Affective Intelligent Systems in Healthcare: A Systematic Review

Abstract

Objectives: To investigate the current state of affective computing in healthcare, focusing on its application contexts, algorithmic trends, and the technical–ethical duality involving data privacy and security. Methods and Results: A systematic review was conducted in two phases (2013–2025) following PRISMA guidelines. A total of 170 peer-reviewed articles were selected from PubMed, IEEE Xplore, Scopus, and Web of Science based on predefined inclusion and exclusion criteria, with the sample restricted to full-text studies in English addressing affective computing in healthcare. No formal risk-of-bias tool was applied due to the computational nature of the studies, and the findings were synthesized descriptively. Discussion: The findings reveal a clear shift from classical machine learning (e.g., SVM, k-NN) toward deep learning and hybrid architectures such as CNN–LSTM and attention-based models for processing complex physiological signals. Recent years have shown a growing interest in multimodal data fusion and privacy-preserving mechanisms such as homomorphic encryption. Evidence remains limited by methodological heterogeneity and inconsistent reporting across studies. A significant gap persists in regulatory compliance, as 57% of recent publications do not adequately address data security or ethical risks associated with sensitive biometric footprints. Conclusions: Although affective computing has reached a certain level of technical maturity, future research must prioritize lightweight, secure, and privacy-by-design architectures to enable ethically aligned and trustworthy deployment in real-world healthcare scenarios.

1. Introduction

Affective computing, a term introduced by Rosalind W. Picard in 1995, is an emerging field of technology that aims to develop systems capable of identifying, interpreting, and responding to human emotional signals, such as sadness, joy, nervousness, or stress [1]. Research in this area focuses on how emotions can be incorporated into models of intelligence, particularly in computational applications and human–computer interactions [2,3,4]. According to Picard, although emotions are not strictly logical, they can significantly influence decision-making processes. Emotions also play a crucial role in the quality of human–computer interaction, directly affecting the effectiveness and intelligence of communication. In this context, emotion recognition can help systems better address human needs.

Affective intelligent systems rely on data from multiple sources, including facial expressions and eye tracking [5,6,7,8,9,10], vocal analysis [11], typing patterns [12], neural signals acquired through brain–computer interface technologies, particularly electroencephalography [13], and social and behavioral data derived from online platforms, where users express emotions, opinions, and affective states through textual and multimodal content [14]. Recent advancements in 2024 and 2025 have shifted the focus from handcrafted feature engineering to end-to-end deep learning architectures [15,16]. These models, particularly attention-based mechanisms and hybrid CNN–LSTM architectures, have significantly enhanced the processing of non-stationary physiological signals [17,18]. Moreover, the integration of contextual, behavioral, and environmental variables has moved beyond simple data augmentation to complex multimodal fusion strategies [15,19,20], which allow for more robust affect inference in uncontrolled in-the-wild scenarios [21]. Furthermore, contemporary architectures have begun to incorporate privacy-preserving mechanisms directly into the sensing pipeline to protect sensitive biometric information [22]. The fundamental premise is that, by understanding human emotions, devices and computational systems can adapt their behaviors and responses, enabling more natural and personalized interactions. While early examples included adaptive games [23], modern implementations increasingly leverage intelligent recommendation and support systems that adapt content and feedback according to users’ inferred affective states [19,24].

The relevance of affective computing has increased as human–computer interaction becomes progressively more complex and pervasive. Over recent years, there has been growing interest in developing robust solutions for stress monitoring and mental health support, particularly through the use of physiological and behavioral data [25,26]. Earlier efforts in this area primarily focused on the preliminary validation of physiological biomarkers and the identification of behavioral patterns, establishing the methodological foundations for the more sophisticated intelligent systems examined in this review [27,28,29]. Emotions are a fundamental component of human intelligence and play a central role in decision-making, social interaction, perception, memory, learning, and creativity [1]. As a result, affective computing has expanded into a wide range of application domains, including healthcare [10], education [30], customer services [12], emotional well-being [31], and transportation systems [7]. In healthcare contexts, affective intelligent systems are particularly relevant, as they can help reduce the stigma associated with psychological support by enabling continuous and real-time emotional assistance through technologies such as wearable devices [8,9,10,32,33]. By allowing systems to recognize and respond to patients’ emotional states through the analysis of biomarkers collected by wearables, such as heart rate and electrodermal activity, it becomes possible to remotely monitor both physical and emotional conditions, thereby extending the reach and effectiveness of telemedicine solutions [9,33]. However, as these systems transition from controlled clinical environments to ubiquitous wearable monitoring, a critical tension emerges between data granularity and user privacy. While the analysis of biomarkers such as electrodermal activity offers substantial diagnostic potential [9], it also exposes highly sensitive biometric information. This challenge highlights a growing gap in regulatory compliance and reinforces the need for privacy-by-design frameworks capable of securing affective data without compromising real-time processing efficiency. Consequently, a systematic evaluation of how current datasets and system architectures balance data richness with security remains essential.

Given the potential impact of affective computing on everyday life and on how individuals interact with technology, it is essential to understand not only its applications but also the social and technical challenges associated with its development [1,2,3]. The main objective of this study is to investigate the current landscape of affective computing through a systematic analysis of research published between 2013 and 2025. By examining studies across this extended time span, the review captures the evolution of sensing technologies, machine learning methods, and application domains, revealing how the field has progressively emphasized the technical–ethical duality of modern affective systems, particularly in healthcare and mental health contexts [4,34,35]. Beyond identifying application domains, this work critically examines the trade-offs between algorithmic performance and data protection, analyzing how the literature has addressed, or neglected, privacy-preserving mechanisms such as homomorphic encryption and federated learning. While ethical considerations are frequently acknowledged at a declarative level, concrete technical implementations remain relatively scarce throughout the analyzed period, including in more recent studies [36,37]. Nevertheless, the most recent contributions provide early evidence of a gradual incorporation of privacy-aware architectures, including encrypted processing pipelines and secure multimodal frameworks [19,22]. Emotional state identification can be used proactively to support treatment recommendations in specific domains [1,21]. In this context, the present review also analyzes the multimodal datasets reported in the literature, the artificial intelligence techniques employed, and the training and testing strategies adopted by emotion recognition systems [3,4,38]. The overarching goal is to provide a comprehensive overview of affective computing research over the past decade, highlighting established application areas while identifying persistent challenges and research gaps that should be addressed by future studies.

The main contributions of this study are summarized as follows:

• This work provides a systematic literature review of 170 articles published between 2013 and 2025, incorporating significant research from 2024 and 2025 to reflect recent technological shifts.
• We identify and categorize the transition from classical machine learning to advanced deep learning architectures, such as CNN–LSTM and attention-based mechanisms, specifically for physiological signal processing.
• The study offers a critical evaluation of data protection strategies, identifying a significant gap in which 57% of recent studies do not explicitly address regulatory compliance or encryption methods.
• We highlight the emergence of privacy-by-design frameworks, such as homomorphic encryption in affective pipelines, and multimodal fusion strategies as a new frontier for trustworthy intelligent systems.
• This research identifies practical constraints to inform the design of future affective systems, ensuring that next-generation wearables meet stringent security standards and data integrity requirements.
• Furthermore, we observed a declining trend in occupational stress studies, revealing that individual predispositions, as well as environmental and emotional determinants of distress, remain critically under-researched.

This study is structured into five sections. Following this introduction, Section 2 describes the systematic literature review protocol adapted from PRISMA guidelines, including the selection of 170 papers based on predefined criteria. Section 3 synthesizes the evidence to answer the research questions. Section 4 provides a critical discussion on algorithms, data privacy, and personalization in stress-related research. The paper concludes in Section 5 with final remarks and implications for the field. Bibliographic references follow the final section.

2. Methods

The study was conducted through a systematic review. The review was adapted from the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 (PRISMA 2020) [39]. Based on the scientific literature, articles addressing applications of affective computing in different fields of knowledge were identified. One of the questions guiding the qualitative analysis concerned the investigation of the social impacts of this topic and the benefits that affective intelligent systems may provide to their users. Additionally, this work explores the different artificial intelligence methods adopted for emotion recognition and related applications. The datasets reported in the selected studies were also analyzed. Finally, given the sensitive nature of the data used by this type of technology, this review investigates the techniques applied to ensure data security and privacy in the selected studies.

