Machine learning-based detection of workplace stress using wearable and multimodal data: a systematic literature review

Abstract

Workplace stress is a significant concern, as it negatively impacts employee wellbeing and organizational productivity and is a major contributor to burnout and job turnover. Detecting stress in real-world work environments remains challenging; however, recent advances in machine learning and deep learning techniques offer promising solutions. Furthermore, the growing availability of multimodal data and wearable sensor technologies may facilitate individual stress tracking. In this paper, a systematic literature review is presented, focusing on machine learning and deep learning approaches for detecting workplace stress using wearable and multimodal data: (a) suggested machine learning and deep learning techniques are considered, (b) dataset characteristics and the sensor modalities employed are examined, (c) detection performance is reviewed, (d) workplace contexts and professional domains are discussed, and (e) gaps and future research directions are identified. The 20 selected studies show that machine learning and deep learning models applied to physiological, behavioral, and multimodal data can effectively detect workplace stress, particularly in high-risk occupations. However, limitations such as small sample sizes, limited dataset diversity, and minimal use of central nervous system signals (e.g., electroencephalogram) remain.

1 Introduction

Stress is a state of worry or mental tension arising from challenging situations (WHO, 2023). According to the World Health Organization, stress, depression, and anxiety result in approximately 12 billion lost workdays each year, with an economic impact of around US$1 trillion (World Health Organization, 2022). Additionally, stress can increase social challenges by straining interpersonal relationships, increasing susceptibility to substance abuse, and undermining community cohesion (Sinha, 2001; Randall and Bodenmann, 2009). Vulnerable populations often experience higher levels of stress due to factors such as socioeconomic disparities and limited access to resources (Turner and Avison, 2003).

Chronic stress also has serious health implications. It contributes to cardiovascular diseases, obesity, cancer, immune system suppression, and overall reductions in wellbeing (Steptoe and Kivimäki, 2013; Tomiyama, 2019). Mental health is similarly affected, with anxiety disorders and depression closely linked to prolonged stress exposure (Cohen et al., 2015). The economic burden associated with stress-related illnesses-including healthcare costs, lost productivity, and disability expenses-is substantial (Hassard et al., 2018).

In recent years, machine learning and deep learning techniques have been increasingly applied to detect stress in various contexts (Doki et al., 2021; Richer et al., 2024), leveraging physiological, behavioral, and multimodal data to identify stress patterns with promising accuracy. However, despite this growing body of research, comparatively little attention has been paid specifically to stress detection in workplace environments, where factors such as occupational demands, high-risk tasks, and organizational pressures create unique stress dynamics.

Several review articles have examined artificial intelligence (AI)-based stress detection, often focusing on general stress recognition across laboratory or non-specific environments (Yashaswini et al., 2025; Nandan et al., 2024; Lakshmanan et al., 2025; Vos et al., 2023; Taskasaplidis et al., 2024; Pinge et al., 2024). However, these reviews typically do not focus explicitly on workplace settings, nor do they systematically analyze key limitations related to occupational applicability and real-world deployment.

The goal of this systematic literature review is to examine the current state of the art in this research area, following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Page et al., 2021) guidelines. By systematically analyzing existing studies, this review aims not only to summarize the machine learning and deep learning approaches employed but also to highlight gaps in the literature, ultimately encouraging further research into multimodal and physiological machine learning solutions tailored to detecting stress in the workplace.

2 Related work

This section presents recent review articles on machine learning applications in stress detection at a general level. The selected works are briefly examined, followed by a synthesis of the machine learning techniques most commonly employed in stress-detection research. Finally, the relevance of our work in relation to the previously discussed reviews is outlined.

A number of review articles have been published on stress detection using machine learning over the last few years. A review published in 2025 (Yashaswini et al., 2025) focuses specifically on papers that use electroencephalography (EEG) data, which is recognized for its reliability, accuracy, and precision, and aims to build complete machine learning-based systems for forecasting stress levels. Another review, Nandan et al. (2024) also includes EEG-based studies but broadens the scope to innovative biomarkers such as salivary cortisol, Mineralocorticoid Receptor-Glucocorticoid Receptor (MR-GR) balance, and galvanic skin response (GSR). These features can all be measured from the head region, such as the head surface, forehead, and mouth. This review summarizes state-of-the-art approaches, identifies relevant biomarkers, and outlines how they relate to stress.

A broader perspective is presented in Lakshmanan et al. (2025), which examines not only physiological signals traditionally used for stress detection but also data-driven approaches involving EEG and electrocardiography (ECG). Additionally, it discusses stress indicators derived from wearable devices, including Galvanic Skin Response and Skin Temperature (ST). The paper further explores the use of machine learning and deep learning methods applied across multiple domains, including workplace settings, education, and automotive environments.

Another review, Vos et al. (2023) focuses on model generalization and evaluates how well-stress detection models perform when trained on publicly available datasets. Similarly, Taskasaplidis et al. (2024) provides a wide-ranging overview of stress detection methods involving the autonomic nervous system (ANS) and the hypothalamic-pituitary-adrenal (HPA) axis, including traditional biomarkers. This review goes further by considering how body movement and posture measurements may also relate to stress.

Finally, Pinge et al. (2024) examines the various types of sensors and wearable devices used to detect and monitor stress. It highlights the widespread reliance on classical physiological signals and discusses machine learning and deep learning techniques used to analyze them.

These related works have led to the identification of certain machine learning techniques often used in stress detection. A wide variety of techniques are namely employed, including support vector machine (SVM), linear regression, logistic regression, random forests, and more advanced techniques such as multilayer perceptron (MLP), long short-term memory (LSTM), convolutional neural networks (CNN), and eXtreme Gradient Boosting (XGBoost). This enumeration of machine learning techniques is not exhaustive, and only presents the techniques most frequently encountered when analyzing existing related works. It aims to provide the reader with an overview of machine learning techniques frequently used in stress detection.

