession (n = 100) were eligible for the study (

Table 1. Sample characteristics.

Parameter	Group 1: Physiotherapists (n = 100, 100%)
Age [years]:
Mean	28.41
SD	4.22
Min	21
Q1	24
Median	28
Q3	31
Max	34
Seniority [years]:
Mean	4.89
SD	1.87
Min	1
Q1	2
Median	4
Q3	7
Max	10
Gender:
Female (F)	60 (60%)
Male (M)	40 (40%)

, Figure 8). This choice was determined by the ease of data collection (convenience sample) and the assumption that the changes discussed would be visible in this group.

The study was approved by the Bioethics Committee No. KB 391/2018 at the Ludwik Rydygier Collegium Medicum in Bydgoszcz, Nicolaus Copernicus University in Toruń.

5.2. Used Clinimetric Scales and VR–BCI

Our approach combines classic clinimetric scales with VR–BCI and ML-based analysis. Such a complex approach that automates and objectifies the assessment of stress, burnout with prediction of future states is currently unheard of, both in preventive medicine (medicine of healthy people) and in traditional approaches in psychiatry and psychology.

The study used three clinimetric scales (

Table 2. Characteristics of scales used in the study.

Scale	Direction of Change	Score
PSS10	Higher score means more stress	1–4: low 5–6: moderate 7–10: high
SWLS	Higher score means higher quality of life	score range 5–35: 5–9 extreme dissatisfaction with life, 20—neutral, 31–35—extreme satisfaction with life
MSQ-SF	Higher score means higher job satisfaction	20–100 at least 50 means job satisfaction

):

• Perceived Stress Score (PSS10),
• Satisfaction with Life Scale (SWLS),
• Minnesota Satisfaction Questionnaire-Short Form (MSQ-SF),

Multiple scales used in the study can yield more accurate ML-based analyses in both the areas of stress and life satisfaction. An original questionnaire was also used to collect demographic data.

A research station consisting of a computer with the necessary software, Esperanza EMV300 (Esperanz sp. j., Ozarów Mazowiecki, Poland) mobile virtual reality goggles and the Emotiv Epoc brain–computer interface was prepared. The choice took into account factors such as the weight of the device and the way it was mounted on the head (the devices used are made of lightweight plastic, equipped with adjustable elastic straps, so that adjusting the field of view was very easy and did not prolong the test time), as well as mobility, based on the way in which the device connects to the desktop application on the computer. It was important that the subject felt comfortable using the brain–computer interface and the VR goggles at the same time.

Emotiv EPOC (Emotiv, San Francisco, CA, USA) is a 16-channelBCI equipped with wet electrodes. Proper operation of the device therefore requires saline, which will moisten the electrodes and ensure adequate conductivity. The device comes with a USB receiver, which needs to be connected to the computer.

The study was performed in a laboratory setting on a prepared test rig consisting of a BCI, VR goggles and a laptop with mouse, with a proprietary application written in C# using the Windows Forms App in the Visual Studio 2022 development environment (Figure 9). In addition, a second external monitor is connected to the computer to help the operator keep track of the study. The application displays instructions and tasks to the practitioner during the study. The cortical brain activity corresponding to the calibrated thought command in the app is recorded, confirmed by a sound from the computer. Task completion times are measured so that the test results can be collated. The Emotiv EPOC device uses the EPOC Control Panel software to calibrate and assign specific thought states to keys on the keyboard or mouse and to perform actions in the authoring application. The connection to the authoring application on the PC is made via the Trinus software (Trinus Card Board, Trinus Corporation, Pasadena, CA, USA) on the PC. The displayed image is already appropriately segmented and prepared for reception via VR goggles. The test is recorded via the OBS Studio program, with which, in addition to the computer screen, the subject’s face is simultaneously recorded from the computer’s camera. In addition to this, the program allows keyboard and mouse activity to be tracked, a process which is also included in the recordings. In this way, when analyzing the study, it is possible to determine when a click was triggered.

The research study utilized a computer game developed in the Unity game engine and executed on a PC in a 2D environment. Its user interface provides options to select either standard control or inverted control schemes (Figure 10a). Upon launching the game, a timer is initiated, and the participant can manipulate the game object (e.g., crate) using the designated keys to intercept the falling brain models (Figure 10b). When a brain model contacts the ground, the value on a counter decreases. The study automatically terminates when the counter reaches zero, at which point the total gameplay duration is recorded.

