The Convergence of Artificial Intelligence in Measuring Attention and Emotion in Digital Technology-Enhanced Tertiary Education: A Scoping Review

Abstract

This scoping review maps AI-based approaches used to infer or measure attention and emotion in technology-enhanced learning (TEL), with a particular focus on tertiary (higher) education and learning analytics-enabled digital environments supporting online and hybrid instruction. Although artificial intelligence (AI) promises personalized digital education, many systems still respond poorly to students’ attentional and emotional fluctuations. We therefore examined the extent to which the literature converges on jointly measuring attention and emotion through AI in educational contexts, especially in virtual and distance-learning settings. Following PRISMA-ScR, we searched Scopus and Web of Science and identified 39 eligible studies. We conducted a methodological quality appraisal using Joanna Briggs Institute tools, a keyword co-occurrence bibliometric analysis, and a narrative synthesis. The evidence shows a rapidly expanding field and a wide range of AI-based techniques, but emotion and attention are typically operationalized and modelled in isolation. Both the bibliometric and narrative results indicate persistent conceptual fragmentation and limitations in the validity of measurement metrics. Overall, the field has not yet established a unified paradigm that integrates attention and emotion within AI-driven educational systems, constraining their adaptive potential. This evidence highlights the need for theory-informed and operational frameworks that enable genuinely holistic, student-centred pedagogical adaptation.

1. Introduction

Digital transformation has reshaped education, enabling personalized learning environments that adjust content and pacing to students’ trajectories. In this context, artificial intelligence (AI) has become a key driver of change by combining machine learning and data analytics to generate adaptive learning paths, continuous formative assessment, and immediate feedback (Deri et al., 2024; Southworth et al., 2023). These advances fuel expectations of more flexible and inclusive education (Kharchenko et al., 2024). However, many AI-based systems still show a form of “blindness” to students’ emotional and attentional fluctuations (Almusaed et al., 2023), despite strong evidence that academic emotions modulate motivation and achievement (Pekrun et al., 2011) and that attention—supported by frontal executive functions—organizes alerting, orienting, and cognitive control (Goldberg, 2001; Posner & Petersen, 1990). Neglecting these dimensions implies operating with an idealized student and increases the risk of frustration, boredom, or anxiety, which can undermine academic success (Martucci et al., 2025; Prinsloo et al., 2023).

To address this disconnect, recent research explores more sophisticated AI integrations that provide learning environments with computational sensitivity. However, as AIClass illustrates (Yuan et al., 2023), simply adding algorithms does not guarantee meaningful improvement: what is required is an integration that explicitly links cognitive, emotional, and attentional dimensions. Automating tasks or personalizing content without understanding how emotion and attention shape cognition underuses AI’s potential. Such integration should be anchored in established frameworks from educational psychology and neuroscience. Control–Value Theory (Pekrun et al., 2011) explains how emotions such as enjoyment, hope, anxiety, or boredom emerge from appraisals of control and value and, in turn, influence motivation, learning strategies, and achievement. Multifaceted attention models (Goldberg, 2001; Posner & Petersen, 1990) describe networks devoted to alerting, orienting, and executive control.

For operational clarity, it is useful to distinguish three layers that are frequently conflated in the literature. Emotion typically refers to relatively discrete states (e.g., enjoyment, anxiety, boredom) with valence and intensity, linked to appraisals and with downstream consequences for cognition and behaviour. Affect often denotes a more diffuse or sustained tone (positive/negative) that can modulate experience without necessarily mapping onto specific categorical emotions. Motivation concerns the direction and energization of goal-oriented behaviour; it can be shaped by emotions and affect, but it is not equivalent to them. This distinction matters because many computational proposals label as “emotion” signals that function primarily as proxies of affect or motivation/engagement, thereby increasing conceptual ambiguity and weakening the educational validity of resulting inferences (Pekrun et al., 2011).

Despite the strength of these pillars, the literature shows fragmented adoption in AI-driven educational systems. Whereas many reviews focus almost exclusively on attention, the present scoping review broadens the lens and, from a neuroeducational perspective, considers both attentional and emotional dimensions. The field remains organized in silos: some studies focus on emotion recognition, and others on attention or engagement, but few articulate both dimensions to inform real-time pedagogical adaptation. This gap is a major barrier to truly intelligent, person-centred online learning. Accordingly, a systematic mapping is needed to document the extent and nature of this fragmentation. This review aims to (1) identify and categorize studies that use AI to measure attention, emotion, or both in educational contexts; (2) critically examine theoretical frameworks, measurement methodologies, and technologies used; and (3) synthesize findings to identify patterns, convergence, and gaps, with particular attention on neuroeducational integration of both constructs.

Crucially, the effectiveness of these AI systems depends not only on their computational accuracy but also on their integration into pedagogical practice and stakeholder acceptance. Therefore, this review encompasses both the technical mechanisms of measurement and the contextual factors (perceptions, adoption barriers) that mediate their deployment.

In this review, the primary contribution is not to expand an inventory of AI applications in education, but to provide a field-level diagnosis. Specifically, our findings support three claims: (i) the literature remains fragmented across emotion-focused and attention/engagement-focused lines of work, with joint modelling still being exceptional; (ii) this fragmentation reflects the absence of integrative operational frameworks that connect emotion–attention dynamics to pedagogically meaningful adaptive decisions; and (iii) recurrent methodological weaknesses—particularly around measurement validity and reporting transparency—continue to constrain generalisability and educational usefulness. These gaps matter because emotion and attention have well-established links to learning-relevant processes within educational psychology and cognitive neuroscience frameworks (Pekrun et al., 2011; Posner & Petersen, 1990; Goldberg, 2001).

Accordingly, the purpose of this study is not merely to ask whether AI can “measure emotions”, but to examine the extent to which research converges on integrating emotion and attention in educationally interpretable ways, the levels of technical evidence and ecological validity reported, and the priority gaps that must be addressed to enable genuinely student-centred adaptive systems

Research questions (RQs).