The results obtained address the following five research questions (RQ):

• (RQ1)—What are the main focus areas of the studies identified in this review?
• (RQ2)—What are the applications and social impacts of affective intelligent systems?
• (RQ3)—How are security and privacy addressed in the selected studies?
• (RQ4)—Which datasets were used, and what types of features were considered?
• (RQ5)—Which artificial intelligence techniques were employed, and which evaluation metrics were used?

2.1. Selection Criteria

IEEE Xplore, Scopus, Web of Science, and PubMed were systematically searched for English peer-reviewed articles. After removing duplicates, two independent reviewers screened the records by title, abstract, and keywords using the Rayyan web-based platform (Rayyan Systems Inc., Cambridge, MA, USA) [40], applying predefined inclusion and exclusion criteria (

Table 1. Inclusion and exclusion criteria for article selection.

Inclusion Criteria	Exclusion Criteria
Articles addressing affective computing	Duplicate articles
Articles written in English	Articles in preprint or early access format
Articles contributing to the research questions	Systematic reviews or systematic mapping studies
Articles available in full text	Articles with inaccessible full text
Peer-reviewed articles	Articles outside the scope of the study

). Disagreements were resolved through discussion or escalated to a third reviewer, with the project coordinator serving as the final arbiter when necessary. Study quality was assessed based on peer-review status and eligibility criteria rather than by post hoc scoring.

2.2. Selection Process

The systematic study selection process was conducted in two phases. The initial search, carried out in March 2024, identified 108 eligible articles. An updated search performed in December 2025 added 63 additional studies published between 2024 and 2025. The second search phase was conducted as a structured update of the original systematic search, using the same databases, search strings, and eligibility criteria, with the only modification being the extension of the publication period. In total, 170 articles met the eligibility criteria, comprising 124 conference papers and 46 journal articles. Conference papers were intentionally included because affective computing is a relatively recent and rapidly evolving research field within computer science and artificial intelligence, in which many relevant methodological advances and state-of-the-art contributions are first disseminated through conference venues. In this context, conference publications often report novel algorithms, architectures, datasets, and experimental results that precede extended journal versions, making them particularly important for identifying emerging trends and current directions in the field. During the initial screening stage, 96 duplicate records were removed. Of the remaining 350 records, 137 were excluded for not meeting the predefined eligibility criteria, resulting in 213 articles selected for full-text assessment. The Rayyan platform was used with blinded independent review to support duplicate removal and ensure a two-reviewer screening process [40]. Among the articles assessed in full text, 5 could not be retrieved due to access limitations, and 37 were excluded because they did not provide sufficient information to address the research questions. The final selection therefore comprised 170 studies aligned with the objectives of this review. The complete study selection process, including both search phases, is summarized in the PRISMA flow diagram shown in Figure 1 [41].

The search strategy was developed based on the research questions and the PCC framework: Population, Concept, and Context [42], defined as follows:

• P—Affective Intelligent Systems in Healthcare.
• C—Emotion recognition, stress, and emotional states.
• C—Application context, including domains, techniques, and aspects related to social impact and security.

To ensure the reproducibility of this systematic review, a structured search strategy was developed and applied consistently across all databases. The search string was organized into three descriptor groups combined with Boolean operators, covering (i) affective computing and artificial emotional intelligence, (ii) healthcare-related terms, and (iii) the target concepts of emotion recognition and stress. The complete search strings used in each database are presented in

Table 2. Search strings used in each database.

Database	Search String
Scopus, Web of Science, IEEE Xplore	(“Affective Computing” OR “Emotion AI” OR “Emotional Technology” OR “artificial emotional intelligence” OR “machine learning”) AND (“Healthcare” OR “Health Care” OR “Health System” OR “Health Problems” OR “Health Problem” OR “Health Insurance” OR “Medical Care” OR “Health” OR “Treatment” OR “Diagnosis” OR “Medicine”) AND (“Emotion recognition” AND “Stress”)
PubMed	(“Affective Computing” OR “Emotion AI” OR “Emotional Technology” OR “artificial emotional intelligence” OR “machine learning”) AND (“Emotion recognition” AND “Stress”)

. For Scopus, Web of Science, and IEEE Xplore, the full three-part query was applied without modification. In PubMed, however, the second descriptor group (healthcare-related terms) was removed. This adjustment was necessary because PubMed relies on MeSH-controlled vocabulary and a different indexing structure, which caused the full query to return incomplete or inconsistent results. By adapting the query while preserving the core conceptual components, we ensured both compatibility with the database and methodological consistency across sources.

This approach allows other researchers to replicate the search process precisely, using the same descriptor groups and Boolean logic detailed in Table 2, while respecting the technical constraints of each database. The review was not registered because, as a computer science protocol, it is not eligible for clinical or health-related registries such as PROSPERO.

2.3. Data Items and Outcome Measures

For each included study, data were extracted according to predefined outcome domains. The primary outcomes were: (1) the proposed affective computing approach, (2) the computational models or algorithms employed, and (3) the evaluation strategies and performance metrics reported. All results aligned with these domains were collected. When multiple results were reported within the same domain, the most comprehensive and clearly described data were selected.

Additional contextual variables were extracted, including publication type, year of publication, application domain, dataset characteristics, experimental design, and the availability of implementation or reproducibility information. Data addressing the research questions (RQ1 to RQ5) were also recorded.

The synthesis was based on descriptive and qualitative comparisons of computational methods, datasets, and reported performance metrics. No formal risk-of-bias or quality assessment tool was applied. In addition to peer-review status and eligibility screening, computational quality indicators were extracted whenever reported, including dataset size, validation protocol (e.g., cross-validation, LOSO, held-out test), and reporting practices that may indicate potential overfitting. These indicators were synthesized descriptively due to heterogeneity and incomplete reporting across studies. In parallel, a qualitative appraisal of methodological clarity and rigor was conducted during the screening and full-text assessment stages, based solely on the information explicitly reported in each article, without applying any formal scoring or weighting scheme. Study quality was therefore addressed through the study selection process, as described in Section 2.1 (Selection Criteria) and Section 2.2 (Selection Process). This process relied on the exclusive inclusion of peer-reviewed articles indexed in curated scientific databases and on strict adherence to predefined inclusion and exclusion criteria. Studies that lacked sufficient methodological detail or did not contribute directly to the research questions were excluded during screening. Independent screening by multiple reviewers and full-text assessment, when required, further supported the rigor of the review.

2.4. Selected Articles

The systematic selection process was conducted in two phases: an initial search in March 2024, which identified 108 articles, followed by an update in December 2025 that added 63 additional studies published between 2024 and 2025. In total, 170 articles met the eligibility criteria, comprising 124 conference papers and 46 journal articles. Conference papers were included because affective computing is a relatively recent and rapidly evolving research field, in which many relevant contributions are first disseminated through conference venues.

2.5. Geographic and Temporal Distribution of the Included Studies

The studies included in this review represent contributions from 31 countries, reflecting the global interest in affective computing applied to mental health. India was the most prolific contributor, with 59 publications, followed by the United States with 18 and China with 12. Sri Lanka contributed 8 studies, and the United Kingdom accounted for 7. Germany, Pakistan, and Canada each contributed 5 publications, while Indonesia, Russia, and Saudi Arabia contributed 4 each. Brazil, France, the Netherlands, Spain, Thailand, the United Arab Emirates, and South Korea each contributed 3 studies. Algeria, Italy, Lebanon, Malaysia, Romania, Japan, and the Philippines contributed 2 articles each. Greece, Kenya, Norway, Switzerland, Poland, and Mexico each contributed one study. The geographic distribution of the included articles is illustrated in Figure 2.

Although the distribution shows a concentration of publications in specific countries, geographic origin is treated in this review as a descriptive attribute. The analysis focuses on computational methods, datasets, and evaluation strategies and does not assume that the country of origin affects methodological quality or validity.

Similarly, differences between conference papers and journal articles are considered solely as descriptive characteristics. Both publication types were assigned equal weight in the analysis, and no hierarchical distinction was made regarding their scientific contribution or methodological robustness.