The VR–BCI study consisted of a specially designed computer game with two different control modes, in which the game play time was recorded from the beginning until the player lost the last trial. During the VR–BCI phase, each participant completed both a control trial and a test trial. In the control trial, players controlled the game character by moving left with the left hand on the left key and moving right with the right hand on the right key. However, in the test trial, the control configuration was reversed: participants moved left with the right hand on the right key and moved right with the left hand on the left key. This control change was designed to assess adaptivity and responsiveness under modified control conditions. This type of research through immersion in VR can initiate a whole series of applications operated in an intuitive way to study cognitive deficits resulting from chronic fatigue, burnout or cognitive aging (Figure 11). In work environments saturated with IT solutions, this will allow for faster detection of employee indisposition and the application of preventive strategies.

Data were recorded using Unicorn software at a rate of 255 samples per second. Each sample was analyzed to determine in which segments of time the test subject performed movement. For each sample collected during the test, the corresponding activity that the subject was currently performing was assigned. The activities were divided into three categories: immobility, box movement to the right, and box movement to the left.

5.3. Statistical Analysis

The results were saved in an MS Excel spreadsheet (.xlsx), statistically analysed using Statistica 13 (StatSoft, Tulsa, OK, USA) and converted to .csv and preprocessed for usability for ML. Incomplete or outlier data were removed. The normality of the data distribution was checked using the Shapiro–Wilk test (α = 0.05). Values for distributions close to the normal distribution were presented by mean and standard deviation (SD). Values for distributions different from the normal distribution were presented by median, minimum value, maximum value and lower quartile (Q1) and upper quartile (Q3). Spearman’s rank correlation coefficient (rho Spearman) was used for correlation analysis. Statistical calculations were performed using ANOVA (analysis of variance) with Tuckey’s post-hoc test applied where appropriate. Significance was set at p < 0.05.

The normality of the data distribution was assessed using the Shapiro–Wilk test, a robust statistical method for evaluating the underlying distribution of the variables. For distributions that were found to be close to the normal distribution, the results were presented using the mean and standard deviation, providing a clear and concise summary of the central tendency and dispersion of the data. For variables with distributions that deviated from the normal distribution, the results were reported using the median, minimum value, maximum value, and lower and upper quartiles, offering a more detailed and informative representation of the data’s characteristics.

To further explore the relationships between the variables, Spearman’s rank correlation coefficient was employed. This nonparametric measure of statistical dependence allowed the researchers to identify any significant associations between the different scales and measures used in the study. Additionally, ANOVA with Tukey’s posthoc test was applied to the data, enabling the researchers to perform a robust statistical comparison of the groups and identify any statistically significant differences. The Tukey method is more conservative than the NIR test, but less so than the Scheffé test. The significance level was set at p < 0.05, a commonly accepted threshold for determining the statistical significance of findings.

5.4. ML Methods

Computational analysis methods used: artificially intelligent models (ML). Software used: Matlab R2022b (Mathworks, Natick, MA, USA).

Predictors were demographics (age, sex) and scores on three scales (one assessing the strength of stress and the other two assessing well-being/life satisfaction, Figure 12, Figure 13 and Figure 14). Accuracy and internal validity of the models were assessed using ordinal logistic regression analysis.

6. Results

In the VR–BCI test, duration of a single trial ranged from 20 to 124 s. In both the classical and inversion trials, the minimum time was 20.4 s. In contrast, the maximum time for the classical trial was 123.5 s, and for the inversion trial it was slightly less, at 115.7 s. The average times, respectively, for the classical trial were 56.4 s and for the inversion trial 53.4 s (Figure 15).

The accuracy and fit of the models were good. The study observed a statistical significance and reflected in the results and computational models greater work-related stress in the older age group, with longer job tenure, and male gender (

Table 3. Results for whole group.

Scale	PSS10	SWLS	MSQ-SF
Mean	5.39	16.42	32.94
SD	1.79	4.17	6.18
Min	2	11	22
Q1	3	14	27
Median	6	18	33
Q3	7	22	43
Max	9	25	57
Distribution	not normal	not normal	not normal
p-value	0.007	0.004	0.005

).