To map the state of the art and characterize the fragmentation identified in the literature, this review is structured around the following research questions:

RQ1. Which variables, signals, and technological frameworks are used to infer, measure, or evaluate the adoption of emotions in digital education settings (with a particular focus on tertiary education)?

RQ2. Which variables, signals, and computational approaches are used to infer or measure attention/engagement, and how are they related to emotion in proposed models?

RQ3. What levels of technical evidence and ecological validity (including user acceptance) are reported, and what priority gaps remain for more reliable learning analytics and adaptive systems?

2. Materials and Methods

We conducted a scoping review with a systematic literature search to identify, map, and synthesize evidence at the intersection of artificial intelligence (AI), emotions, attention, engagement, and personalized learning in digital education, with a particular focus on tertiary education. The process followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidance (Tricco et al., 2018), ensuring transparent reporting across identification, screening, data extraction, and evidence synthesis.

Protocol and registration: No review protocol was registered prior to conducting this scoping review. The methods were defined a priori by the authors and are reported in accordance with the PRISMA-ScR guidance.

2.1. Eligibility Criteria

Eligibility criteria were defined a priori to ensure the relevance and quality of the final corpus. We included the following peer-reviewed sources: (a) Works that were published between 2012 and 2025 inclusive were included; the 2012–2025 time window was selected for two complementary reasons: (i) from 2012 onwards, the educational AI ecosystem increasingly reflects contemporary conditions relevant to this review (scalable learning analytics, online/hybrid TEL environments, and machine learning approaches applied to digital traces and multimodal signals); and (ii) this period captures the most relevant growth phase for assessing whether the field has progressed from isolated models toward operational emotion–attention integration with adaptive potential. The interval therefore maximizes relevance and comparability for addressing the research questions while avoiding dilution by pre-digital or TEL-misaligned approaches. (b) We also included studies that explicitly focused on AI applications in educational contexts; (c) considered emotions and/or attention as study variables; (d) addressed AI applications in digital education, including measurement/inference of emotion, attention and/or engagement, or the evaluation of AI-driven tools designed for these purposes (including adoption and perception studies); (e) were published as journal articles, peer-reviewed conference proceedings, or peer-reviewed scholarly book chapters retrieved via Web of Science and indexed in the searched databases; and (f) were written in English or Spanish. We excluded editorials, brief commentaries, or opinion pieces without a clear methodology or verifiable conceptual contribution; studies that focused on non-educational contexts; and sources without access to full text.

Deviation Management: Although the review primarily focused on tertiary education, a limited subset of studies conducted in upper-secondary contexts (e.g., Quigley et al., 2017) or using generalized datasets was retained. These sources were included exclusively when they tested transferable AI architectures (such as clickstream analytics or wearable IoT frameworks) whose technical validity is independent of the educational level and which address specific gaps in measuring attention/engagement not yet fully covered in the tertiary-specific literature.

2.2. Information Sources and Search Strategy

We conducted a systematic search in two high-impact electronic databases, Scopus and Web of Science (WoS), selected for their broad coverage across education, psychology, neuroscience, and computer science. The search strategy combined three conceptual pillars (artificial intelligence; emotion and attention; educational context). Titles and abstracts were screened using the Rayyan platform. The search strings below were adapted to each database’s syntax.

2.2.1. Scopus

(TITLE-ABS-KEY((“artificial intelligence” OR “machine learning” OR “adaptive learning” OR “intelligent tutoring system”) AND (“emotion recognition” OR “achievement emotion” OR “student engagement” OR “attention tracking” OR “real-time feedback”) AND (“online classroom” OR “virtual learning” OR “higher education” OR “ai-enhanced learning” OR “learning analytics”))) AND PUBYEAR > 2011 AND PUBYEAR < 2026 AND (LIMIT-TO(DOCTYPE, “ar”) OR LIMIT-TO(DOCTYPE, “cp”)) AND (LIMIT-TO(SUBJAREA, “EDUC”) OR LIMIT-TO(SUBJAREA, “COMP”) OR LIMIT-TO(SUBJAREA, “PSYC”)).

2.2.2. Web of Science (WOS)

TS = ((“artificial intelligence” OR “machine learning” OR “adaptive learning” OR “intelligent tutoring system”) AND (“emotion recognition” OR “achievement emotion” OR “student engagement” OR “attention tracking” OR “real-time feedback”) AND (“online classroom” OR “virtual learning” OR “higher education” OR “ai-enhanced learning” OR “learning analytics”)) AND PY = (2012–2025).

2.3. Study Selection and Data Extraction

The study selection process, illustrated in the PRISMA-ScR flow diagram (Figure 1), comprised three stages: (1) identification, including removal of duplicates from search results; (2) screening, in which two independent reviewers assessed titles and abstracts using Rayyan and excluded clearly irrelevant studies; and (3) eligibility, in which the same two reviewers examined full texts of the remaining sources for final inclusion. Disagreements at any stage were resolved through discussion and consensus.

After selecting the final corpus of 39 studies, data extraction was conducted systematically using a standardized spreadsheet. For each included source, we extracted: (a) bibliographic information (authors, publication year, title, journal or conference); (b) study design (e.g., experimental, quasi-experimental, review), sample size, and participant characteristics; (c) AI intervention (if applicable), including components and objectives; (d) outcomes and measures, including metrics used to assess emotions, attention, engagement, and/or performance; and (e) data analysis methods, including statistical or qualitative techniques. In line with PRISMA-ScR, quality appraisal was used for descriptive and interpretive purposes rather than as an exclusion criterion.

2.4. Methodological Quality Appraisal

To assess rigour and risk of bias, each included study was critically appraised using the Joanna Briggs Institute (JBI) tools (Joanna Briggs Institute, 2020). The specific JBI checklist was selected according to study design (e.g., JBI Critical Appraisal Checklist for Systematic Reviews, Qualitative Research, or Quasi-Experimental Studies). Two reviewers independently appraised each article against the relevant checklist items. Results were compared and disagreements resolved through consensus discussion to ensure reliability of the final judgements.