While these previous reviews have examined stress detection in general contexts, covering laboratory settings, daily-life monitoring, multi-domain datasets, wearable devices, or EEG-based biomarkers, none have concentrated specifically on stress detection within real workplace environments. Existing reviews either focus on specific sensor modalities (e.g., EEG), broad physiological monitoring, or general affective computing, but they do not analyze how machine learning models are applied in the workplace, what types of workers are studied, which data modalities are collected in occupational settings, or how these models perform under real-world job conditions.

In contrast to the previously mentioned reviews, this systematic review addresses this gap by focusing on multimodal and physiological machine learning workplace stress detection and providing a structured examination of the current state of the art, identifying trends, limitations, and opportunities specific to occupational contexts.

3 Methods

This section describes the methodology used to conduct the systematic literature review. First, the employed methodology is outlined in Subsection 3.1, followed by the research questions guiding the review in Subsection 3.2. The selection of databases and the formulation of research queries, adjusted for different databases, are explained in Subsections 3.3 and 3.4. Next, the inclusion and exclusion criteria applied during the study selection process are detailed in Subsection 3.5. Finally, the overall selection process is presented with the help of a flow diagram in Subsection 3.6.

3.1 Methodology

The present systematic review was conducted following the (Page et al., 2021) guidelines. PRISMA is a widely accepted framework that provides tools such as checklists and flow diagrams to improve the transparency, rigor, and reproducibility of systematic reviews and meta-analyses. The main objective of applying this methodology is to enhance the quality and reliability of the review process. No registration number is associated with this review.

3.2 Research questions

The present systematic literature review is guided by the following research questions (RQs):

• RQ1: What machine learning and deep learning techniques are applied for workplace stress detection?

• RQ2: What types of data or sensor modalities are used in these studies?

• RQ3: What performance levels do these models achieve?

• RQ4: What workplace contexts are considered?

3.3 Databases

Papers published between 2017 and 2025 were considered for this review. This period was chosen to capture the most recent advances in workplace stress detection using machine learning (ML) and deep learning (DL), as significant developments in AI techniques and wearable/multimodal systems have occurred during these years.

• ACM Digital Library (ACM, 2025)

• IEEE Xplore (ieeexplore, 2025)

• Science Direct (Sciencedirect, 2025)

• Springer Link (Springer, 2025)

• PubMed (Pubmed, 2025)

• Google Scholar (Google, 2025)

These databases were chosen because they cover a broad range of computer science, engineering, and healthcare journals, ensuring comprehensive coverage of relevant literature.

3.4 Research query

To formulate the research query, the main concepts of workplace stress detection, machine learning, and wearable/multimodal data were first identified. Synonyms and related terms were then derived to ensure comprehensive coverage of the literature. Table 1 presents the identified concepts along with their respective synonyms and the final search query.

3.5 Eligibility criteria

The following criteria were applied to select studies for this systematic review. Table 2 summarizes the inclusion and exclusion criteria.

For a paper to be included in this review, the terms defined in the research query had to appear in the title, keywords, or abstract. Duplicate records were excluded. The focus of this systematic review is on workplace stress detection using machine learning or deep learning, particularly in studies employing wearable or multimodal data. Studies that did not involve workplace settings or that relied solely on self-reported stress without objective measurements were excluded.

3.6 PRISMA flow diagram

The study selection process is summarized in the PRISMA flow diagram (Figure 1). This diagram illustrates the number of records identified, screened, included, and excluded at each stage of the review.

Figure 1

The identification and screening processes were carried out between September and October 2025. During the identification phase, all papers resulting from the search queries across the selected databases were considered, totaling 219 records. The screening phase began with a manual analysis of paper titles, which led to the exclusion of 155 papers. Subsequently, abstracts were manually analyzed, resulting in the exclusion of 16 additional papers. After full-text reading of the remaining 43 papers, 24 were excluded because they did not meet the defined criteria, leaving 19 papers included in this systematic review.

The critical appraisal was conducted solely by the author, who worked independently in selecting studies. The most frequent reason for exclusion was that the papers did not focus on workplace stress.

4 Results

In this section, after presenting the distribution across different countries of the papers, we will analyze them according to the research questions defined in Subsection 3.2. Each research question is addressed in a separate subsection.

4.1 Country distribution

Figure 2 shows the distribution of the geographic locations of the authors. The country of the university or organization to which the author is affiliated is considered. If the authors of a paper have different geographic locations, the majority country is taken into consideration.

Figure 2

The geographical distribution of the analyzed studies shows that research on multimodal and physiological machine learning workplace stress detection is highly concentrated in a few regions. India is the most represented country, followed by several European nations such as Switzerland, Belgium, Poland, Portugal, Spain, and Norway, along with the United States, Japan, and Mexico.

This distribution reveals a noticeable lack of global diversity, as large parts of the world (particularly Africa, South America, and sections of Asia) are absent from the current research landscape. Such imbalance raises concerns regarding cultural and occupational bias, since stress responses and workplace dynamics differ significantly across countries. The dominance of studies from technologically advanced or research-intensive regions suggests that workplace stress detection remains an emerging focus primarily in countries with strong academic and innovation ecosystems.

Broader international participation would be essential for creating more generalizable and culturally robust stress detection models.

4.2 Research question 1: What machine learning and deep learning techniques are applied for work-place stress detection?

The analysis of the selected papers revealed a diverse range of machine learning and deep learning techniques applied to workplace stress detection tasks. Figure 3 illustrates the distribution of the most frequently employed models across the included studies. Traditional machine learning approaches remain prevalent.