This decision is methodologically justified because, in computer science and artificial intelligence research, high-impact peer-reviewed conferences constitute primary dissemination venues for novel computational methods. Many influential contributions in affective computing are first published in conference proceedings, often under rigorous review standards comparable to journals. Therefore, assigning equal analytical weight to conference and journal publications ensures a balanced and field-appropriate synthesis of the literature.

The selected studies were published between 2013 and 2025. A clear increase in publication volume was observed after 2023. The period from 2023 to 2025 accounts for 84 conference papers and 24 journal articles, corresponding to 66% of the total sample. This trend indicates a recent and growing interest in affective computing research. The temporal distribution of publications is shown in Figure 3.

3. Results

The results presented in this section detail the categorical distributions and data extracted from the 170 selected articles, focusing on the research questions defined in Section 2.

3.1. RQ1—What Are the Main Focus Areas of the Studies Identified in This Review?

The systematic mapping of the 170 selected studies reveals clear trends in affective computing applications within the health domain over the review period (2013–2025). Research has been predominantly focused on stress and emotion recognition, which together account for 65% of the analyzed studies. At the same time, the literature shows a gradual expansion toward broader mental health monitoring scenarios and the modeling of more complex affective constructs, such as emotional states represented by arousal and valence dimensions. Across the reviewed period, this evolution is accompanied by an increasing use of multimodal data fusion strategies, for example, combining EEG with heart rate variability and other physiological signals, aiming to improve robustness and classification performance. Figure 4 summarizes the overall distribution of application domains identified in the included studies.

Regarding RQ1, the distribution of studies confirms that Stress and Depression remain the most investigated affective states within the period 2013–2023. A wide diversity of applications for affective intelligent systems in healthcare was identified, exhibiting a clear predominance of research focused on mental health monitoring. These applications were categorized into six primary domains: Stress, Emotions, Emotional States (specifically those modeling arousal and valence), Depression, Anxiety, and an Other category, which groups studies addressing specific conditions such as Post-Traumatic Stress Disorder (PTSD) and general well-being. Figure 4 illustrates the quantitative distribution of these categories, highlighting the thematic concentration of the analyzed studies. Some studies addressed more than one category, as reflected in the appendix table. Among the 108 studies published between 2013 and 2023, 124 category descriptors were identified. A total of 91 studies were classified into a single category, while 15 were assigned to two or more categories. Specifically, studies such as [43,44,45,46] were classified under Stress and Emotions; [26,47] under Emotions and Depression; [12,21,31] under Emotions and Other; [48] under Stress and Emotional States; [49] under Emotions and Emotional States; and [50] under Depression and Other. Two studies [10,51] were classified under Stress, Depression, and Anxiety, while one study [52] included four descriptors: Stress, Emotions, Depression, and Anxiety. As detailed in

Table 3. This table organizes the references used in the analysis of the application areas discussed between 2013 to 2023. Each study was classified into one or more categories according to its primary approach.

References	Application Area
[2,8,9,10,21,33,43,44,45,46,48,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99]	Stress
[3,4,6,12,15,19,20,22,24,25,26,30,32,34,36,43,44,45,46,47,49,52,78,98,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140]	Emotions
[7,8,20,48,97,98,100,111,127,135,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159]	Emotional States (arousal and valence)
[4,10,26,35,47,50,51,52,160,161,162,163,164]	Depression
[10,51,52,165,166]	Anxiety
[11,12,19,21,23,31,37,38,50,129,132,139,167,168,169,170,171,172,173,174,175]	Other

, these studies are categorized across the different research areas identified during this systematic review.

3.1.1. Emotions and Emotional States

Emotion recognition and emotional state modeling represent distinct analytical approaches in affective computing. Emotion recognition is based on discrete classification, assigning labels such as happiness, sadness, anger, or fear based on observable cues. These models prioritize interpretability and are commonly used in interaction-oriented systems. Emotional state modeling follows dimensional affect theories, in which affect is represented along continuous axes, most often valence and arousal. These models estimate activation levels and affective intensity over time, supporting longitudinal psychological monitoring and mental health assessment. This conceptual distinction is well established in affective computing research and underpins many physiological and multimodal affect modeling studies [1,110,141,144]. Studies classified under Emotions and Emotional States constitute a major portion of the analyzed literature, reflecting the central role of affect analysis in healthcare-oriented intelligent systems. Together, these categories capture complementary perspectives on human affect, ranging from categorical emotion recognition to continuous affective state estimation.

The Emotions category includes studies focused on the recognition of discrete affective classes. Most approaches rely on facial expression analysis, speech processing, text-based sentiment analysis, electroencephalography, or multimodal physiological sensing. Several studies demonstrate the use of facial and speech features for emotion recognition in human–computer interaction and mental health contexts [12,101,121]. Others employ EEG and physiological signals to support emotion classification and monitoring [43,44,111,124].

More recent work in this category emphasizes system robustness and deployment in real-world settings. These studies introduce multimodal fusion architectures, physiological emotion datasets, and secure or privacy-aware pipelines. Examples include multimodal emotion recognition systems combining physiological and behavioral data [15,20], encrypted emotion recognition frameworks [22], and EEG- or speech-based emotion classification models applied to clinical, educational, and daily-life scenarios, including tweet messages [129,133,134,159].

In contrast, studies categorized under Emotional States focus on modeling affect as a continuous process. These works commonly adopt arousal and valence representations to capture affective intensity and activation. Early studies relied on physiological signals to estimate affective dimensions in real time [100,143,145]. Subsequent research expanded these approaches by integrating EEG, ECG, heart rate variability, and electrodermal activity to infer arousal–valence dynamics [120,141,144,146,152].

More recent studies extend dimensional affect modeling through multimodal and context-aware sensing strategies. Examples include EEG-based modeling of perceived stress and affect [48,111], hybrid architectures combining physiological and behavioral signals [7,49], and applications targeting mood and affective intensity in healthcare and daily-life contexts [142,151]. These studies illustrate the suitability of dimensional representations for capturing affective variations that are not adequately described by discrete emotion labels.

Overall, the combined analysis of Emotions and Emotional States highlights methodological complementarity. Discrete emotion recognition remains prevalent due to its clarity and applicability in interaction-driven systems. Dimensional affect modeling is increasingly adopted for continuous monitoring and fine-grained assessment of psychological states. This shift reflects the growing demand for affective computing systems that represent emotional processes as dynamic and continuous phenomena in healthcare applications.

3.1.2. Stress, Depression, and Anxiety

Stress, depression, and anxiety constitute central targets of healthcare-oriented affective computing. Studies in these categories focus on mental health monitoring, risk detection, and continuous assessment, often relying on physiological sensing and machine learning-based inference.

Stress studies address the detection of physiological responses to stressors and the monitoring of stress levels over time. Most approaches use wearable or ambient sensors to capture electrodermal activity, heart rate variability, electroencephalography, and skin temperature. Several systems are designed for real-time operation in educational, occupational, or daily-life contexts [9,75,76,79,80,81,82]. A recurring limitation concerns model generalization. Many studies rely on population-level patterns and do not incorporate subject-specific baselines, which limits personalization and robustness across users and contexts.

Depression studies aim to identify depressive symptoms or risk markers using behavioral and physiological data. The reviewed works employ speech, text, facial cues, and physiological measures to support early screening and continuous monitoring [4,10,26,35,47,50,51,52,160,161,162,163,164]. These systems are commonly framed as complementary tools to clinical assessment and emphasize non-invasive data acquisition. Methodological heterogeneity remains high, thereby limiting cross-study comparability.

Anxiety studies are less frequent but address a distinct clinical construct. Whereas stress detection focuses on physiological reactivity to stressors, anxiety monitoring relates to persistent affective symptoms and therefore requires different validation settings. The reviewed anxiety studies apply sensor-based or IoT-based solutions and use physiological and behavioral parameters to infer anxiety-related states [10,51,52,165,166]. The limited number of studies and the absence of standardized evaluation protocols remain key challenges.

Occupational stress emerges as a recurrent application context; however, it remains underrepresented within the broader stress literature.

Table 4. Occupational stress studies and recurrent methodological limitations identified in the reviewed literature.