Consistent with PRISMA-ScR (Tricco et al., 2018), the JBI appraisal was used for descriptive and interpretive purposes rather than as an exclusion criterion (Joanna Briggs Institute, 2020). Specifically, appraisal outcomes were used to: (i) characterize recurring methodological strengths and weaknesses in the field (e.g., measurement validity, reporting transparency), (ii) weight thematic inferences with appropriate caution, and (iii) inform the discussion regarding scientific maturity and ecological validity of proposed approaches. Purely technical algorithm validation studies were appraised descriptively, as their assumptions and performance metrics are not always directly comparable to educational or psychosocial study designs.

2.5. Data Analysis and Synthesis

Given the heterogeneity of interventions, populations, and outcome measures, a quantitative meta-analysis was not appropriate. Instead, we adopted a dual synthesis approach. First, we conducted a thematic narrative synthesis to organize findings around recurring patterns and themes. Second, we performed a bibliometric keyword co-occurrence analysis using VOSviewer (version 1.6.20) (van Eck & Waltman, 2010) to visualize the intellectual structure of the field, identify thematic clusters, and detect disconnections between research areas, providing visual evidence of conceptual gaps.

2.6. Study Selection

The coordinated search in Scopus and Web of Science, conducted in May 2025, yielded 372 records (Web of Science = 206; Scopus = 166). Prior to screening, a formal data-cleaning step removed 253 records: 32 duplicates and 221 records were excluded during preliminary cleaning based on recoverability and formal eligibility (e.g., insufficient metadata for screening and/or non-recoverable full text) and/or an evident mismatch with the AI–education–attention/emotion/engagement focus based on available metadata.

Preliminary cleaning (n = 221). Before title/abstract screening, we performed a technical pre-processing step to ensure the evaluability and comparability of imported records (e.g., verification of minimum metadata required for screening, normalization of bibliographic fields, and checking full-text recoverability). We also removed records clearly outside the scope based on available metadata (source outlet, subject area, and dominant terms). This pre-processing—common in scoping reviews with broad search strategies—helped consolidate a homogeneous set of records suitable for screening.

After preliminary cleaning, 119 unique records proceeded to title/abstract screening, and 15 were excluded for not meeting inclusion criteria. We sought full-text retrieval for 104 records, but 2 could not be obtained.

Consequently, 102 full texts were assessed in detail for final eligibility. At this stage, 63 documents were excluded for the following reasons: 32 did not address the AI–education intersection with measurement/inference/modelling of attention, emotion, and/or engagement; 11 focused on AI policy or management without a direct educational application or target variables; 10 focused on language learning without a relevant affective–attentional component or without analytics/AI aligned with the objective; 5 addressed ICT competences or perceptions/attitudes without measurement or inference of the target constructs; and 5 applied AI in education without a substantive relationship to adaptation/assessment of learning or to the target constructs (attention/emotion/engagement). The final corpus comprised 39 sources included in the qualitative synthesis and bibliometric analysis. The full selection flow is presented in the PRISMA-ScR diagram (Figure 1).

3. Results

This section presents the review results. We first report the source selection flow (Figure 1), followed by the characterization of the included corpus, methodological quality appraisal, bibliometric analysis of the field, and, finally, a narrative synthesis of thematic findings.

3.1. PRISMA-ScR Flow Diagram (Study Selection)

The identification, screening, and eligibility process is summarized in Figure 1 (PRISMA-ScR). In brief, 372 records were identified (Scopus = 166; Web of Science = 206). After removing 32 duplicates and applying preliminary cleaning (221 records), 119 records were screened by title and abstract and 15 were excluded. We sought 104 full texts and could not retrieve 2. Finally, 102 full texts were assessed; 63 were excluded for the reasons noted in Figure 1, and 39 sources were included in the qualitative synthesis and bibliometric analysis.

3.2. Characteristics of Included Studies

The final corpus comprised 39 scientific documents published between 2017 and 2025, with a notable concentration from 2021 onwards, suggesting a field in accelerated expansion. Evidence shows high methodological heterogeneity, consistent with the emerging and interdisciplinary nature of the AI–affective/cognitive measurement–education intersection. The corpus includes experimental and quasi-experimental designs (e.g., technology interventions in higher education; Huang et al., 2025; Yang et al., 2023), qualitative studies focused on teachers’ perceptions and practices (e.g., Ulla et al., 2024), conceptual proposals or design frameworks without controlled empirical implementation (e.g., Zaibout et al., 2024), and psychometric/instrument validation studies supported by analytical and statistical/computational models (e.g., Moreira-Choez et al., 2025), as well as systematic reviews and synthesis studies providing relevant secondary evidence (e.g., Prinsloo et al., 2023; Ren & Wu, 2025).

Geographically, the output is global, with notable contributions from Asia, Europe, and North America. In terms of educational scope, populations and settings are predominantly higher education, and to a lesser extent secondary education, alongside studies centred on teachers or on datasets (e.g., technical investigations validating models on multimodal datasets). Application contexts include online learning platforms, intelligent tutoring systems, instant messaging supported by AI, immersive/virtual reality environments, multimodal learning analytics, and tools based on generative models.

To enable cross-study comparison despite design diversity,

Table 1. Characteristics of the sources included in the scoping review.