Figure 3

Among the most common techniques, Random Forests were the most frequently used model, appearing in four studies. CNNs and SVMs followed closely, each used in three studies. Other ensemble approaches, including Gradient Boosting, AdaBoost, and Extra Trees, appeared twice each, demonstrating the growing preference for boosting and bagging strategies to improve prediction accuracy and generalization.

Several papers also experimented with hybrid or more advanced deep learning methods. For instance, one study proposed an Ensemble Stacking Model combining Decision Trees, Random Forests, XGBoost, and Multi-Layer Perceptrons (MLP). Recurrent architectures such as bi-directional long short-term Memory and attention mechanism (BiLSTM-AM), recurrent neural network (RNN), and LSTM were also explored for modeling temporal dependencies in physiological signals.

Less frequent, yet notable methods included adaptive neuro fuzzy inference system (ANFIS) and Fuzzy Logic approaches, which leverage rule-based and adaptive systems for stress level classification. Other unique models included Deep Neural Networks (DNN), Autoencoders, and a Von Neumann-inspired Online Multi-Task Learning (OMTL) algorithm, each used in a single study. Classical classifiers such as Naïve Bayes, K-Nearest Neighbors (KNN), Gaussian Discriminant Analysis, Decision Tree, and Decision Jungle were also found in isolated works.

In general, the reviewed literature demonstrates that no single model dominates workplace stress detection research. Instead, the choice of technique often depends on the data modality, such as physiological signals or multi-modal inputs and the study's objective, whether focused on interpretability, accuracy, or real-time applicability.

4.3 Research question 2: What types of data or sensor modalities are used in these studies?

Figure 4 shows the distribution of the types of data used across the studies.

Figure 4

Multimodal and physiological data were the most commonly used types, each appearing in nine studies. It is worth noting that most multimodal datasets also include physiological signals. This reflects a clear trend toward combining physiological measurements with other data sources to improve stress detection accuracy, while also highlighting the central role of physiological signals in this research area.

Physiological data was collected using a variety of sensor modalities, including:

• Heart Rate (HR)

• Heart Rate Variability (HRV)

• Electroencephalography (EEG)

• Blood Volume Pulse (BVP)

• Blood Pressure Levels (BPL)

• Body Temperature (BT)

• Electro-dermal Activity (EDA)

• Inter-beat Interval (IBI)

• Skin Temperature (ST)

• Skin Conductance (SC)

• Skin Conductance Response (SCR)

• Accelerometer (ACC)

• Motion Activity (MA)

These physiological signals allow researchers to capture subtle changes in the autonomic nervous system and other bodily responses that correlate with stress. Multimodal approaches often combine physiological signals with behavioral data, images, text, or other modalities to increase the reliability and robustness of stress detection.

Datasets

Table 3 summarizes the datasets used in the reviewed studies, with references provided where available. For each dataset, the dataset name and its availability, data modality, and corresponding signals and features are specified.

As observed in Table 3, physiological data dominates the literature, both as a single modality (Mathur et al., 2024; Pasha et al., 2024; Jebelli et al., 2019a; Hossain et al., 2025; Jebelli et al., 2019b, 2018; Oskooei et al., 2021; Satapathy et al., 2022; Vadym, 2020) and as a core component in multimodal approaches (Adamopoulos et al., 2025; Rodrigues and Correia, 2024; Mihirette et al., 2025; Nkurikiyeyezu et al., 2019; Nägelin, 2023; Martinez et al., 2018; Meelu and Assres, 2024; Bolliger et al., 2020; Cvetković et al., 2017). Many studies rely on a rich combination of physiological signals indicating a strong preference for capturing stress through internal biological responses rather than external or contextual factors.

Most studies created their own datasets (Rodrigues and Correia, 2024; Jebelli et al., 2019a; Hossain et al., 2025; Jebelli et al., 2019b, 2018; Nägelin, 2023; Martinez et al., 2018; Oskooei et al., 2021; Satapathy et al., 2022; Vadym, 2020; Bolliger et al., 2020; Cvetković et al., 2017), whereas some (Adamopoulos et al., 2025; Mathur et al., 2024; Pasha et al., 2024; Mihirette et al., 2025; Meelu and Assres, 2024; Schneegass et al., 2013; Schmidt et al., 2018) relied on publicly available datasets [Affective Road System Dataset (Haouij et al., 2018), SWELL-KW (Koldijk et al., 2014), EmpathicSchool (Hosseini et al., 2022), WESAD (Schmidt et al., 2018), Video Rating Dataset, data from National Oceanic and Atmospheric Administration (NOAA), National Aeronautics and Space Administration NASA and National Centers for Environmental Information (NCEI-NOAA)] and others used a combination of public and self-created datasets (Nkurikiyeyezu et al., 2019). All reviewed papers disclosed the dataset used for training except for paper (Manikandan et al., 2024).

4.4 Research question 3: What performance levels do these models achieve?

The reviewed studies report varying performance levels for workplace stress detection models. Table 4 summarizes the F1-scores and accuracies achieved by different machine learning and deep learning approaches.

Random Forests and Extra Trees are among the most frequently used classical machine learning models, often achieving high accuracy (between 70 and 85.3 %) but with mixed F1-scores (some of them not even reported). Gradient Boosting and AdaBoost also show strong performance, with F1-scores ranging from 63.1% to 81.0% and accuracy generally above 77%. SVM and KNN demonstrate moderate accuracy, though F1-scores are sometimes not reported. Decision Tree, despite being a classical model, achieved one of the highest performance levels in terms of both accuracy and F1-score, showing that interpretable models can still be very effective. Deep learning models like DNN, LSTM, and BiLSTM-AM achieve competitive results, with F1-scores as high as 95% and accuracies up to 96%–86%. Some ensemble approaches, like the Stacking Model combining decision trees, random forest, XGBoost, and MLP, show high performance, suggesting that combining multiple methods can enhance stress detection. More specialized architectures (e.g., OMTL VonNeumann, Decision Jungle) are less common but still achieve moderate to high accuracy.