Context	Main Methodological Limitation	References
Occupational stress	Most studies rely on generalized models trained on population-level data and do not incorporate subject-specific baselines or individualized calibration. This limitation constrains personalization, reduces robustness across users, and weakens applicability in real-world workplace environments.	[3,7,11,36,44,51,57,70,73,86,94,99,101,107,110,117,120,124,168]

summarizes occupational stress studies and highlights a shared methodological limitation. Most models do not incorporate individual characteristics or subject-specific calibration, relying instead on generalized benchmarks. Only a small subset of studies focuses explicitly on workplace environments and job-related stressors [86,94,99]. Many recent works evaluate stress in general daily-life settings or in specific populations, such as students, social-interaction scenarios, children undergoing physiotherapy, or patients requiring medical support [79,80,87,88,96]. This pattern indicates that stress monitoring research remains dominated by generalized models rather than context-sensitive and individualized occupational approaches.

3.1.3. Other

The Other category groups studies that do not align directly with the primary affective constructs analyzed in this review, such as stress, emotions, emotional states, depression, or anxiety. These works typically address specific psychological conditions, peripheral affective phenomena, or system-level approaches that fall outside the dominant analytical categories.

A subset of studies in this category focuses on post-traumatic stress disorder (PTSD), exploring the use of physiological signals to monitor intense or trauma-related emotional responses [37,38]. These approaches target vulnerable populations, including individuals exposed to high-risk or high-pressure environments, and emphasize continuous affective monitoring rather than discrete emotion recognition.

Other studies address less frequently investigated affective or psychological phenomena, such as anger detection via electromyography [167], distress monitoring [23], negative social interactions identified through acoustic analysis [11], suicide risk assessment [169], and general well-being monitoring [168]. Several contributions also fall into this category because they present affective computing frameworks, support platforms, or methodological systems without a specific emotional or clinical target.

Although the Other category represents a smaller proportion of the analyzed literature, these studies illustrate the breadth of affective computing applications in healthcare. They highlight emerging research directions and complementary use cases that extend beyond the dominant focus on stress and mental health monitoring.

3.2. RQ2—What Are the Applications and Social Impacts of Affective Intelligent Systems?

Across the 170 included studies, affective intelligent systems are applied in mental health, education, workplace settings, and daily-life monitoring. These applications aim to support continuous assessment, early risk detection, and adaptive human–computer interaction. However, many studies emphasize algorithm validation, which limits the discussion of practical impact and user-facing benefits.

3.2.1. Emotion Detection and Practical Applications

Of the 170 analyzed studies, 72 (42%) focused primarily on validating algorithms for emotion or stress detection. These studies use physiological signals, facial expressions, speech, and text to infer affective information. Several works apply facial and speech cues to emotion-aware interaction scenarios [12,121,133]. Other studies use EEG and multimodal physiological sensing to classify affective responses in controlled and semi-natural settings [43,44,124,134].

A smaller subset integrates affect recognition into application scenarios with explicit user-facing goals. Examples include multimodal affect recognition architectures for daily-life monitoring [15,20] and rehabilitation-oriented systems that incorporate automated facial emotion recognition [16]. Recent work also explores encrypted emotion recognition pipelines, which support affective inference under privacy constraints [22].

3.2.2. Mental Health and Well-Being

Mental health monitoring constitutes one of the most prominent application domains in the reviewed literature. Early studies proposed emotion-aware assistants and stress management tools designed to support mental well-being through affect recognition and user interaction [52,69]. These systems demonstrated the feasibility of combining affective computing with conversational interfaces and decision-support mechanisms to assist users outside traditional clinical settings.

A substantial body of work investigates speech-based stress detection and adaptive support mechanisms, highlighting the relevance of vocal features as non-invasive indicators of psychological states [83]. These approaches are particularly attractive for mental health applications because they reduce the need for specialized hardware and enable continuous monitoring through everyday devices.

Wearable-based monitoring platforms are also widely adopted, enabling continuous physiological sensing in naturalistic environments [9,176]. By leveraging signals such as heart rate variability and electrodermal activity, these systems support longitudinal observation of stress and emotional fluctuations. This shift toward continuous sensing expands the potential reach of mental health monitoring beyond episodic assessments and controlled clinical evaluations.

More recent studies propose integrated monitoring platforms that combine emotion recognition with conversational agents or multimodal interfaces to provide personalized mental health support [19,92]. These systems reflect a growing interest in holistic and user-centered solutions, in which affective data are used not only for detection but also to guide adaptive feedback and intervention strategies.

Overall, these applications indicate a gradual transition from feasibility-driven studies toward systems that aim for continuous monitoring and real-world deployment. However, many contributions still lack explicit evaluation of long-term outcomes, clinical effectiveness, or measurable user benefits. As a result, although affective computing shows strong potential for mental health support, its social impact remains constrained by limited validation in real-world care scenarios and by the absence of standardized assessment protocols.

3.2.3. Education, Work, and High-Performance Scenarios

Applications of affective computing in education and work environments aim to assess stress, cognitive load, and affective states in performance-critical contexts. In educational settings, early studies proposed affect-aware platforms to support learning processes, monitor student engagement, and provide adaptive feedback based on emotional responses [59]. These systems demonstrated the feasibility of integrating physiological and behavioral sensing into learning environments to improve personalization and emotional support.

In workplace contexts, initial research explored mood and stress recognition to support employee well-being and organizational awareness. Examples include smartphone- and wearable-based monitoring platforms designed to capture affective states during routine professional activities [120]. These approaches anticipated many of the challenges later observed in occupational stress research, particularly regarding contextual interpretation and individual variability.

High-performance scenarios represent a distinct application domain, in which affective monitoring is applied to competitive or high-demand activities. Studies in this area investigated affect recognition in contexts such as e-sports and performance analysis, demonstrating how emotional states relate to teamwork, decision-making, and performance outcomes [106]. These settings emphasize the sensitivity of affective responses under pressure and reinforce the need for robust and low-latency monitoring solutions.

More recent research extends these applications through stress-aware learning systems and academic stress analysis. Studies focusing on students integrate affect recognition with intelligent tutoring systems and academic analytics to monitor stress and emotional well-being during learning activities [87,154]. These works reflect a growing interest in continuous affect monitoring beyond controlled laboratory conditions.

In occupational scenarios, several studies explicitly address workplace stress and professional well-being, as identified in the RQ1 analysis. These include systems designed for socio-economic stress detection, employee mental health support, and stress monitoring in industrial or professional settings [86,94,99]. While these contributions highlight the relevance of affective monitoring for safety, productivity, and well-being, they largely rely on generalized stress models and limited contextual variables. This limitation mirrors the broader gap identified in RQ1, in which the lack of individualized and context-aware modeling restricts the real-world applicability of occupational stress monitoring systems.

3.2.4. Integration of IoT and Affective Computing

The integration of Internet of Things technologies has been instrumental in shifting affective computing from laboratory-based experiments to continuous monitoring in real-world environments. Early IoT-oriented frameworks demonstrated the feasibility of distributed sensing and real-time affect inference using embedded and networked devices, establishing the foundations for pervasive affective systems [74,126]. These studies showed that physiological signals could be captured outside controlled settings, enabling affect monitoring in everyday contexts.

Subsequent research emphasized wearable-based stress monitoring and multimodal sensing pipelines, combining physiological signals such as electrodermal activity, heart rate variability, and motion data [71,75,76]. These approaches extended affective computing toward long-term observation and ecological validity, supporting applications in daily life, educational environments, and occupational contexts.

More recent systems focus on integrating multiple physiological streams to improve robustness under real-world noise and variability [43,44,81]. These studies reflect a transition from proof-of-concept implementations to deployable architectures operating on resource-constrained devices. However, this transition introduces structural trade-offs. Real-time inference on wearables requires balancing model complexity, energy consumption, and communication overhead, thereby constraining the adoption of computationally intensive architectures.

Closed-loop IoT-based systems further extend this paradigm by coupling affect detection with automated intervention mechanisms. Examples include affective brain–computer interface models designed for stress intervention, in which physiological sensing directly informs adaptive system responses [95]. These systems move affective computing from passive monitoring toward active support, reinforcing its potential social impact in healthcare and mental well-being applications.