ID	Year	Study Title	Authors	Source	Study Design	Population/Unit of Analysis	AI/Technology	Key Outcomes/ Measures
1	2025	Serious Games Analytics in VR Environments: A Two-Stage Systematic Literature Review	(Ismayilzada et al., 2025)	Journal of Interactive Learning Research	Systematic literature review (two-stage SLR).	Not applicable (synthesis; no empirical sample). Unit: 10 included studies (LAK + JLA) on VR serious games analytics.	Synthesis of AI/learning analytics methods used across included studies (e.g., ML, time series, social network analysis).	Reported constructs across included studies include cognitive engagement and visual attention (e.g., EEG and eye tracking).
2	2023	Semantic fusion of facial expressions and textual opinions from different datasets for learning-centered emotion recognition	(Cárdenas-López et al., 2023)	Soft Computing	Technical/computational study (multimodal model benchmarking).	Public datasets (facial images + text); no participant sample (not reported, NR).	Deep learning multimodal fusion (facial expression + text sentiment/opinion) for emotion recognition in learning-centred contexts.	Model performance metrics (e.g., accuracy/F1); emotion recognition (attention not directly measured).
3	2025	Exploring the impact of artificial intelligence on curriculum development in global higher education institutions	(Abbasi et al., 2025)	Education and Information Technologies	Cross-sectional survey study.	Higher education stakeholders (faculty and students), n ≈ 2000.	No experimental intervention; assesses AI use/impact perceptions in curriculum development.	Perceptions and use patterns; perceived benefits/challenges (includes engagement/innovation-related perceptions; no direct attention/emotion measures).
4	2020	Comparative Analysis for Boosting Classifiers in the Context of Higher Education	(Abou Gamie et al., 2020)	International Journal of Emerging Technologies in Learning (iJET)	Technical/computational comparison of algorithms.	Higher education academic datasets (Not reported, NR).	Boosting classifiers for predicting academic success.	Algorithm performance (accuracy, efficiency, validation); no direct attention/emotion measurement.
5	2024	Leveraging Artificial Intelligence in Higher Educational Institutions: A Comprehensive Overview	(Deri et al., 2024)	Revista de Educación y Derecho (Education and Law Review)	Narrative review/overview.	Not applicable (synthesis; no empirical sample).	Descriptive synthesis of AI applications in higher education (personalization, assessment/feedback, analytics, administration) and related ethical issues.	Conceptual discussion of benefits/risks (engagement, effectiveness, ethics/privacy); no primary quantitative measures.
6	2023	Research on student engagement in distance learning in sustainability science to design an online intelligent assessment system	(Fang et al., 2023)	Frontiers in Psychology	Qualitative study with system-design component.	n = 42 students and n = 2 educators (semi-structured interviews).	Design of an online intelligent assessment approach; uses ML to detect low engagement and support teacher feedback.	Operationalization of engagement; identification of low-engagement patterns; recommendations/support logic (no direct attention/emotion instrumentation).
7	2022	Formative Assessment Tasks as Indicators of Student Engagement for Predicting At-risk Students in Programming Courses	(Veerasamy et al., 2022)	Informatics in Education	Educational data mining/predictive modelling study.	Higher education programming course data (not reported, NR).	Analytics/ML to predict at-risk students using engagement indicators derived from formative tasks.	Engagement indicators + academic risk/prediction outcomes; no direct emotion/attention measures.
8	2023	Philosophy of self-learning education implemented in a virtual education system	(Ma et al., 2023)	Learning and Motivation	Applied implementation study.	Learners in an asynchronous virtual learning system (not reported, NR).	AI-supported virtual system with text-based emotion inference for personalization.	Use/retention and inferred affective signals; learning experience indicators (attention not directly measured).
9	2024	Towards advanced digital assessments: Artificial intelligence, gamification, and learning analytics	(Zaibout et al., 2024)	International Journal on Technical and Physical Problems of Engineering (IJTPE)	Conceptual/theoretical model proposal.	Not applicable (conceptual model; no empirical sample).	Proposes an assessment framework integrating AI, gamification, and learning analytics.	Expected outcomes discussed conceptually (engagement, motivation, performance); no direct attention/emotion measures.
10	2023	Multimodal learning analytics—In-between student privacy and encroachment: A systematic review	(Prinsloo et al., 2023)	British Journal of Educational Technology	Systematic review.	Not applicable (synthesis; no empirical sample).	Synthesizes multimodal data capture/analytics and associated privacy/ethical concerns.	Reported constructs across the literature (engagement, learning outcomes, proxy affect/attention signals) + ethical/privacy themes.
11	2025	Determinants of Adopting 3D Technology Integrated With Artificial Intelligence in STEM Higher Education: A UTAUT2 Model Approach	(Truong & Pham, 2025)	Computer Applications in Engineering Education	Cross-sectional survey (UTAUT2).	Higher education STEM participants (students and/or instructors), n ≈ 300.	No intervention; assesses adoption/intention regarding 3D technology integrated with AI.	UTAUT2 constructs; no direct attention/emotion measures.
12	2024	Can Multimodal Large Language Models Enhance Performance Benefits Among Higher Education Students? An Investigation Based on the Task–Technology Fit Theory and the Artificial Intelligence Device Use Acceptance Model	(Al-Dokhny et al., 2024)	Sustainability	Cross-sectional survey with SEM.	n = 550 higher education students.	Evaluates perceived benefits/acceptance of multimodal LLMs using TTF/AIDUA-based model.	Perceived performance benefits and acceptance/use; no direct attention/emotion instrumentation.
13	2025	Predicting Student Academic Success Using Deep Learning: A Multi-Factor Approach to Performance Prediction	(Al Husaini et al., 2025)	Journal of Logistics, Informatics and Service Science	Quantitative predictive modelling study.	OULAD dataset (n = 32,593 students; learning analytics traces).	Deep learning model (MLP) to predict academic success using multiple factors.	Prediction performance (classification metrics) and factor contribution; no direct emotion/attention measures.
14	2024	University Teachers’ Views on the Adoption and Integration of Generative AI Tools for Student Assessment in Higher Education	(Khlaif et al., 2024)	Education Sciences (MDPI)	Cross-sectional survey.	Higher education teachers (faculty) (not reported, NR).	Perception/adoption of generative AI tools for student assessment (no intervention).	Adoption drivers/barriers, perceived risks/benefits; no direct student emotion/attention measures.
15	2025	A Comparative Analysis of On-Device AI-Driven, Self-Regulated Learning and Traditional Pedagogy in University Health Sciences Education	(Chun et al., 2025)	Applied Sciences (MDPI)	Quasi-experimental mixed-methods study (AI group vs. control).	n = 86 higher education health sciences students.	On-device GenAI-supported learning (LLM-based) compared to traditional pedagogy.	Academic performance, learning time, satisfaction; engagement/perceptions (no direct attention/emotion instrumentation).
16	2023	Developing a model for AI Across the curriculum: Transforming the higher education landscape via innovation in AI literacy	(Southworth et al., 2023)	Computers and Education: Artificial Intelligence	Position paper/curricular framework proposal.	Not applicable (conceptual model; no empirical sample).	AI literacy model and recommendations for integrating AI across higher education curricula.	Conceptual outcomes (competencies, implementation guidance); no direct emotion/attention measures.
17	2025	A complex dynamical system approach to student engagement	(Li et al., 2025)	Learning and Instruction	Analytical multimodal modelling study.	n = 61 third-year medical students (higher education) in an ITS context.	OpenFace-based multimodal extraction (gaze/head pose/AUs) + dynamical modelling of engagement.	Cognitive engagement prediction using multimodal proxies (attention via gaze; affective facial cues).
18	2018	Adaptive quizzes to increase motivation, engagement and learning outcomes in a first year accounting unit	(Ross et al., 2018)	International Journal of Educational Technology in Higher Education	Quasi-experimental cohort comparison.	Higher education first-year accounting students (online unit) (not reported, NR).	Adaptive quizzes integrated into course platform.	Motivation and engagement (survey) + learning outcomes/grades.