4.4.1 Observations

• Models from papers (Manikandan et al., 2024; Oskooei et al., 2021; Satapathy et al., 2022; Meelu and Assres, 2024), weren't include in the table because they didnt report any performance metric results

• Accuracy is the most frequently reported metric across studies, while F1-score is less consistently reported.

• Classical machine learning models tend to be more interpretable and still achieve high accuracy, whereas deep learning approaches offer slightly higher F1-scores, likely due to their ability to capture complex patterns in physiological or multimodal data.

Overall, both classical and deep learning models are effective for workplace stress detection. Table 5, Table 6 illustrate the top three models in terms of F1-score and accuracy, showing the variety of approaches available to achieve strong predictive performance in workplace stress detection.

4.5 Research question 4: What workplace contexts are considered?

The reviewed studies covered a variety of workplace contexts, though not uniformly. As shown in Table 7, several papers focused on specific occupational groups, including software engineers (IT sector), nurses and healthcare professionals (health sector), construction workers (construction sector), firefighters and public health inspectors (public sector), and agriculturists (agriculture sector). These studies emphasize stress detection in professions characterized by high workload, safety-critical tasks, or exposure to unpredictable work conditions.

However, a number of studies (two in total) did not explicitly define a professional context, and this information could not be inferred from the datasets used.

Four additional studies used generalized datasets, that is, datasets that do not target a specific profession. This indicates that while some research targets occupationally unique stressors, many stress-detection models are still developed using context-neutral or undefined workplace settings.

Overall, workplace contexts vary widely, but most studies cluster around healthcare, construction and public service, with a considerable portion lacking explicit contextual definition. This highlights a gap in the literature: many stress-detection approaches are not yet validated in real-world, domain-specific workplace environments.

5 Discussion

In this paper, a systematic literature review following the PRISMA guidelines (Page et al., 2021) was conducted with the aim of identifying the current state of workplace stress detection using AI. A total of 19 papers were selected and analyzed according to the research questions defined in Subsection 3.2.

Unlike prior reviews that address general stress detection, this study highlights several critical gaps that are particularly relevant to real-world occupational settings.

First, the sectors that currently receive the most research attention, particularly high-risk occupations are discussed in Subsection 5.2. This is followed by an examination of the broad variety of other workplace contexts that remain underexplored, as presented in Subsection 5.3.

Subsection 5.4 addresses the lack of workplace-specific datasets, as many studies rely on general stress datasets not tailored to real-world occupational environments. The need for studies with larger and more diverse participant samples is discussed in Subsection 5.5. Finally, Subsection 5.6 highlights the potential for expanding the types of physiological and behavioral variables measured to improve model performance and ecological validity.

5.1 Comparative analysis of the reviewed studies

The reviewed studies differ mainly in the types of data they use, the models they apply, and the performance they report. Some studies rely on physiological signals such as heart rate, heart rate variability, and electrodermal activity. Others combine these with additional data sources such as EEG, behavioral data, or images. In general, studies that use multiple data modalities tend to achieve better results than those relying on a single type of data.

In terms of models, classical machine learning methods such as Decision Trees, Random Forests, boosting methods, and ensemble approaches are most commonly used. Only a smaller number of studies apply deep learning models such as CNNs, LSTMs, BiLSTM-AM, and DNN.

Reported performance varies significantly across studies. Some papers report very high results (up to 97%–99%), while most studies achieve more moderate performance, typically above 70%.

The majority of high-performing models in workplace stress detection are classical machine learning approaches, particularly tree-based ensembles.

Overall, direct comparison between studies is difficult because they use different datasets, methods, and evaluation setups. This limits the ability to identify a single best-performing approach and highlights the need for more standardized evaluation protocols in future research.

Multimodal deep learning approaches for stress detection.

Within the studies included in this review, the use of multimodal deep learning approaches remains relatively limited. Only three papers (Adamopoulos et al., 2025; Jebelli et al., 2019a; Manikandan et al., 2024) employ deep learning techniques, specifically LSTM (Adamopoulos et al., 2025), DNNs (Jebelli et al., 2019a), and CNNs combined with RNNs (Manikandan et al., 2024).

Among these studies, one study (Jebelli et al., 2019a) employing a DNN uses physiological signals as input data, specifically EDA, HR, ST, ACC, IBI, and BVP. In contrast, Adamopoulos et al. (2025) does not rely on physiological signals; instead, it combines survey data, real-time environmental data (e.g., temperature, humidity, and pollutant levels), and historical workplace incident data. For Manikandan et al. (2024), limited details are provided regarding the input data, although a CNN- and RNN-based approach is reported.

The reported performance achieved by these models was promising. The LSTM model trained on surveys and real-time environmental data achieved an accuracy of 87%, while the DNN model trained on physiological signal data achieved an accuracy of 86.6%. Although these results are not the highest reported among the reviewed models, they nevertheless demonstrate promising potential for stress detection in workplace settings using multimodal deep learning approaches.