Despite these advances, IoT-based affective architectures also amplify challenges related to computation, energy efficiency, and data privacy. Continuous sensing increases the exposure of sensitive biometric data, while resource limitations restrict the use of advanced security mechanisms. As a result, IoT integration expands the feasibility and social relevance of long-term affect monitoring, while simultaneously reinforcing the need for lightweight, secure, and privacy-aware system designs [58,177].

3.3. RQ3—How Are Security and Privacy Addressed in the Studies?

The analysis of the selected studies reveals a wide range of approaches for addressing security and privacy concerns in affective intelligent systems. However, the depth, consistency, and technical maturity of these approaches vary substantially across the reviewed literature, exposing persistent gaps throughout the entire period analyzed (2013–2025). For the RQ3 analysis, studies were coded as did not address when no security or privacy measure was mentioned or described in the text. When a study explicitly stated that privacy protection was unnecessary or minimized privacy risks, it was recorded separately as an explicit dismissal case.

Considering the complete set of 170 studies reviewed, a substantial majority ($n=121$) did not explicitly mention any security or privacy measures. This result indicates that, despite the inherently sensitive nature of affective, physiological, and biometric data, protection mechanisms remain largely peripheral in affective computing research. Figure 5 summarizes the distribution of strategies adopted by the subset of studies that did address these concerns.

Among the studies that reported some form of protection strategy, references to ethical guidelines constitute the most frequently adopted approach ($n=43$), as detailed in Table 5. These references typically include approval by ethics committees, adherence to institutional review protocols, or general ethical principles governing research involving human participants [3,34]. In contrast, concrete technical mechanisms remain considerably less common. Only a small number of studies explicitly reported anonymization procedures ($n=4$) [3,112,145,161], encryption mechanisms ($n=3$) [10,30,45], explicit informed consent handling ($n=4$) [2,3,112,161], or blockchain-based solutions ($n=1$) [52]. This distribution suggests that ethical awareness is present in the literature but is rarely operationalized into enforceable, system-level security or privacy mechanisms.

Across the entire review period, this pattern remains largely unchanged. Although the number of publications has increased and application domains have diversified, particularly in stress detection and mental health monitoring, security and privacy considerations continue to lag behind algorithmic development. This trend persists in more recent studies, where 57% of publications still fail to incorporate any explicit data protection strategy. In most cases, this corresponds to not reported (i.e., no privacy or security measure described), rather than not implemented; explicit dismissals were rare and recorded separately (e.g., [12]).

Nevertheless, only a very limited number of studies point toward more robust protection strategies in affective computing systems. A notable example is the use of real-time encrypted emotion recognition pipelines based on homomorphic encryption, which enable inference directly over encrypted data without exposing raw biometric signals [22]. Despite representing a significant technical advance, such approaches remain isolated cases rather than established practice within the field.

The limited adoption of concrete security and privacy mechanisms is particularly concerning in light of the current regulatory landscape. The European Union Artificial Intelligence Act [178] and Brazil’s General Data Protection Law [179] classify biometric and affective data as high-risk or sensitive, imposing requirements for transparency, accountability, and robust data protection. The absence of documented security protocols in a large proportion of the reviewed studies therefore stands in direct tension with these regulatory obligations.

Finally, the findings indicate that security and privacy considerations have not progressed at the same pace as sensing technologies and machine learning methods in affective computing. Across the entire review period, protection mechanisms are frequently subordinated to performance optimization and application feasibility. This gap highlights the need for future research to integrate privacy and security strategies from the earliest stages of system design, ensuring that affective intelligent systems are not only technically effective but also ethically and legally compliant.

Table 5. Summary of identified security and privacy strategies in the reviewed literature.

References	Security and Privacy Strategies
[2,3,4,8,9,21,24,25,26,31,34,35,36,37,38,46,49,50,60,62,68,70,71,72,73,74,76,77,78,111,112,116,120,123,126,127,145,161,162,168,169,180,181]	Ethical guidelines
[3,21,36,37,60]	Privacy protected (no technical details reported)
[10,60,112,145,157]	Anonymization
[2,3,112,161]	Informed consent
[10,45,52]	Encryption
[9]	Local processing
[52]	Blockchain

Table 5. Summary of identified security and privacy strategies in the reviewed literature.

References	Security and Privacy Strategies
[2,3,4,8,9,21,24,25,26,31,34,35,36,37,38,46,49,50,60,62,68,70,71,72,73,74,76,77,78,111,112,116,120,123,126,127,145,161,162,168,169,180,181]	Ethical guidelines
[3,21,36,37,60]	Privacy protected (no technical details reported)
[10,60,112,145,157]	Anonymization
[2,3,112,161]	Informed consent
[10,45,52]	Encryption
[9]	Local processing
[52]	Blockchain

3.3.1. Adopted Strategies

The reviewed studies presented different technical and ethical approaches to mitigating privacy and security risks associated with sensitive biometric and physiological data. Table 5 summarizes only studies that explicitly mentioned at least one security or privacy-related aspect.

Data Encryption: A limited number of studies explicitly reported the use of encryption mechanisms to protect data integrity [10,45,52]. A notable recent advancement is the exploration of homomorphic encryption for emotion recognition, which enables computation directly over encrypted data without decrypting the source information [22]. However, most studies that mention encryption do not provide details regarding computational overhead, latency, or scalability, which remain critical challenges for real-time affective computing systems.

Anonymization and De-identification: Anonymization was occasionally reported in the reviewed literature [10,60,112,145]. In most cases, anonymization was mentioned at a declarative level, without specifying the techniques employed, such as k-anonymity or differential privacy. This lack of methodological detail limits reproducibility and prevents a rigorous assessment of re-identification risks.

Privacy Statements without Technical Detail: Several studies stated that privacy was protected but did not describe concrete technical or procedural mechanisms to support this claim [3,21,36,37,60]. These statements were therefore classified separately, as they indicate ethical awareness but do not provide sufficient information to evaluate actual data protection practices.

Blockchain and Decentralization: Blockchain-based approaches were explored in isolated cases [52]. While these solutions align with transparency and accountability principles present in data protection regulations such as the LGPD [179], their practical adoption remains limited due to implementation complexity and the energy constraints of wearable and IoT-based affective systems.

Local (Edge) Processing: A small number of studies adopted local processing strategies to reduce data exposure during transmission [9]. Although effective from a privacy-by-design perspective, these approaches often overlook the trade-offs imposed by limited computational resources when processing complex multimodal affective data on-device.

Terminology note: In this review, anonymisation refers to the irreversible removal of identifiers, whereas de-identification refers to masking or removing direct identifiers but may still allow re-identification under certain conditions. Pseudonymisation denotes replacing identifiers with pseudonyms under a separately stored key. No primary study explicitly reported pseudonymisation as a distinct strategy; when authors used broad terms such as “anonymised” or “de-identified” without technical detail, we retained their wording and classified the study under the closest applicable category.

3.3.2. Identified Gaps and Shortcomings

Despite the strategies discussed above, a substantial portion of the reviewed literature continues to overlook security and privacy considerations. Among the 170 studies included in this systematic review, 121 articles (70.76%) did not explicitly address any aspect of security or privacy. In our coding scheme, “did not address” indicates not reported: the article does not mention or describe any privacy or security measure. Explicit statements that protection was unnecessary were uncommon and treated separately as explicit dismissals (e.g., [12]).

Even in application-oriented research, such as stress monitoring wearables [176] or multimodal affective systems [19], privacy and security are often treated as peripheral documentation requirements rather than architectural design priorities. Among studies reporting adherence to institutional ethical guidelines (33.6%), a critical methodological opacity remains. In many cases, ethical approval is mentioned without operationalizing these principles for sensitive physiological or biometric data, resulting in the absence of clearly defined data retention policies, access control mechanisms, or signal obfuscation strategies.

A notable conceptual shortcoming is observed in studies such as [12], which argued that voice recordings did not require privacy considerations, restricting such concerns to physiological or visual data. This assumption overlooks well-documented biometric re-identification risks, as vocal features can be exploited to infer identity, health status, or behavioral traits, directly conflicting with the principle of data minimization. Such positions reveal a limited understanding of contemporary privacy threats associated with multimodal affective data.