19	2024	Integrating AI in college education: Positive yet mixed experiences with ChatGPT	(Song et al., 2024)	Meta-Radiology	Case study/descriptive evaluation.	Higher education course students (not reported, NR).	Integration of ChatGPT (ChatGPT-4 Turbo–enhanced teaching application integrated into course activities).	Usage and perceptions (usefulness, accuracy, concerns) and perceived engagement; no direct attention/emotion instrumentation.
20	2025	Emotional engagement in a humor-understanding reading task: an AI study perspective	(Martucci et al., 2025)	Current Psychology	Cross-sectional AI-based affect measurement study.	n = 132 university students.	Automated facial-expression analysis during humour-reading task.	Detected emotions/affective engagement + task performance; attention not directly measured.
21	2023	Developing a gamified artificial intelligence educational robot to promote learning effectiveness and behavior in laboratory safety courses for undergraduate students	(Yang et al., 2023)	International Journal of Educational Technology in Higher Education	Randomized controlled trial (RCT).	n = 53 undergraduate students (higher education).	Gamified AI educational robot intervention.	Motivation, flow, cognitive load, and performance; engagement via questionnaires/observations.
22	2024	Adaptive learning in computer science education: A scoping review	(Barbosa et al., 2024)	Education and Information Technologies	Scoping review.	Not applicable (synthesis; no empirical sample).	Maps adaptive learning approaches and reported metrics/technologies.	Reported constructs across studies (engagement, outcomes, methods); no direct attention/emotion measures.
23	2025	Application and optimization of digital situated teaching in university finance courses from a constructivist perspective: An analysis based on machine learning algorithms	(Liu et al., 2025)	Education and Information Technologies	Quasi-experimental/longitudinal analysis.	n = 514 university students (finance courses; multi-year dataset).	ML methods (clustering/trees/GBM) to analyze/optimize digital situated teaching.	Learning outcomes + participation/engagement indicators; no direct attention/emotion measures.
24	2023	Investigating the Association Between Student Engagement With Video Content and Their Learnings	(Amashi et al., 2023)	IEEE Transactions on Education	Analytical association study.	Students in video-based learning context (not reported, NR).	Learning analytics applied to video interaction data.	Engagement with video + learning outcomes; no direct attention/emotion instrumentation.
25	2024	Artificial Intelligence Use in Feedback: A Qualitative Analysis	(Pang et al., 2024)	Journal of University Teaching and Learning Practice	Qualitative study.	Higher education teachers (not reported, NR).	Explores AI use in feedback practices (no controlled intervention).	Qualitative themes: perceived utility, risks, adoption considerations; no direct attention/emotion measures.
26	2022	Assessing how QAA accreditation reflects student experience	(Rybinski, 2022)	Higher Education Research & Development	Document analysis/large-scale analytical study.	Accreditation documents + large-scale student experience evidence (dataset-based).	Analytics/AI to compare accreditation indicators with student experience metrics.	Experience/satisfaction indicators; no direct attention/emotion measures.
27	2025	Validation of a teaching model instrument for university education in Ecuador through an artificial intelligence algorithm	(Moreira-Choez et al., 2025)	Frontiers in Education	Instrument validation (quantitative; SEM).	n = 413 higher education teachers (Ecuador).	AI/SEM used as analytical support to validate an instrument (not an educational intervention).	Reliability/validity metrics; no direct attention/emotion measures.
28	2024	Integrating AI in Higher Education: Applications, Strategies, Ethical Considerations	(Al Daraai et al., 2024)	IGI Global	Narrative synthesis/position.	Not applicable (synthesis; no empirical sample).	Discusses AI applications, strategies, and ethical considerations in higher education.	Conceptual implications (ethics/privacy, adoption, engagement); no primary measures.
29	2025	ChatGPT-enhanced mobile instant messaging in online learning: Effects on student outcomes and perceptions	(Huang et al., 2025)	Computers in Human Behavior	Randomized controlled trial (RCT).	n = 63 graduate students (16-week intervention).	ChatMIM mobile instant messaging enhanced with ChatGPT for online learning.	Learning outcomes + perceptions/usage + engagement measures (not direct attention/emotion instrumentation).
30	2020	Utilizing Multimodal Data Through fsQCA to Explain Engagement in Adaptive Learning	(Papamitsiou et al., 2020)	IEEE Transactions on Learning Technologies	Case study with configurational analysis (fsQCA).	Adaptive learning multimodal dataset (not reported, NR).	fsQCA to identify configurations leading to high/low engagement.	Engagement as outcome; multimodal predictors (proxy signals).
31	2024	Generative AI in Higher Education Assessments: Examining Risk and Tech-Savviness on Student’s Adoption	(Oc et al., 2024)	Journal of Marketing Education	Longitudinal qualitative study.	Higher education students (not reported, NR).	Examines GenAI use in assessments; adoption shaped by perceived risk and tech-savviness.	Risk perception, adoption/use behaviours, and reported affective factors; no direct attention measures.
32	2024	Co-designing enduring learning analytics prediction and support tools in undergraduate biology courses	(Plumley et al., 2024)	British Journal of Educational Technology	Design and development (longitudinal co-design cycles).	Undergraduate biology course context (higher education) (not reported, NR).	Co-designed predictive analytics and support tools (incl. explainability/XAI).	Prediction/support for SRL/engagement; explainability outcomes; no direct attention/emotion instrumentation.
33	2024	Digital Technologies as a Factor of Transformation of Learning in the University Education	(Kharchenko et al., 2024)	Revista Românească para Educaţie Multidimensională	Mixed-methods study.	n = 200 teachers + n = 500 students (higher education).	Examines digital transformation and GenAI tools (ChatGPT, Gemini, and Copilot; versions not reported) in university learning contexts.	Perceived benefits/challenges and engagement-related perceptions; no direct attention/emotion instrumentation.
34	2025	Examining Teaching Competencies and Challenges While Integrating Artificial Intelligence in Higher Education	(Ren & Wu, 2025)	TechTrends	Systematic review.	Not applicable (synthesis; no empirical sample).	Synthesizes evidence on teaching competencies and challenges for AI integration in HE.	Themes/competencies/barriers; no primary attention/emotion measures.
35	2024	How can GenAI foster an inclusive language classroom? A critical language pedagogy perspective from Philippine university teachers	(Ulla et al., 2024)	Computers and Education: Artificial Intelligence	Qualitative study.	n = 14 university teachers (language classroom context).	Explores GenAI (e.g., ChatGPT) for inclusive pedagogy and critical language teaching.	Qualitative themes (inclusion, participation, confidence); no direct student attention/emotion instrumentation.
36	2017	Using learning analytics to understand scientific modeling in the classroom	(Quigley et al., 2017)	Frontiers in ICT	Case study (learning analytics).	Secondary school biology students using EcoSurvey (not reported, NR).	Clickstream analytics + ML to characterize scientific modelling processes.	Modelling complexity/completeness + engagement patterns (clickstreams); no direct attention/emotion instrumentation.
37	2020	Data Capture and Multimodal Learning Analytics Focused on Engagement with a New Wearable IoT Approach	(Camacho et al., 2020)	IEEE Transactions on Learning Technologies	Technical/applied system proposal with evaluation.	Educational context with wearable IoT data capture (not reported, NR).	Wearable IoT architecture for multimodal capture + learning analytics to infer engagement.	Engagement inferred from multimodal/proxy physiological signals; participation indicators.
38	2024	Evaluating English Teaching Quality in Colleges Using Fuzzy Logic and Online Game-Based Learning	(Wu & Luo, 2024)	Computer-Aided Design & Applications	Applied modelling study (fuzzy logic/IF-AHP).	Higher education (“colleges”); evaluation dataset (not reported, NR).	IF-AHP (fuzzy logic + AHP) applied to teaching-quality evaluation in an online game-based learning context.	Quality indicators and engagement-related metrics; no direct attention/emotion instrumentation.
39	2024	Inclusive Learning and Assessment in the Era of AI	(Nayak et al., 2024)	SN Computer Science	Survey/review article (conceptual synthesis).	Not applicable (synthesis; no empirical sample).	Discusses inclusive learning and assessment strategies in the AI era.	Conceptual themes: inclusion, assessment, participation, fairness; affective/engagement implications discussed (no direct measures).