Significant correlations were observed between SWLS and MSF-SF scores (

Table 4. Correlations between tests results for whole group.

Scale	PSS10	SWLS	MSQ-SF
PSS10	-	n.s.	n.s.
SWLS	n.s.	-	0.889 p = 0.002
MSQ-SF	n.s.	0.878 p = 0.001	-

n.s. = not significant.

).

The results of the ANOVA test together with the Tukey’s post-hoc test showed that the aforementioned difference was statistically significant.

When testing models based on different algorithms, it was noted that there is a large variation in results not only in terms of accuracy, but also in terms of learning time (by up to several hundred %), which can be important when choosing a specific algorithm (

Table 5. The results after searching the set of 57 algorithms for a match to the dataset at hand.

Algorithm	MicroAccuracy	MacroAccuracy	Duration
SdcaMaximumEntropyMulti	0.8622	0.8711	4.3
FastTreeOva	0.9005	0.9178	5.6
LbfgsMaximumEntropyMulti	0.7102	0.7443	1.9
SdcaLogisticRegressionOva	0.8133	0.8284	12.4
LbfgsLogisticRegressionOva	0.8602	0.8748	2.3
FastForestOva	0.7819	0.8116	7.2
LightGbmMulti	0.8712	0.8983	1.8

). Nevertheless, at this stage, an accuracy of more than 90% has already been achieved, which seems sufficient to start fine-tuning the hyperparameters to better fit the model to the application and data.

Tuning the hyperparameters of the certain models resulted in an increase in accuracy of 3–5%, with a short learning time for the models (

Table 6. The results for the best five algorithms after tuning of hyperparameters.

Algorithm	MicroAccuracy	MacroAccuracy	Duration
LbfgsLogisticRegressionOva	0.9223	0.9334	2.5
SdcaMaximumEntropyMulti	0.8789	0.8955	2.6
LbfgsMaximumEntropyMulti	0.7378	0.7478	1.9

).

Cross-validation showed that the best model wasLbfgsLogisticRegressionOva (Figure 16).

The survey data were prepared in such a way that it could be further used in computational models (including the patient’s digital twin) in subsequent studies and in the construction of further, more accurate computational models.

7. Discussion

The main contribution of the paper is the comprehensive description and application of machine learning (ML) methods tailored to the analysis of specific datasets, providing a structured approach to data preprocessing, feature selection, and model training. By describing the datasets in detail, the study allows for improved repeatability and comparability, helping others understand the characteristics and challenges of the data. The inclusion of robust statistical analysis adds rigor to the study, confirming the reliability and performance significance of ML models. Finally, the paper advances the field by illustrating a repeatable framework that combines ML techniques with statistical methods to derive meaningful insights from complex data.

Despite the fact that research on affective computing itself has been conducted for more than 40 years, a search of the major bibliographic databases with the keywords ‘machine learning’ and ‘affective computing’ has not yielded any publications. The results of the studies conducted are statistically significant and interesting. They can provide a basis for replication and a starting point for extending them to other groups (occupational, age), larger groups, or other ways of studying and analyzing the results (including ML-based). It may also inspire the development of computational solutions that collect and preprocess data for ML analysis. ML is a research method that enables early prediction of unexpected events, suggesting potential new scientific applications in the analysis of work–life balance [38,39,40]. Over recent years, studies have been conducted on brain activity while playing computer games. These studies have shown that high-resolution EEG is a useful quantitative analysis tool for studying dynamic brain activity [41,42,43,44]. Studies have focused on the role of emotions, detecting disorders such as ADHD or cybernetic disorders, and examining social interactions in games [43,45,46,47,48,49]. Much of the research on games using BCI for control is based on attention control signals and meditation or emotion recognition, and much less often uses imagined movement or P300 potential. The results show that there are still many open questions and research opportunities regarding BCI-based games, as most evaluations have only looked at quantitative aspects of BCI systems, while very few studies have analyzed usability and qualitative aspects of users’ interactions with games [49,50]. Previous research on BCI-enabled games shows that game genre is important, as is the level of player experience. Action games, RPGs and vehicle simulations may not be suitable when BCI reaction time is very slow, while turn-based strategy games, CMS and puzzle games are ideal for BCI because there are usually no time constraints [49]. The research underlying this article focuses on detecting the EEG signal associated with motion planning and provides a basis for implementing new solutions for brain–computer interfaces [48]. Artificial intelligence (primarily machine learning) is used in reasoning, imaging and prediction to improve quality of life, including increasing the work–life balance of employees and reducing their stress at work caused by a multitude of tasks. The article presents a predictive tool based on machine learning to search for factors influencing the subjective sense of work–life balance. The possible results could help develop commercial, widely used preventive medicine systems used to predict and prevent future occurrences of work–life imbalance. Such a solution would perhaps shift the burden of controlling the depression epidemic from treatment to prevention, providing direct benefits to the health care system and improving the quality of life of societies [51,52,53].