The scarcity of concrete technical solutions further highlights this gap. With the exception of isolated efforts, such as the real-time encrypted emotion recognition pipeline based on homomorphic encryption proposed in [22], the literature has largely failed to bridge the gap between algorithmic accuracy and cyber-physical security. The absence of comparable implementations incorporating encryption, anonymization, local processing, or blockchain-based architectures in recent studies reinforces the conclusion that technical protection mechanisms remain significantly underexplored.

Ethical frameworks such as the Declaration of Helsinki are frequently cited as guiding references for research involving human participants. However, although referenced in several studies (e.g., [3,34]), their practical application varies substantially and often lacks sufficient detail. This limitation is also evident in more recent publications, where ethical approval is commonly reported without accompanying descriptions of data handling procedures, storage safeguards, or participant protection mechanisms.

Only 3.4% of the reviewed studies explicitly reported obtaining informed consent from participants [2,3,161], and this proportion does not improve in the most recent years of the review period. Similarly, a subset of studies claimed that privacy was protected but provided no technical or procedural details to substantiate this claim [3,21,36,37,111]. Such vague statements hinder reproducibility, limit methodological transparency, and undermine confidence in the reported results.

Taken together, these findings demonstrate that security and privacy considerations remain largely decoupled from system design in affective computing research. While ethical awareness is increasingly acknowledged at a declarative level, it is seldom translated into enforceable technical safeguards. This persistent gap underscores the need for future research to move beyond ethical compliance as a formal requirement and toward the systematic integration of privacy- and security-by-design principles throughout the lifecycle of affective intelligent systems.

3.4. RQ4—Which Datasets Were Used, and What Types of Resources Were Employed?

Datasets Used

The analysis of the review period (2013–2025) identifies a persistent challenge regarding data availability and standardization.

Table 6. Classification of reviewed experimental studies by dataset type ($N=170$). Private/custom datasets are the dominant source, representing 60.59% ($n=103$) of categorized papers.

References	Dataset Type
[2,3,4,7,8,11,12,16,18,19,21,23,24,30,31,33,34,35,36,37,38,45,46,51,52,55,59,60,62,63,67,68,73,74,75,77,78,79,80,82,83,84,85,86,87,88,89,91,94,95,97,98,99,100,101,108,110,111,112,115,116,117,118,120,122,123,124,127,128,133,135,137,142,143,144,145,146,147,148,149,150,151,153,154,155,156,158,160,161,162,163,164,166,167,168,169,170,174,180,182,183,184]	Private/Custom
[6,56]	JAFFE Dataset (Zenodo)
[9,17,20,22,43,44,48,49,58,61,65,71,72,76,81,93,104,132,136,140,171]	WESAD
[10,25,26,44,47,50,53,57,64,66,90,92,96,125,175]	FER2013
[70]	MAUS Dataset (IEEE Dataport)
[93,102,139,152]	DREAMER (Zenodo)
[152]	GAMEEMO (Kaggle)
[10,26,69,103,106,107,121,126]	RAVDESS (Kaggle)
[105]	Emotions (Kaggle)
[113]	SAVEE/SemEval2018 (Kaggle)
[129,134,138]	EEG Brainwave (Kaggle)
[130,131,140]	CASE
[141]	Neurological Status (Physionet)/PsPM-HRA1 (Zenodo)
[171]	Anxiety Phases Dataset (APD)
[172]	COVID-19 Tweets
[157,181]	DEAP (Kaggle)
[165]	DASPS (IEEE Dataport)
[159]	Tweet/Emotion Analysis (Kaggle)

summarizes the experimental studies reviewed ($N=170$). We identified $n=178$ dataset entries because several papers utilized multiple datasets [10,26,44,93,140,152,171]. Private or custom datasets represent 60.59% ($n=103$) of the categorized entries. This category includes research based on bespoke data collection as well as studies that merged four or more datasets to create hybrid repositories. This trend indicates a focus on domain-specific research but constrains reproducibility in affective computing [4,21,74,76,120]. Public datasets account for 34.71% ($n=59$) of the mentions. An additional 8 studies (4.71%) provide insufficient information or references regarding their data sources [32,54,97,109,114,119,173,176], a lack of documentation that exposes significant transparency gaps. Furthermore, limited sharing of raw data and inconsistent preprocessing protocols restrict performance comparison and hinder the establishment of robust benchmarks.

Wearable Stress and Affect Detection (WESAD) is the most frequently reported public benchmark ($n=21$) [9,17,20,22,76,132]. It provides multimodal physiological recordings typically captured via wearable devices for stress recognition research. Similarly, FER2013 remains a primary benchmark for facial emotion recognition ($n=15$), frequently employed for training convolutional neural networks [25,26,96]. Although it contains over 35,000 labeled images, its laboratory-curated nature limits the capture of spontaneous behavior found in real-world environments. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) dataset ($n=8$) also reflects a growing interest in vocal and audiovisual emotion recognition [69,106,126]. Other specialized resources include the DREAMER dataset ($n=4$) for EEG/ECG signals [93,102], and the CASE dataset for high-temporal-resolution physiological recordings [130]. The EEG Brainwave Dataset (Kaggle) offers 2549 feature representations [129,138], while works such as [152] combine DREAMER with the GAMEEMO dataset. Additional benchmarks include JAFFE and MAUS for facial expressions and mental workload, respectively [56,70]. Multimodal physiology is also addressed by DEAP [157], DASPS for anxiety elicitation [165], and social media corpora for textual emotion inference [159]. Notably, [141] combined the Skin Conductance Responses in Fear Conditioning with Visual CS and Electrical US (PsPM-HRA1) dataset with the Non-EEG Dataset for Assessment of Neurological Status. Despite the availability of these benchmarks, private or custom datasets remain prevalent due to specific contextual requirements, such as virtual reality or clinical monitoring. While bespoke data enhance ecological validity, limited accessibility and inconsistent reporting (e.g., [133,135]) continue to hinder cross-study comparability and transparency in the field.

3.5. RQ5—Which Artificial Intelligence Techniques Were Identified, and Which Evaluation Metrics Were Used?

The analysis of the reviewed journal articles published between 2013 and 2025 reveals a methodological landscape dominated by classical machine learning and deep learning techniques applied to affective computing tasks. Overall, the identified approaches reflect a gradual evolution from interpretable, traditional algorithms toward more complex and hybrid architectures designed to handle high-dimensional, non-stationary, and multimodal data.

Table 7. Extracted data for answering RQ5 (2013–2025).