uses standardized synthesis categories for: (a) document type and study design; (b) population/unit of analysis (human participants, teachers, institutions, or datasets); (c) role of AI/technology (educational intervention, analytical support, or algorithm evaluation/benchmarking); and (d) key outcomes/measures (performance, engagement/use, perceptions, and, where applicable, affective or attentional indicators). This harmonization is intentional and used as a mapping strategy to identify patterns and gaps, without replacing each study’s specific description. Table 1 summarizes the main characteristics of the 39 included documents.

3.3. Risk of Bias and Quality Assessment Results

Aggregated results of the risk-of-bias appraisal are presented in Figure 2. To ensure statistical consistency, the percentages reported in this section were calculated based on the subset of 34 studies with standardizable clinical or social science designs, excluding the five technical algorithm validation studies which were appraised descriptively. The analysis reveals systematic strengths and weaknesses across the examined literature. The domains with the highest methodological adherence were “Clarity of the question/objective” (85% of the rated studies showed low risk of bias) and “Methodology and design” (78%). In contrast, the domains with the most frequent deficiencies were “Search Strategy and Biases” (32% rated as high risk), often due to limitations in searching the grey literature or language restrictions. “Measurement validity and reliability” also presented challenges (25% high risk), reflecting the frequent use of ad hoc instruments. Meanwhile, “Sample and Participants” showed moderate methodological adherence, with only 12% of studies rated as high risk in this domain.

3.4. Bibliometric Analysis

To visualize the intellectual structure of the field, we conducted a keyword co-occurrence analysis using VOSviewer. Figure 3 presents the resulting network map, where nodes represent keywords, node size reflects occurrence, and links indicate co-occurrence relationships (shorter distances suggest stronger relatedness). The network is anchored by high-frequency terms such as “artificial intelligence”, “learning”, “education”, “assessment”, and “higher education”, while “learning analytics” occupies a bridging position connecting assessment-, prediction-, and support-tool-related terms.

Figure 4 provides the cluster-based thematic view of the same keyword set, where colours denote clusters. Four clusters are evident: (i) a cluster centred on learning analytics and system-level support tools; (ii) a cluster focused on emotion-related concepts (e.g., emotion recognition and emotional engagement), which appears comparatively peripheral and weakly connected to the learning analytics backbone; (iii) a cluster emphasizing assessment and academic outcomes; and (iv) a cluster capturing predictive modelling and performance-related terms. Overall, the coloured clustering highlights that emotion- and attention-related constructs are rarely modelled jointly in the retrieved literature and tend to appear in separate thematic strands, supporting the broader finding of conceptual fragmentation across attention–emotion research in technology-enhanced learning.