5.2 Importance of high risk occupations in stress detection

As shown in Subsection Table 7, when excluding studies with undefined professions or generalized populations, approximately 53% of the included papers focus on high-risk workers (Hossain et al., 2025; Jebelli et al., 2018, 2019b; Pasha et al., 2024; Oskooei et al., 2021; Mathur et al., 2024; Jebelli et al., 2019a). High-risk professions tend to show stronger and more consistent physiological stress responses, which results in clearer and more distinguishable stress patterns. While these pronounced patterns don't directly translate to low-risk occupations, they provide valuable insights that can guide feature selection, sensor choices, and modeling strategies. Consequently, knowledge derived from high-risk workers can support the development of models capable of detecting the more subtle and heterogeneous stress responses typically observed in low-risk workplace environments.

5.3 Room for other workplace contexts and professions

From the 19 papers included in this review, only nine distinct professions were represented. Additionally, during the PRISMA filtering process, a large number of papers had to be excluded because their data did not originate from workplace environments. This highlights a substantial gap: many workplace contexts remain unexplored despite their potential to contribute valuable insights into stress detection patterns.

Even within sectors that do appear in the literature, the representation is narrow. For example, while several studies include nurses, other closely related professions such as doctors, surgeons, paramedics, or laboratory technicians are entirely absent. Similar gaps exist in other sectors, where only one role is typically studied despite the presence of diverse occupations that may experience stress differently.

This limited coverage suggests that current research captures only a small fraction of real-world workplace variability. Expanding the range of professions studied would not only improve generalizability but also enable the development of more robust and occupation-specific stress detection models.

5.4 Lack of suitable datasets

A noticeable pattern in some of the analyzed studies is the usage of pre-existing stress detection datasets such as the SWELL dataset (Koldijk et al., 2014) and WESAD (Schmidt et al., 2018). Several papers in our review (Mihirette et al., 2025; Meelu and Assres, 2024; Nkurikiyeyezu et al., 2019) make use of these datasets, despite their limitations in terms of workplace relevance and diversity.

The SWELL dataset, while valuable, suffers from limited variability. It includes only office workers, involves approximately 3 h of recording per participant, and contains data from just 25 individuals. This narrow scope restricts its applicability to broader workplace contexts and reduces the generalizability of models trained on it.

Similarly, the WESAD dataset was designed to detect general affective states in a laboratory setting rather than workplace stress. It contains data from only 15 participants and lacks any real-world professional context. Although WESAD appeared frequently among the papers excluded during the PRISMA screening, it was used in only one of the studies included in this review (Mihirette et al., 2025). This highlights a common issue in the field: many studies claim to investigate workplace stress detection but rely on datasets collected outside workplace environments, which undermines the ecological validity of their findings.

Overall, these limitations demonstrate a clear need for more suitable and robust datasets, ones that capture the variability of stress responses across different workplace environments, professions, and task demands. Without such datasets, the development of accurate and generalizable workplace stress detection models remains significantly constrained.

5.5 Need for more participants in the studies

As shown in Table 8, the majority of the reviewed studies rely on very small participant samples, with many experiments involving fewer than 30 individuals. Several studies—such as Hossain et al. (2025), Jebelli et al. (2019b), and Jebelli et al. (2018)—include only 7–20 participants, which is insufficient for capturing the variability of physiological or behavioral stress patterns across different workers. Not all papers where included in the Table 8, since not all of them mentioned explicitly how many participants they had in their data collection processes.

Only three studies stand out for using larger samples: Satapathy et al. (2022) with 200 participants, Nägelin (2023) with approximately 100 participants, and Oskooei et al. (2021) with 90 participants. These represent the minority. Most other studies, including those that collected their own physiological measurements, remain limited in scale and therefore they could face challenges in model generalization.

Small participant samples introduce several problems. First, they reduce the generalizability of machine learning models. Stress responses are highly individual and influenced by factors such as age, physical condition, personality, and job role; with limited participants, it becomes difficult for models to learn patterns representative of the broader working population. Second, small datasets increase the risk of overfitting, where models perform well on the training data but fail to recognize stress effectively in real-world scenarios. Third, limited sample diversity restricts the ability to capture sector-specific stress patterns, which are particularly important for occupational stress detection.

Overall, these results demonstrate a clear need for future studies to involve larger and more diverse participant groups, ideally across multiple workplace sectors. Increasing sample sizes would improve model robustness, enhance cross-worker generalization, and ultimately support the development of AI systems that are reliable for real-world workplace stress monitoring.

5.6 Expanding measured variables

As seen in Table 3, most of the studies included in this review focus on peripheral physiological signals such as heart HR, HRV, and EDA. However, stress originates in the brain, particularly through activation of regions such as the amygdala, prefrontal cortex, and hypothalamus (Harvard Health Publishing, 2024), suggesting that central nervous system signals could provide more direct and informative indicators of stress.

One promising direction is the use of EEG, which provides real-time information on cognitive load, emotional arousal, and stress-related neural patterns (Acharya et al., 2025). Expanding the set of measured variables to include modalities like EEG offers several potential benefits. First, it can enable models to detect stress earlier, as neural markers precede physiological responses such as sweating or heart-rate changes. Second, EEG can help disambiguate stress from other states with similar peripheral signatures (e.g., physical activity), improving model specificity. Third, combining EEG with existing peripheral sensors could enable multimodal models that better capture the complexity of stress responses across diverse workplace environments.

Some studies have already explored this approach: (Jebelli et al., 2019a,b, 2018) used EEG-based monitoring to capture brain activity related to stress, demonstrating the feasibility and potential advantages of incorporating central nervous system measures in workplace stress detection.

While EEG-based measurement poses challenges, such as intrusiveness, cost, and limited feasibility for certain occupations, recent advances in dry-electrode, wearable, and low-profile EEG devices (Zhang et al., 2025; Xiong et al., 2025) make workplace integration increasingly realistic. Broadening the range of monitored variables, particularly toward brain-level indicators, may therefore be essential for developing more accurate, generalizable, and context-aware workplace stress-detection systems.