The use of VR and brain–computer interfaces (BCI) to assess the work–life balance (WLB) of employees introduces a groundbreaking approach to understanding and improving employee well-being. This novel combination of immersive VR environments and BCI technology makes several important contributions: traditional methods for assessing WLB, such as surveys and interviews, often rely on subjective self-assessments that can be limited by participant biases and interpretations. VR and BCI enable objective, real-time monitoring of physiological and neurological responses during simulated work–life scenarios. By immersing employees in realistic work–life situations, VR captures authentic responses, while BCI tracks brain activity, stress levels, and cognitive load, providing unprecedented insight into mental states and emotional well-being. The integration of VR and BCI enables continuous, non-invasive monitoring of employees’ cognitive and emotional states across a variety of work-related and personal scenarios. This real-time data collection ensures that a more nuanced understanding of stressors, triggers, and recovery times can be captured. It also enables dynamic adjustments in virtual work environments to better reflect or optimize WLB. Applying machine learning to the collected sensor data is a significant step forward in this field of research. By analyzing complex patterns of brain activity, physical movements, and reactions, machine learning algorithms can uncover previously undetectable correlations between work and life conditions and employee well-being. This can lead to more personalized WLB strategies, predictive models of burnout, and recommendations for proactive interventions tailored to individual needs. By combining VR, BCI, and machine learning, this approach enables highly personalized WLB assessments. Unlike generic workplace surveys, this method can provide insight into how specific job demands or personal responsibilities impact an individual’s mental state, productivity, and overall balance. This level of detail offers the potential for targeted recommendations that address unique stressors, leading to more effective work–life balance strategies. The integration of these advanced technologies opens the way to a new era of employee well-being assessment, positioning this method as a state-of-the-art tool for organizations to track, predict and improve WLB performance. This system can become an invaluable tool for HR professionals, enabling proactive monitoring of employee health and reducing the risk of burnout, absenteeism or reduced productivity, ultimately benefiting both employees and organizations. Overall, the novelty of using VR and BCI lies in their ability to combine subjective and objective data in WLB assessment, offering unparalleled precision, personalization and predictive capabilities through machine learning analysis. This multidisciplinary approach represents a significant contribution to both employee well-being research and applied neurotechnology [54,55,56,57,58].

7.1. Potential Impacts of the Predictive Models on QoL, HRQoL and SWB

Quality of life (QoL), health-related quality of life (HRQoL), and subjective well-being (SWB) are key indicators for assessing overall life satisfaction, health status, and personal happiness. Predictive models using these indicators often aim to predict future health outcomes, psychological states, and life satisfaction based on factors such as socioeconomic status, health behaviors, and environmental conditions. QoL typically encompasses broader aspects of life, including emotional, social, and physical domains, while HRQoL focuses specifically on the impact of health on an individual’s daily life.SWB, meanwhile, captures an individual’s personal assessment of happiness and life satisfaction. Predictive models can reveal correlations or cause-and-effect relationships between these indicators and various life factors, but they can also change the way definitions are applied, emphasizing measurable or quantitative aspects. These changes can result in oversimplifications, such as focusing on physical health over mental or emotional well-being in HRQoL, or standardizing happiness in SWB, potentially omitting personal or cultural nuances. Ultimately, the use of predictive models can redefine these concepts by prioritizing elements that are easier to predict, potentially neglecting important subjective or qualitative aspects.