Because the full VOSviewer output for the keyword co-occurrence network was visually dense and difficult to interpret at publication scale, we present the bibliometric results in two complementary views: a simplified network map to convey overall structure (Figure 3) and a cluster-based thematic visualization to improve readability and interpretability (Figure 4). This separation is purely for graphical clarity; both figures are derived from the same keyword set and co-occurrence matrix.

3.5. Narrative Synthesis

The analysis of the 39 included sources reveals that AI is addressing attention and emotion through three distinct but complementary approaches: algorithmic inference at a technical level, focused on signal processing (e.g., facial recognition, sensors) to detect internal states directly; interaction analytics at a behavioural level, using log data and clickstreams as proxies for engagement; and contextual and perceptual evaluation at an ecological level, assessing the adoption, risks, and perceived utility of these AI tools by stakeholders, which provides the necessary context for real-world implementation. Within this structure, three interconnected thematic axes emerge alongside a cross-cutting area of challenges. The first recurring theme is the use of AI for multimodal measurement of affective and cognitive states. A substantial body of literature focuses on this task. For example, Cárdenas-López et al. (2023) showed that fusing image and text data improves emotion-recognition accuracy compared with single-modality approaches. Similarly, studies such as Li et al. (2025) and Camacho et al. (2020) explore facial expressions, eye tracking, and wearable biometric data to infer student engagement, demonstrating the technical feasibility of capturing indicators of students’ internal states through multiple data channels.

Interpretively, this theme reflects strong technical innovation but conceptual heterogeneity: attention, emotion, and engagement are operationalised through diverse proxies, often without a shared theoretical frame clarifying what each signal represents. As a result, cross-study comparability is limited, and the educational validity of inferences depends not only on model accuracy but also on the quality of the construct-to-metric mapping. This pattern suggests a field still consolidating, where instrumentation advances faster than conceptual standardization.

The second central theme is the application of AI to trigger adaptive pedagogical interventions. Beyond measurement, several works investigate how inferred data can be used to personalize the learning experience. In a randomized controlled trial, Huang et al. (2025) implemented a chatbot system that adapted feedback, leading to a significant increase in student engagement. Other works, such as Yang et al. (2023), explore gamification as an AI-enabled intervention strategy and report improvements in motivation and students’ flow experience when adaptive game elements are used.

Structurally, intervention studies indicate adaptive potential, yet they rarely make explicit how emotional and attentional signals are integrated into interpretable pedagogical rules; adaptation is often driven by aggregated engagement or performance indicators. This limits pedagogical traceability: even when adaptation “works”, it is not always clear whether it responds to emotion–attention dynamics or to broader behavioural patterns. Advancing field maturity therefore requires moving from “detect-and-adjust” to decision models that connect inferred internal states to theoretically justified instructional strategies.

The third prominent theme is the centrality of student engagement as a key construct. This term emerges as a prevalent target in the analyzed research. Studies such as Fang et al. (2023) and the review by Ren and Wu (2025) explicitly define engagement across behavioural, cognitive, and emotional dimensions and position it as a primary indicator of the success of an AI intervention. Al-Dokhny et al. (2024) similarly identify engagement as a key expected outcome from implementing large language models in education.

While engagement acts as a bridging construct, its widespread use also reveals operational ambiguity: in some studies it functions as a behavioural proxy of attention, in others it incorporates affective components, and in others it is treated as a global outcome. This flexibility supports broad uptake of the term, yet it complicates emotion–attention integration and can yield non-equivalent conclusions under the same label. The field would therefore benefit from more stable operational definitions and explicit triangulation between engagement, attention, and emotion.

Across these themes, the literature consistently reports technical and ethical challenges. Model accuracy and reliability are recurrent concerns, as reflected in the study by Song et al. (2024) on students’ perceptions of ChatGPT. Privacy and the ethics of using sensitive data are also highlighted as critical barriers in reviews such as Prinsloo et al. (2023), which notes limited evidence on scaling multimodal learning analytics while preserving student privacy.

Taken together, these challenges are not peripheral: they indicate that progress toward genuinely adaptive systems depends as much on improving metrics and measurement validity as on ensuring ethical implementation conditions (transparency, minimisation of sensitive data capture, and pedagogical oversight). Without these safeguards, technical sophistication may translate into fragile inferences or adoption resistance, ultimately limiting real-world educational impact (Prinsloo et al., 2023; Al Daraai et al., 2024).

4. Discussion and Conclusions

4.1. Interpretive Summary of Key Findings

This review maps a rapidly expanding research field on applying AI to measure and adapt learning processes. Across the 39 studies, three convergent axes emerge: the technical feasibility of multimodal measurement of affective and cognitive states; the use of such information to trigger adaptive pedagogical interventions; and the consolidation of student engagement as a primary success indicator. Cárdenas-López et al. (2023) and Camacho et al. (2020) show that indicators of students’ internal state can be captured through semantic fusion and physiological data, whereas the trials by Huang et al. (2025) and Yang et al. (2023) demonstrate the potential of chatbots and gamified robots to personalize educational experience. Nevertheless, triangulating the narrative synthesis with the methodological quality appraisal (Figure 2) and the bibliometric maps (Figure 3 and Figure 4) reveals a more fragmented landscape: keyword co-occurrence shows a clear disconnection between emotion and attention clusters, indicating that although both dimensions are often measured separately, their systematic integration is exceptional. This conceptual fragmentation is reinforced by methodological weaknesses in measurement validity and in the transparency of search strategies, which may be perpetuating research silos.