5.7 Key challenges in workplace stress detection

Based on the findings of this review, several key challenges in machine learning-based detection of workplace stress using wearable and multimodal data can be identified.

First, the current literature focuses mostly on a limited set of professions, particularly high-risk occupations, leaving many workplace contexts underexplored.

Second, the lack of workplace-specific datasets and the reliance on laboratory or general stress datasets limit the ecological validity of existing models.

Third, many studies are based on small participant samples, which restricts generalizability and increases the risk of overfitting.

Finally, the range of measured variables remains limited, with relatively few studies exploring central nervous system approaches (e.g., EEG). Addressing these challenges is essential for developing robust and scalable stress detection systems suitable for real-world workplace environments.

6 Conclusion

6.1 Key findings

This systematic literature review identified several key findings on the current state of artificial intelligence for workplace stress detection. Across the studies, both classical machine learning methods (such as Random Forests, SVM, Gradient Boosting, and Decision Trees) and deep learning techniques (including CNNs, LSTMs, autoencoders, and DNNs) were widely adopted. Many models achieved promising performance levels, with several reporting accuracies above 90%, demonstrating the feasibility of AI-driven stress detection in controlled and semi-controlled environments. The analysis of data types and sensor modalities revealed that most studies relied on physiological signals such as HR, HRV, EDA, BVP, IBI, and temperature, often collected through wearable sensors. Multimodal and physiological data represented the most common approaches. To enhance robustness, researchers frequently combined multiple physiological signals with behavioral indicators, environmental signals and questionnaire data. However, central nervous system data, particularly EEG, remains underutilized, despite evidence that stress originates in the brain (American Psychological Association, 2024). Only a few studies incorporated EEG, indicating an opportunity for expanding the range of measured variables to improve early detection and model specificity.

Regarding workplace contexts, the findings show a limited representation of professions. High-risk occupations such as healthcare workers, firefighters, construction workers, and public safety personnel appeared frequently, representing over half of the studies with clearly defined professions. While this focus is understandable due to the clearer physiological stress signatures in high-demand environments, it highlights the lack of research on low-risk or office-based occupations, which constitute a large portion of the workforce. Additionally, several studies relied on laboratory conditions or generic datasets such as SWELL and WESAD, which limits ecological validity and raises concerns about the generalizability of existing models to real workplace settings.

A recurring limitation across the reviewed studies is the small number of participants. Many experiments included fewer than 20 subjects, and even widely used datasets contained fewer than 30 participants. This restricts the statistical power, robustness, and ability of models to generalize across populations, job types, and cultural contexts. The geographic distribution of studies further reinforces this limitation, as most contributions originated from a small cluster of countries, leaving significant gaps in global representation.

6.2 Future directions

Overall, the findings underscore the promise of multimodal and physiological machine learning workplace stress detection but also highlight substantial gaps that must be addressed to advance the field. Future research should prioritize:

• Larger and more diverse participant samples,

• Workplace-specific datasets collected in real working environments,

• Broader inclusion of professional sectors, including low-risk occupations,

• Integration of central nervous system measurements, such as EEG, and

• Cross-cultural and cross-context validation to ensure robustness.

By addressing these limitations, future systems can move closer to delivering reliable, real-time stress monitoring tools capable of supporting worker wellbeing, improving productivity, and enabling healthier workplace environments.

References

• 1

  AcharyaS.KhosraviA.CreightonD.AlizadehsaniR.AcharyaU. R. (2025). Neurostressology: a systematic review of EEG-based automated mental stress perspectives. Inf. Fusion124:103368. doi: 10.1016/j.inffus.2025.103368

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  ACM (2025). ACM Digital Library. Available online at: https://dl.acm.org/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 3

  AdamopoulosI.ValamontesA.TsirkasP.DouniasG. (2025). Predicting workplace hazard, stress and burnout among public health inspectors: an AI-driven analysis in the context of climate change. Eur. J. Invest. Health Psychol. Educ. 15:65. doi: 10.3390/ejihpe15050065

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  American Psychological Association (2024). Stress Effects on the Body. Available online at: https://www.apa.org/topics/stress/body (Accessed October 21, 2024).

  • Google Scholar
  • View reference in article

• 5

  BolligerL.LukanJ.LuštrekM.BacquerD. D.ClaysE. (2020). Protocol of the stress at work (straw) project: how to disentangle day-to-day occupational stress among academics based on ema, physiological data, and smartphone sensor and usage data. Int. J. Environ. Res. Public Health17:8835. doi: 10.3390/ijerph17238835

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  CohenB. E.EdmondsonD.KronishI. M. (2015). State of the art review: depression, stress, anxiety, and cardiovascular disease. Am. J. Hypertens. 28, 1295-1302. doi: 10.1093/ajh/hpv047

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  CvetkovičB.GjoreskiM.ŠornJ.FrešerM.BogdaňskiM.JackowskaK.et al. (2017). “Management of physical, mental and environmental stress at the workplace,” in 2017 International Conference on Intelligent Environments (IE) (Seoul), 76–83. doi: 10.1109/IE.2017.20

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  DokiS.SasaharaS.HoriD.OiY.TakahashiT.ShirakiN.et al. (2021). Comparison of predicted psychological distress among workers between artificial intelligence and psychiatrists: a cross-sectional study in Tsukuba science city, Japan. BMJ Open11:e046265. doi: 10.1136/bmjopen-2020-046265

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  Google (2025). Google Scholar. Available online at: https://scholar.google.com/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 10

  HaouijN. E.PoggiJ.-M.Sevestre-GhalilaS.GhoziR.JaïdaneM. (2018). “Affective road system and database to assess driver's attention,” in Proceedings of the 33rd Annual ACM Symposium on Applied Computing, SAC '18 (New York, NY: Association for Computing Machinery), 800–803. doi: 10.1145/3167132.3167395

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  Harvard Health Publishing (2024). Understanding the Stress Response. Available online at: https://www.health.harvard.edu/healthy-aging-and-longevity/understanding-the-stress-response (Accessed November 11, 2025).