7.2. Limitations of Current and Previous Studies

The area relating to emotions, stress and well-being is a particularly sensitive research and clinical area, and all ML-based analyzes are an auxiliary tool for a medical specialist making a clinical decision and making a diagnosis. The small sample size may also be a limitation of the study, but it is comparable to the lower limits for this type of study. Technological limitations in the field of ML, which are surmountable with the development of intelligent technologies, may also play an important role. ML-based tools can both improve the understanding of the mechanisms of stress and well-being, as well as data collection, analysis, prediction and assessment of the possibility of changing the trend (e.g., to a beneficial one in prevention or therapy) [53,59,60,61,62,63]. Research to date shows that, while ML can be an important tool for predicting work-related stress and work–life balance, it also has several of the following limitations:

• Complexity of human behavior: human behavior is complex and depends on various factors (personal, social and organizational).ML models may have difficulty accurately capturing the nuances and intricacies of human behavior because not all mechanisms and features describing it are known. Factors such as individual differences, cultural differences and external events can influence work-related stress and work–life balance in unpredictable ways.
• Data quality: the effectiveness of ML models depends largely on the quality of the data used for training. Incomplete, inaccurate or biased data (e.g., aggregated from a specific population with its characteristics) may lead to unreliable predictions. Collecting high-quality data on work-related stress and work–life balance can be challenging because it is often based on information provided by employees themselves, which may be subjective or incomplete. It is also possible to hide mental problems (e.g., depression) for fear of the consequences.
• Interpretability: Many ML models, especially complex ones (e.g., traditional or deep neural networks), lack interpretability, making it difficult to understand how the models arrive at their predictions. This lack of transparency can be troublesome, particularly in sensitive areas such as the prediction of work-related stress, where stakeholders may need to understand the reasoning behind the predictions and how they relate to individuals’ current actual behavior.
• Limited generalization: ML models trained on historical data may not generalize well to new or unexpected situations. Work environments and individual circumstances may change over time, and models may fail to adapt to these changes, leading to inaccurate predictions.
• Bias and fairness: Machine learning models can embed biases present in the data used for training, leading to unfair or discriminatory results. Data biases, such as underrepresentation of certain demographic groups or stereotypes related to gender, race, occupations, branch, location, education, etc., can result in biased predictions about work-related stress and work–life balance.
• Human expert intervention and contextual understanding: ML models can automate the prediction process to some extent, but often require human intervention to interpret results and take appropriate action. Understanding the context of work-related stress and work–life balance issues may require human expertise that cannot be fully replicated by machines.
• Ethical considerations: Predicting work-related stress and work–life balance raises ethical concerns related to privacy, consent, and potential discrimination. Employers must ensure that employee data is collected and used responsibly, with appropriate safeguards in place to protect the rights of individuals and prevent the misuse of sensitive information [63,64,65,66,67].

A limitation of the VR–BCI stage may have been that the game code did not include a readout of the key pressed along with a timestamp, which would have greatly facilitated data analysis and increased the accuracy of the study. Determining when the subject pressed a key was done manually by analyzing the recordings and observing the moving box and EEG graph.

However, it seems that the key limitations in this area of research and practice may be due to ignorance and/or low acceptance of ML solutions in the analysis of emotions, stress and human well-being, both by users themselves and their families, as well as by medical staff [68,69,70,71,72,73]. Participants’ previous experience with immersive technologies can affect their level of comfort and ease in navigating the virtual environment, potentially giving some participants an advantage in completing the tasks. Familiarity with these technologies can also shorten the initial learning curve, leading to faster adaptation and potentially skewed performance results. Participants with extensive experience may approach the experiment with preconceived expectations or strategies that influence how they interact with and respond to the technology. Consequently, these discrepancies in previous experience can introduce variability into the results, complicating the interpretation of experimental results.

A purely statistical research framework for work–life parameters and stress levels could integrate clinimetric testing, VR-based assessments, BCI monitoring, and AI-based data analysis to create a comprehensive understanding of staff stress. Research would benefit from clinimetric testing for validated stress measurement, while VR provides immersive environments that simulate real-world work stressors and BCI enables real-time tracking of physiological responses to stress.AI models could analyze the complex interactions in these data streams, highlighting individual and group trends. However, weaknesses may include variability in participants’ experiences with VR or BCI, potentially introducing bias, and the challenge of integrating large volumes of data from different sources. Future research could focus on improving VR and BCI protocols to standardize data collection, improving AI models to interpret nuanced behavioral responses, and extending the framework to study adaptive coping mechanisms in real time.