Ref.	Data Type	Target	Algorithm
[2]	Speech	Stress	MLP, SVM, k-NN
[3]	Phys. Signals	Emotions	k-NN, PSO
[4]	Text	Emotions	PrefixSpan
[6]	Images	Emotions	CNN
[7]	Eye Tracking, Vehicle Data, Environmental Context	Dominance, Arousal, Valence	ConvLSTM, Hybrid Attention Mechanism
[8]	Phys. Signals	Stress	LSTM, RF
[9]	Phys. Signals	Stress	CNN, CondConv, Matching Network, SNN, SVM
[10]	Video, Audio, Facial Features	Stress, Anxiety, Depression	DNN, Ensemble Learning, Transfer Learning
[16]	Movement and Bio-signals	Emotions	ML Classifiers and Fusion Models
[17]	Physiological Signals	Emotions, Stress	Hybrid CNN-LSTM and Attention Mechanisms
[20]	Multimodal (HRV, EDA, Temp, Acc)	Emotions, Stress, States	Self-Supervised and Contrastive Learning
[21]	Text, Physiology, Smartphone Data, Weather Data	Mood, Stress	HBLR, MTMKL, MTL
[22]	Encrypted Phys. Signals	Others (Privacy), Emotions	CNN on Fully Homomorphic Encryption
[33]	Phys. Signals	Stress	DNN
[34]	Phys. Signals	Emotions	JMI, PCA, k-NN
[35]	Text	Depression	Deep Multimodal Multitask System
[38]	Text	PTSD	Boosted Trees, CART, Neural Networks, RF, SVM
[43]	Phys. Signals	Emotions, Stress	CNN-LSTM, Encoder, FCN, MCDCNN, MLP, MLP-LSTM, ResNet, StresNet, TPE, TPEFCN, Time-CNN
[55]	EEG Signals	Stress	ELM, IELM, AdaBoost
[56]	Facial Data	Stress	SVM, Tree-based Algorithm
[58]	Multimodal	Stress	K-NN, LDA, RF, SVM, Naive Bayes, NN and Ensemble Learning
[59]	Phys. Signals	Stress	Rule-based
[63]	Phys. Signals	Stress	ANN, ANFIS, SVR, SVM, k-NN
[75]	Phys. Signals	Stress	k-NN, Logistic Regression, Naive Bayes, RF, SVM
[76]	Phys. Signals	Stress	AdaBoost, RF, SELF-CARE
[93]	Multimodal (ECG, EDA, EEG)	Emotions, Stress	CNN-LSTM, Random Forest, and XGBoost
[96]	Facial Images (Video)	Stress	Modified VGG-Face (Deep Learning)
[102]	Phys. Signals	Emotions	CNN
[106]	Speech and Game Logs	Emotions	CNN, LSTM, MLP, RF, SVM
[110]	Speech	Emotions	LDA, k-NN
[111]	Phys. Signals	Dominance, Arousal, Valence	DCNNER
[112]	Speech	Emotions	GMM, RNN-LSTM
[117]	Speech, MFCC	Emotions	NSL
[124]	Phys. Signals	Emotions	ResNet50, CNN
[130]	Multimodal	Emotions	RF, SVM, XGBoost, CNN, ATTN
[131]	Multimodal	Emotions	LR, SVM, RF
[134]	EEG and Speech	Emotions	Roberts Similarity and PSO Selection
[135]	Multimodal (videos)	Emotional States	RF, Leave-One-Group-Out (LOGO)
[140]	Physiological Signals (EDA)	Emotions	LR, RF, XGB, MLP, RFECV
[141]	Phys. Signals	Arousal	Bayesian Filtering, Point Process State-Space Model
[145]	Phys. Signals	Arousal, Valence	DT, LDA, LSVM, SVM-RBF, k-NN
[151]	Physiological (HRV, EEG, GSR)	Emotions, States	Review of ML/DL models (CNN, RNN)
[155]	Phys. Signals	Emotional States	BDNN-CSMHPM
[157]	Multimodal	Emotions, high × low emotional valence	Supervised Contrastive Learning (SCL), SHAP, t-SNE
[158]	Body Movement (Kinematics)	Emotional States	Linear and Non-linear Dynamics Analysis
[159]	Text (Social Media)	Emotions, Depression, Anxiety	RF, DT, LR, LightGBM, GCN, IGCN
[161]	Speech	Depression	XGBoost
[165]	Phys. Signals	Anxiety	FFT, RF
[168]	Phys. Signals	Panic	F2D-CapsNetF
[181]	Phys. Signals	Emotions	BiLSTM

summarizes the extracted information regarding data types, prediction targets, and artificial intelligence techniques employed across the full analysis period.

Early studies in the reviewed corpus predominantly relied on classical machine learning methods such as Support Vector Machines (SVMs) and k-Nearest Neighbors (k-NN). These algorithms were frequently adopted due to their robustness, relatively low computational cost, and suitability for physiological and EEG data. Their use is consistently reported across multiple years and application domains, including stress detection, emotion recognition, and arousal analysis [2,3,34,63,75,145]. Ensemble-based methods, particularly Random Forest, also appear repeatedly throughout the period, being applied to text analysis, anxiety detection, and physiological stress prediction [8,38,75,76,165].

As the field matured, deep learning approaches became increasingly prominent. Convolutional Neural Networks (CNNs) were widely used for processing physiological signals, facial images, and multimodal sensor data [6,9,102,124]. Recurrent architectures, especially Long Short-Term Memory (LSTM) networks, were adopted to capture temporal dependencies in physiological and EEG signals [112,181]. Hybrid models combining CNN and LSTM components emerged as a dominant trend, enabling joint spatial and temporal feature learning in affective datasets [17,43,81,106].

More recent studies introduced additional architectural complexity, including attention mechanisms, capsule networks, and multimodal fusion frameworks. Examples include ConvLSTM models with attention layers for emotion recognition [7], CapsNet-based architectures for panic detection [168], and contrastive or self-supervised learning strategies applied to physiological and behavioral data [20]. Multimodal fusion approaches were increasingly explored to combine speech, text, physiological signals, and movement data, aiming to improve robustness and generalization across contexts [16,19,35].

Although performance optimization remains the primary focus of most studies, a small subset of recent works has begun to integrate security-aware design choices within the artificial intelligence pipeline. Notably, [22] proposed a convolutional neural network operating over fully homomorphically encrypted physiological signals, representing the only identified study that explicitly combines advanced cryptographic techniques with affective model training. However, such approaches remain isolated and are not yet representative of common practice.

Regarding evaluation metrics, classification accuracy, precision, recall, F1-score, and area under the ROC curve are the most frequently reported measures across the reviewed studies. Regression-oriented tasks commonly employ mean squared error, mean absolute error, or correlation coefficients. Despite this apparent consistency, evaluation protocols vary considerably, particularly with respect to dataset partitioning, cross-validation strategies, and class imbalance handling, thereby limiting direct comparability between studies. Overall, the findings indicate that while artificial intelligence techniques applied to affective computing have evolved substantially over the 2013–2025 period (Table 7), methodological heterogeneity remains high. Classical machine learning methods continue to coexist with increasingly sophisticated deep learning architectures, while evaluation practices lack standardization. This diversity underscores the need for clearer methodological reporting and harmonized evaluation frameworks to support reproducibility and longitudinal comparison across affective computing research.

4. Discussion

As with many systematic reviews in computer science, this study did not apply a standardized quality assessment instrument typically used in clinical research. Instead, the notion of study quality was operationalized through the selection of well-established, peer-reviewed scientific databases and the rigorous application of predefined inclusion and exclusion criteria. This approach ensures a consistent baseline of editorial and scientific quality across the selected studies. Nevertheless, heterogeneity in experimental design, dataset size, and evaluation protocols may still influence the strength and comparability of individual findings. This limitation reinforces the descriptive and exploratory nature of the present review.

4.1. The Evolution of Affective Architectures and Temporal Modeling

The analysis of the 170 reviewed articles reveals a clear and accelerating shift toward deep learning and hybrid architectures. While classical algorithms such as k-NN and SVM remain relevant due to their interpretability and low computational overhead [28,145], the expanded corpus from 2024 and 2025 shows an increasing adoption of models capable of capturing complex temporal and spatial dependencies in non-stationary physiological data [7,17]. This evolution corroborates research trajectories identified in earlier studies [28,29], which argue that static modeling approaches are insufficient to represent the dynamic nature of emotional and affective signals. To synthesize these findings, Figure 6 illustrates the contemporary affective computing pipeline derived from this review, highlighting the transition from multimodal data acquisition to privacy-preserving inference mechanisms. In this context, an emerging paradigm is the integration of federated learning (FL) directly into the data acquisition stage [185]. By enabling decentralized model training on local devices, FL ensures that sensitive raw signals remain on the user’s hardware, thereby addressing privacy concerns at the source before the data are processed by deep learning models.

The transition to hybrid architectures (CNN–LSTM) and attention mechanisms is empirically supported by the 2024–2025 works (see Table 7), which highlight the use of these models to capture the non-stationary nature of physiological signals [15,17,20,93]. Recent models achieve high performance by leveraging temporal dependencies in biomarkers, which were a major limitation in earlier studies [17]. For instance, while earlier work achieved 90.35% accuracy for valence using classical classifiers, more recent hybrid models based on CNN–LSTM architectures and attention mechanisms report accuracies above 95% by explicitly modeling temporal dynamics [17,93]. This performance leap suggests that the future of affective computing lies in architectures capable of prioritizing salient features within heterogeneous and multimodal data [15,20].

However, a significant performance–complexity trade-off persists. While deep CNN–LSTM models with attention mechanisms achieve high accuracy, their computational cost remains a barrier for deployment on resource-constrained devices [17]. In contrast, the SELF-CARE framework achieved a competitive 94% accuracy using ensemble methods [76], demonstrating that optimized classical and ensemble-based approaches remain a robust alternative for wearable and edge-based affective systems.