  • Google Scholar
  • View reference in article

• 12

  HassardJ.TeohK. R.VisockaiteG.DeweP.CoxT. (2018). The cost of work-related stress to society: a systematic review. J. Occup. Health Psychol. 23:1. doi: 10.1037/ocp0000069

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  HossainT.RahamanM. S.YamashitaK.KonoA.GuoweiM.OyamaS.YonezawaT. (2025). “Quantifying work-stress in surgical intensive care unit nurses: a multimodal activity-aware physiological analysis,” in Proceedings of the 2025 ACM International Symposium on Wearable Computers (ISWC '25), (New York, NY: ACM), 127–133. doi: 10.1145/3715071.3750406

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  HosseiniS.GottumukkalaR.KatragaddaS.BhupatirajuR. T.AshkarZ.BorstC. W.et al. (2022). A multimodal sensor dataset for continuous stress detection of nurses in a hospital. Sci. Data9:255. doi: 10.1038/s41597-022-01361-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  Ieeexplore (2025). IEEE Xplore Digital Library. Available online at: https://ieeexplore.ieee.org/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 16

  JebelliH.HwangS.LeeS. (2018). EEG-based workers' stress recognition at construction sites. Autom. Constr. 93, 315–324. doi: 10.1016/j.autcon.2018.05.027

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  JebelliH.KhaliliM. M.LeeS. (2019a). “Mobile EEG-based workers' stress recognition by applying deep neural network,” in Advances in Informatics and Computing in Civil and Construction Engineering, eds. I. Mutis and T. Hartmann, (Cham: Springer International Publishing), 173–180. doi: 10.1007/978-3-030-00220-6_21

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  JebelliH.Mahdi KhaliliM.LeeS. (2019b). A continuously updated, computationally efficient stress recognition framework using electroencephalogram (EEG) by applying online multitask learning algorithms (omtl). IEEE J. Biomed. Health Inf . 23, 1928–1939. doi: 10.1109/JBHI.2018.2870963

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  KoldijkS.SappelliM.VerberneS.NeerincxM.KraaijW. (2014). “The SWELL knowledge work dataset for stress and user modeling research,” in Proceedings of the 16th ACM International Conference on Multimodal Interaction (ICMI 2014) (Istanbul). doi: 10.1145/2663204.2663257

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  LakshmananL. L. K.LeelasankarK.SubbiyanB. (2025). Stress detection using machine learning and deep learning techniques: a systematic review and meta-analysis. Arch. Comput. Methods Eng. 33, 4183–4215. doi: 10.1007/s11831-025-10429-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  ManikandanN. K.ManivannanD.KavithaM. (2024). “Stress monitoring with computer vision and machine learning for software employees,” in 2024 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS) (Coimbatore: IEEE), 1016–1021. doi: 10.1109/ICSCSS60660.2024.10625637

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  MartinezA.SanchezW.BenitezR.GonzalezY.MejiaM.OrtizJ. (2018). A job stress predictive model evaluation through classifier's algorithms. IEEE Latin Am. Trans.16, 178–185. doi: 10.1109/TLA.2018.8291471

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  MathurA.ShikhaSethia, D. (2024). “Body sensor-based multimodal nurse stress detection using machine learning,” in 2024 16th International Conference on Communication Systems and NETworkS (COMSNETS) (Bengaluru), 67–73. doi: 10.1109/COMSNETS59351.2024.10427006

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  MeeluL.AssresG. M. (2024). “Explaining predicted stress levels in employed individuals,” in Proceedings of the 2024 9th International Conference on Machine Learning Technologies, ICMLT '24 (New York, NY: Association for Computing Machinery), 133–137. doi: 10.1145/3674029.3674051

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  MihiretteS.De la CalE. A.TanQ.SedanoJ. (2025). Cross-contextual stress prediction: simple methodology for comparing features and sample domain adaptation techniques in vital sign analysis. Appl. Intell. 55:420. doi: 10.1007/s10489-025-06277-9

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  NǎgelinM. L. (2023). The Stress in your data: towards the automated detection of work-related stress (Doctoral thesis), ETH Dissertation No. 29790, ETH Zurich, Zürich.

  • Google Scholar
  • View reference in article

• 27

  NandanS.MandalS.GhosalP. (2024). “Stress detection and monitoring: a systematic review,” in 2024 IEEE International Symposium on Smart Electronic Systems (iSES) (New Delhi), 309–314. doi: 10.1109/iSES63344.2024.00072

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  NkurikiyeyezuK.ShojiK.YokokuboA.LopezG. (2019). “Thermal comfort and stress recognition in office environment,” in Proceedings of the 12th International Conference on Health Informatics (HEALTHINF 2019) (Prague: SCITEPRESS – Science and Technology Publications), 256–263. doi: 10.5220/0007368802560263

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  OskooeiA.ChauS. M.WeissJ.SridharA.MartínezM. R.MichelB. (2021). “Destress: deep learning for unsupervised identification of mental stress in firefighters from heart-rate variability (HRV) data,” in Studies in Computational Intelligence (Cham: Springer International Publishing), 93–105. doi: 10.1007/978-3-030-53352-6_9

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  PageM. J.McKenzieJ. E.BossuytP. M.BoutronI.HoffmannT. C.MulrowC. D.et al. (2021). The Prisma 2020 statement: an updated guideline for reporting systematic reviews. BMJ372:n71. doi: 10.1136/bmj.n71