7.3. Directions for Further Research

During the VR–BCI study, it was observed that the study group consisted of people who do not regularly play computer games that require proficiency in concrete and schematic key-pressing on the keyboard. Expanding the study to include another group, this time of people who regularly play arcade games, would allow conclusions to be drawn about the effect of playing on the EEG signal when pressing particular keys with specific meaning (in this case, moving to one side). Further research aims to use the BCI to move the box and develop the proposed algorithm more toward industrial applications, such as operating machines with BCIs in a way that requires concentration [74,75,76]. To do this, the user will have to focus his attention on a specific point on the screen (for example, a displayed left or right arrow)and, when this point is momentarily highlighted, the crate will be moved in a specific direction (using the potential of the P300). This will also increase the versatility of the proposed solution and expand its adaptability [77,78]. Further directions of research on ML applications in the field of predicting work-related stress and work–life balance, important for the development of this area of knowledge and experience and crucial for their dissemination in society, include:

• Collecting detailed, ML-enabled data, including integrating data from various sources (surveys, emails, calendars, social media, physiological sensors, and more) to gain a comprehensive understanding of individuals’ experiences.
• Longitudinal studies: tracking changes in work-related stress and work–life balance over time. This can help identify patterns, trends, and causal relationships that may not be obvious from cross-sectional data analysis.
• Causal inference: understanding the causal relationships between various factors and work-related stress or work–life balance (as association alone does not imply a cause-effect relationship) based on Bayesian causal networks or counterfactual analysis to discover the underlying mechanisms at their core.
• Personalized ML models that account for individual differences in response to stressors and work-related interventions based on transfer learning, reinforcement learning, or online learning to tailor models to individuals’ unique characteristics and preferences.
• Integrated intervention strategies: integrating ML predictions with existing and emerging intervention strategies to support employees in managing work-related stress and improving work–life balance can include the design of personalized interventions based on individual prediction results and assessing their effectiveness via using randomized trials.
• Explainable AI: further improving the interpretability and clarity of ML models for predicting work-related stress and work–life balance. This may include developing methods for generating human-readable explanations of model predictions and identifying actionable insights for stakeholders.
• Ethical AI: developing approaches to mitigating bias in ML models for predicting work-related stress and work–life balance by developing fairness-sensitive learning algorithms, checking model decisions for bias, and incorporating fairness constraints into the model training process [79,80,81].

Wider use of BCIs will effectively replace the mouse and keyboard or corresponding touch screens with hands-free solutions, useful for a wide range of applications. This will significantly expand the use of many larger systems, including those based on artificial intelligence, to control devices and vehicles, making them not only faster but also safer (e.g., with no possibility of being triggered by another person, thanks to technologies similar to Brain Fingerprinting). This aspect of cyber security using BCIs has rarely been considered so far.

Interdisciplinary research is needed with computer scientists, psychologists, sociologists, and organizational experts to gain a holistic understanding of work-related stress and work–life balance, and to integrate insights from different disciplines to develop more comprehensive predictive models and intervention strategies (

Table 7. SWOT analysis for proposed hybrid ML-based well-being monitoring system [82,83,84,85,86,87].

Strengths	Weaknesses
Automatization of data collection Intuitive use Individualized use 24/7 well-being monitoring and prediction IoT/smart home support ML-based analysis and prediction Built-in warnings and alerts Multipurpose use (work, fitness, care) Relatively low cost per device/machine	Limited number and quality of data sets to begin Lack of historical data sets Introduction requires educated specialists
Opportunities	Threats
Objectivization of work-related stress and work life balance assessment Reduced workload toward optimization Early diagnosis Preventive intervention Easier testing Novel diagnostic methods Possibility of standardization Quick further development Part of bigger systems (e.g., eHealth, smart home)	Non-acceptance of AI/ML Fear of being a part of surveillance society

) [77,78,79,80,81].