Beyond physiological sensing, the reviewed literature reinforces the role of heterogeneous input modalities as a defining characteristic of contemporary affective computing systems. In addition to wearable biosignals, several studies leverage eye tracking to capture attentional and cognitive correlates of affective states [7,8], neural signals acquired through EEG-based brain–computer interface technologies to model emotional and mental states at the neural level [145,165], and social or behavioral data derived from online platforms to infer affective patterns expressed through text and interaction dynamics [4,50]. The convergence of these modalities within unified learning architectures highlights a shift toward richer multimodal representations, in which affective inference emerges from the fusion of physiological, neural, behavioral, and contextual signals rather than from isolated data sources.

4.2. Individual Biomarkers and Environmental Context in Occupational Stress

A critical synthesis of the occupational stress literature (Table 4) indicates that stress monitoring should be understood as a multifaceted phenomenon rather than a binary classification problem. Occupational stress emerges from a dynamic interaction between individual biological characteristics and environmental triggers, a complexity that is often simplified in computational models [186,187].

As illustrated in the proposed conceptual framework (see Figure 7), effective identification of occupational stress requires a multidimensional approach. Individual characteristics, such as physiological baselines and personal biomarkers, directly influence how a subject responds to a given workload [29]. In parallel, emotional factors and environmental conditions, including workplace demands, noise levels, task structure, and professional context [188], act as external modulators that can intensify or mitigate stress responses [186].

When crossing occupational stress studies with the dataset analysis from RQ4, a consistent pattern becomes evident. Most studies addressing occupational stress rely on private or custom datasets (e.g., [7,36,70,73,86,94,120,168]), whereas only a small subset explicitly reports the use of public benchmark datasets. The WESAD dataset appears in occupational stress research in a limited number of cases (e.g., [44]), and FER2013 is used even less frequently (e.g., [57]).

Beyond physiological benchmarks, some studies employ publicly available visual and multimodal datasets, such as facial image and video collections [6,10,56], as well as datasets combining audio, video, and contextual information [15,20]. Text-based and social media datasets are also used for affective and emotional state inference, particularly in studies focused on depression, well-being, and emotional expression [4,50,128]. While these resources support reproducibility and comparative evaluation, they typically represent aggregated patterns across limited participant cohorts and are often collected under controlled or semi-controlled conditions.

As a result, most current models addressing occupational stress learn generalized representations of stress responses that do not explicitly account for inter-individual variability or worker-specific physiological baselines [7,36,70,73,86,94,120,168]. Evidence of experiments based on individualized, longitudinal, or worker-specific datasets remains limited. In many occupational stress studies, the lack of detailed reporting on the data used, data accessibility, or calibration procedures makes it difficult to verify whether models were effectively personalized at the individual level. As a result, it is often not possible to determine whether stress identification is based on subject-specific physiological baselines or whether it meaningfully incorporates workplace context, job-related conditions, or other psychological and emotional stressors [187]. This reliance on generic, aggregated, or insufficiently documented datasets constrains the assessment of personalization and provides a concrete explanation for its limited adoption across the occupational stress literature.

4.3. Social and Ethical Implications: The Privacy-by-Design Milestone

The expansion of affective computing into healthcare, occupational, and educational settings reveals a persistent gap between algorithmic progress and the implementation of effective ethical safeguards. In the 2024–2025 update, 57% of the reviewed studies did not address data security or regulatory compliance, which constitutes a critical omission given the requirements established by the AI Act [178]. This regulation prohibits emotion recognition systems in workplace and educational environments, except for narrowly defined medical or safety purposes. When permitted, such systems are classified as high risk and must comply with strict obligations related to transparency, human oversight, and data governance.

In Brazil, the LGPD [179] reinforces these requirements by defining biometric data used for emotion recognition as sensitive personal data. This classification mandates enhanced security measures and grants individuals the right to contest automated emotional assessments. Together, the AI Act and the LGPD establish a regulatory framework that demands a transition from generic ethical statements toward verifiable technical enforcement mechanisms.

The 2025 research landscape reflects early steps toward this transition through the consolidation of the Privacy-by-Design paradigm [189]. A central development is the integration of Fully Homomorphic Encryption into convolutional neural network pipelines [190]. Cha et al. [22] demonstrated that affective inference can be performed directly on encrypted tensors, ensuring that sensitive biometric information remains protected throughout the inference process. This approach addresses the long-standing tension between high-fidelity physiological sensing and the legal requirement to preserve user privacy.

Complementing this line of work, Sagar et al. [157] introduced an explainable contrastive learning framework for privacy-preserving consumer devices. Their approach employs a reduced EEG electrode montage to support wearable deployment while using SHAP for post hoc interpretability. This design aligns with transparency and accountability requirements imposed by both the AI Act and the LGPD, while preventing exposure of raw neural data.

A parallel research direction focuses on the combination of Federated Learning and Homomorphic Encryption to support secure collaborative healthcare analytics [191]. Recent implementations of the PPFLHE framework demonstrate that multiple institutions can jointly train affective models without sharing raw patient data. Secure aggregation mechanisms based on CKKS or Paillier encryption maintain predictive accuracy and computational efficiency, while ensuring data confidentiality, a prerequisite for large-scale clinical deployment [190].

Advancing toward a trustworthy affective AI ecosystem requires the adoption of explicit technical standards rather than reliance on institutional ethical approvals alone. This includes the specification of robust encryption protocols, such as AES-256, and the implementation of validated de-identification procedures to mitigate biometric re-identification risks [185]. As affective systems transition from controlled laboratory environments to real-world medical and occupational applications, ensuring data integrity, algorithmic accountability, and compliance with international regulatory frameworks becomes a fundamental requirement for adoption. Future research must also address vulnerabilities in wireless communication channels, such as Bluetooth and Wi-Fi, and incorporate lightweight cryptographic schemes tailored to IoT-based physiological monitoring [177].

5. Conclusions

The primary objective of this study was to examine how affective computing has been applied across different contexts and to assess the impacts of these applications on users’ lives, with particular attention paid to the healthcare domain. Through a systematic literature review of 170 peer-reviewed articles retrieved from four well-established scientific databases, this work synthesized key challenges, methodological trends, and application areas in affective computing over the period from 2013 to 2025 [1,2,3]. The analysis revealed a broad range of use cases, with a strong emphasis on affective intelligent systems and their implications for mental health, personal well-being, and data protection [4,10,35].

Across the reviewed literature, a substantial proportion of studies focused primarily on the technical validation of algorithms for detecting emotional or affective states. While these contributions provide an essential methodological foundation, many earlier works exhibited limited connection to practical or socially impactful applications [2,145]. Over time, however, the field has shown a gradual shift toward more application-oriented research. Studies published in later years increasingly report multimodal systems and closed-loop architectures designed for deployment in real-world settings, such as clinical monitoring, occupational stress management, and personalized mental health support [19,21,25]. Despite these technical advances, significant gaps persist with respect to social, ethical, and regulatory considerations. Although frameworks such as the European Union Artificial Intelligence Act and the Brazilian General Data Protection Law impose strict requirements for the handling of biometric and affective data, explicit discussion of these regulations remains limited in the reviewed literature [178,179].

Notably, more than half of the studies published in the most recent period still fail to adequately address data security or privacy issues, underscoring a persistent disconnect between algorithmic development and responsible data governance [34,36,37]. At the same time, recent research introduces promising directions for addressing these challenges. Emerging approaches include privacy-preserving pipelines based on homomorphic encryption, as well as advanced multimodal deep learning fusion strategies aimed at improving robustness while reducing data exposure [17,18,22]. These developments demonstrate the potential of affective computing systems to enhance quality of life while maintaining ethical integrity, provided that protection mechanisms are integrated into system design rather than treated as secondary considerations.

In summary, affective computing has reached a high level of algorithmic sophistication over the past decade, but its long-term success depends on bridging the gap between predictive performance and practical feasibility. This includes the development of lightweight and energy-efficient architectures suitable for wearable and embedded devices, alongside the establishment of robust and standardized protocols for data security and privacy [9,31]. Addressing these challenges is essential for transforming affective intelligent systems into trustworthy, ethically aligned, and socially beneficial technologies.