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  PashaS. T.HalderN.BadrulT.SetuJ. H.IslamA.AlamM. Z. (2024). “Physiological signal data-driven workplace stress detection among healthcare professionals using BiLSTM–AM and ensemble stacking models,” in 2024 Advances in Science and Engineering Technology International Conferences (ASET) (Abu Dhabi,: IEEE), 1–10. doi: 10.1109/ASET60340.2024.10708646

  • CrossRef
  • Google Scholar

• 32

  PingeA.GadV.JaisinghaniD.GhoshS.SenS. (2024). Detection and monitoring of stress using wearables: a systematic review. Front. Comput. Sci. 6:1478851. doi: 10.3389/fcomp.2024.1478851

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  Pubmed (2025). Pubmed. Available online at: Available online at: https://pubmed.ncbi.nlm.nih.gov/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 34

  RandallA. K.BodenmannG. (2009). The role of stress on close relationships and marital satisfaction. Clin. Psychol. Rev. 29, 105–115. doi: 10.1016/j.cpr.2008.10.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  RicherR.KochV.AbelL.HauckF.KurzM.RinggoldV.et al. (2024). Machine learning-based detection of acute psychosocial stress from body posture and movements. Sci. Rep. 14:8251. doi: 10.1038/s41598-024-59043-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  RodriguesF.CorreiaH. (2024). Semi-supervised and ensemble learning to predict work-related stress. J. Intell. Inf. Syst. 62, 77–90. doi: 10.1007/s10844-023-00806-z

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  SatapathyS.MishraD.Realyvásquez VargasA. (2022). “Measurement and prediction of mental stress through innovative equipment in agriculture,” in Innovation in Agriculture with IoT and AI (Cham: Springer International Publishing), 53–68. doi: 10.1007/978-3-030-88828-2_4

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  SchmidtP.ReissA.DuerichenR.MarbergerC.Van LaerhovenK. (2018). “Introducing WESAD, a multimodal dataset for wearable stress and affect detection,” in Proceedings of the 20th ACM International Conference on Multimodal Interaction, ICMI '18 (New York, NY: Association for Computing Machinery), 400–408. doi: 10.1145/3242969.3242985

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  SchneegassS.PflegingB.BroyN.SchmidtA.HeinrichF. (2013). “A data set of real world driving to assess driver workload,” in Proceedings of the 5th International Conference on Automotive User Interfaces and Interactive Vehicular Applications (AutomotiveUI '13) (New York, NY: ACM), 150–157. doi: 10.1145/2516540.2516561

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  Sciencedirect (2025). Sciencedirect. Available online at: https://www.sciencedirect.com/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 41

  SinhaR. (2001). How does stress increase risk of drug abuse and relapse?Psychopharmacology158, 343–359. doi: 10.1007/s002130100917

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  Springer (2025). Springerlink. Available online at: https://link.springer.com/ (Accessed October 19, 2025).

  • Google Scholar
  • View reference in article

• 43

  SteptoeA.KivimäkiM. (2013). Stress and cardiovascular disease: an update on current knowledge. Annu. Rev. Public Health34, 337–354. doi: 10.1146/annurev-publhealth-031912-114452

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  TaskasaplidisG.FotiadisD. A.BamidisP. D. (2024). Review of stress detection methods using wearable sensors. IEEE Access12, 38219–38246. doi: 10.1109/ACCESS.2024.3373010

  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  TomiyamaA. J. (2019). Stress and obesity. Annu. Rev. Psychol. 70, 703–718. doi: 10.1146/annurev-psych-010418-102936

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  Turner R. J. and Avison, W. R.. (2003). Status variations in stress exposure: implications for the interpretation of research on race, socioeconomic status, and gender. J. Health Soc. Behav. 44, 488–505. doi: 10.2307/1519795

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  VadymM. (2020). “Exploration of scenario-based simulations for stress benchmarking in Swiss Public Service,” in 2020 7th Swiss Conference on Data Science (SDS) (Luzern), 23–28. doi: 10.1109/SDS49233.2020.00012

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  VosG.TrinhK.SarnyaiZ.Rahimi AzghadiM. (2023). Generalizable machine learning for stress monitoring from wearable devices: a systematic literature review. Int. J. Med. Inform. 173:105026. doi: 10.1016/j.ijmedinf.2023.105026

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  WHO (2023). Stress. Geneva: WHO'S Publications.

  • Google Scholar
  • View reference in article

• 50

  World Health Organization (2022). Mental Health at Work: Policy Brief . Geneva: WHO'S Publications.

  • Google Scholar
  • View reference in article

• 51

  XiongF.FanM.FengY.LiY.YangC.ZhengJ.et al. (2025). Advancements in dry and semi-dry EEG electrodes: design, interface characteristics, and performance evaluation. AIP Adv. 15:040703. doi: 10.1063/5.0228644

  • CrossRef
  • Google Scholar
  • View reference in article

• 52

  YashaswiniY.SinchanaT. G.NikithaB.PrahithaM.ReddyN. V. U. (2025). “Stress detection based on EEG values: a systematic literature review,” in Computing Technologies for Sustainable Development, eds. P. D. Sivakumar, R. Ramachandran, C. Pasupathi, and P. Balakrishnan (Cham: Springer Nature Switzerland), 319–331. doi: 10.1007/978-3-031-82386-2_25

  • CrossRef
  • Google Scholar
  • View reference in article

• 53

  ZhangM.QianB.GaoJ.ZhaoS.CuiY.LuoZ.et al. (2025). Recent advances in portable dry electrode EEG: architecture and applications in brain-computer interfaces. Sensors25:5215. doi: 10.3390/s25165215

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article