The proposed hybrid study is a pilot study. The data collected allowed us to observe differences in traditional and reversed control. This may mean that the actual direction of what we see in front of our eyes is important to the human brain, rather than the side of the keyboard or the hand used. The results should take into account that the people involved in the study are not active gamers and therefore do not have the trained reflexes associated with using the keys on the keyboard in controlling in specific directions. At this stage, it is difficult to conclude that, if the trained reflex were stronger, contradictions would be visible in the EEG signal and it would be difficult to classify whether the person intends to turn right or left. This research reconnaissance allowed us to determine the direction of further research and to adapt the research tool (the game) in terms of using brain activity to freely steer in any direction. Industry 5.0 envisions a future where human workers and advanced technologies collaborate more closely to achieve higher levels of customization and flexibility. BCIs could play a critical role in enabling this collaboration:

• Basic prevention mechanism: assessment of employee’s condition and mood;
• Cooperative work: BCIs could facilitate seamless cooperation between human workers and robots or AI systems. For example, workers could communicate their intentions directly to machines using their thoughts, enhancing coordination and efficiency;
• Customization: Industry 5.0 emphasizes the customization of products to meet individual customer needs. BCIs could allow workers to provide real-time input and adjustments to manufacturing processes, enabling rapid customization;
• Cognitive augmentation: BCIs could provide workers with cognitive augmentation, enhancing their problem-solving abilities, creativity, and decision-making skills. This could lead to more agile and adaptive production processes;
• Adaptive automation: BCIs could help adapt automation levels based on the real-time cognitive states of workers. Tasks that require human intuition and decision-making could be retained, while repetitive or physically demanding tasks could be automated;
• Responsive work environments: BCIs could contribute to creating work environments that respond to workers’ mental states. For instance, if a worker is experiencing high stress, the environment could adjust to reduce stressors;
• More efficient training: BCIs could accelerate the onboarding process for new workers by providing immediate feedback and guidance during training.

As these concepts continue to evolve, it is important to address challenges, such as data security, user privacy, and ethical considerations, when implementing BCIs in industrial settings. Collaborations between researchers, industrial experts, and ethicists will be essential to ensure the responsible and effective integration of BCIs into Industry 4.0 and Industry 5.0 scenarios.

VR/AR-BCIs are already seen as a key component of hands-free control diagnostic systems in many different areas: from Industry 4.0 and 5.0 to clinical applications, as in our case. For the reasons mentioned above, the security and privacy of BCIs are a major focus of efforts in the area of cyber security and the threats posed by hybrid and cyber warfare. Gaps in knowledge and practice in this area should be addressed without delay, as they can represent not only economic advantages, but also survival. For the reasons mentioned above, BCI security research has the same importance as that of space technologies.

8. Conclusions

It is worth systematically assessing and improving the work environment, adapting the workload to position and taking into account work–life balance. The accuracy of 0.9334 we have achieved is sufficient for clinical use, but the method itself needs to be further refined to ensure consistency and reproducibility of results.

ML can play a significant role in understanding, anticipating and alleviating work-related stress, while promoting a healthy work–life balance for employees. However, it is important to provide accessible, easy-to-use and socially popular tools for this purpose on mobile and IoT devices, as well as handling sensitive data of employees and others in an ethical manner and ensuring that privacy and confidentiality are maintained throughout the process.

ML-based systems can provide accurate and insightful analysis and complete database penetration, but they do not have the human expert’s ability to follow a hunch that intuitively reduces diagnostic risk, hence ML will remain an advisory solution.

By exploring these lines of research, we can advance the use of ML to predict work-related stress and work–life balance, ultimately leading to more effective interventions and support systems that promote employee well-being.

Research findings on human-centric technologies and AI-based solutions will play a key role in the development of Industry 5.0, which focuses on collaboration between humans and machines. Biometrics such as facial recognition, fingerprint scanning, and voice recognition will increase automation and security while personalizing the interaction between workers and AI systems. Integrating advanced biometrics will enable more seamless and adaptive collaboration between humans and machines, increasing productivity and safety in industrial environments. Industry 5.0 will emphasize ethical considerations and the privacy of personal data, ensuring the responsible and safe use of biometric technologies. Ultimately, advances in research will support the design of systems that prioritize human well-being while achieving greater efficiency with biometrics in smart factories and industrial environments.