From an interpretive standpoint, the evidence converges on a dual fragmentation: (i) conceptual, as emotion, attention, and engagement are defined and labelled inconsistently across disparate theoretical frames; and (ii) methodological, as measurement metrics and procedures vary widely and frequently present limitations in validity and reporting. Consequently, systematic emotion–attention integration remains exceptional, which constrains the field’s progression from superficial personalisation toward robust, explainable, and pedagogically justified adaptive systems.

4.2. Comparison with Existing Literature

Placing these findings within prior research shows both convergence with established trends and a divergence that underscores this review’s added value. The centrality of student engagement as a key outcome aligns with reviews such as Ren and Wu (2025), which identify self-regulated learning and engagement as core targets of AI in education. Likewise, the emphasis on ethical and privacy challenges—especially in multimodal learning analytics—converges with the warnings raised by Prinsloo et al. (2023).

However, this review’s distinctive contribution lies in its integrative scope and neuroeducational lens. Unlike reviews such as Roig-Vila et al. (2025), or qualitative studies such as Ulla et al. (2024), our analysis explicitly seeks the intersection between attentional and emotional dimensions.

Moreover, whereas studies such as Barbosa et al. (2024) and Ismayilzada et al. (2025) map AI techniques in specific domains, our cross-cutting analysis suggests that the fragmentation between emotion measurement—shown as feasible in studies such as Martucci et al. (2025)—and performance optimization—explored by Ma et al. (2023)—is a general phenomenon. The finding by Ross et al. (2018), documenting a dissociation between improved perceived motivation and gains in academic performance, is particularly revealing: rather than an anomaly, this discrepancy may be interpreted as a symptom of the gap identified here. Without a model that integrates and responds to the emotion–attention dynamic, adaptive interventions risk being superficial and ineffective.

This pattern reflects persistent theoretical tensions: technical measurement (e.g., classifiers or behavioural proxies) does not always translate into actionable pedagogical interpretation, particularly when constructs are not anchored in stable operational definitions. Moreover, engagement often functions as an umbrella concept: it facilitates interdisciplinary communication, yet it may conceal substantive differences between sustained attention, emotional involvement, and motivation. Addressing these tensions requires models that explicitly connect signals, constructs, and instructional decisions, grounded in educational validity criteria rather than algorithmic performance alone.

4.3. Strengths of the Review

The validity and robustness of the conclusions are supported by several methodological strengths. First, strict adherence to PRISMA-ScR (Tricco et al., 2018) ensured a transparent, systematic, and reproducible process across all stages—from search strategy formulation to final study selection.

Second, as a key differentiator relative to other reviews in the area, we conducted a rigorous risk-of-bias appraisal of each of the 39 included studies using (JBI) tools (Joanna Briggs Institute, 2020). This quality analysis not only helps weigh evidence reliability but also provides data to interpret the maturity of the field.

Third, the dual synthesis approach—integrating an in-depth narrative synthesis with a quantitative bibliometric analysis—offers a more comprehensive view. While narrative synthesis captures qualitative detail, keyword co-occurrence mapping with VOSviewer (van Eck & Waltman, 2010) provides an objective visualization of the field’s structure, making it possible to identify relationships and, crucially, disconnections between major research domains. Finally, the breadth and recency of the corpus, including diverse designs and publications up to 2025, supports a timely picture of this rapidly evolving field.

4.4. Limitations of the Review and the Evidence

Several limitations should be acknowledged. Regarding the evidence base, the quality appraisal (Figure 2) reveals important weaknesses: restricted search strategies (e.g., Ismayilzada et al., 2025) that may bias conclusions; limitations in measurement validity and reliability in a meaningful share of studies—particularly when dealing with constructs such as emotion and attention; and insufficient reporting of ethical aspects in more than one third of studies, a concern also highlighted by Prinsloo et al. (2023). Regarding our review process, restricting inclusion to publications in English and Spanish may have excluded relevant research in other languages, and the typical publication bias associated with the academic literature remains. Moreover, the rapid evolution of AI—especially generative models, as discussed by Khlaif et al. (2024) and Al-Dokhny et al. (2024)—means that results should be interpreted as a snapshot of a moving target.

Overall, the observed risk-of-bias pattern suggests that key bottlenecks are not only technical but also psychometric and transparency-related. Limitations in measurement validity and reliability imply that portions of the evidence may capture partial correlates (e.g., activity or interaction) rather than well-defined emotional or attentional states, and incomplete reporting undermines replication and cross-context comparison. Consequently, claims about “adaptation” should be interpreted cautiously when the link between signal, construct, and pedagogical decision is not made explicit, or when measurement quality is insufficiently reported (Joanna Briggs Institute, 2020).

4.5. Implications and Future Directions

This scoping review has implications for educational practice, technology development, and the research agenda. From a practice and development perspective, the findings confirm that creating adaptive, affect-sensitive online learning systems is both technically feasible and pedagogically desirable.

Evidence from studies such as Veerasamy et al. (2022) and Abou Gamie et al. (2020), which demonstrate AI’s ability to predict at-risk students, supports the utility of such approaches. However, our synthesis also calls for caution and a shift in focus. Educational technology developers should not only improve the accuracy of emotion or attention classifiers in isolation, but also design architectures that integrate these data sources in educationally meaningful ways. The qualitative evidence reported by Pang et al. (2024) and users’ perceptions of chatbots in Song et al. (2024) remind us that technical effectiveness does not guarantee pedagogical benefit or user acceptance unless human factors are addressed.

For educators, this review suggests that while tools such as those discussed by Southworth et al. (2023) and Al Daraai et al. (2024) offer substantial potential to understand and support students, implementation should be critical and supported by strong AI literacy—an issue central to Deri et al. (2024).

For emotion–attention integration to be pedagogically meaningful, implications must be stated operationally. At the curriculum level, AI literacy should include criteria for interpreting signals (what they indicate, what they do not, and with what uncertainty) and for evaluating the impact of adaptation on learning and well-being—not only on short-term performance (Southworth et al., 2023). In teacher education, core competencies include: (i) reading dashboards and system explanations, (ii) recognizing measurement bias and uncertainty, and (iii) deciding when to intervene and when not to, aligned with teachers’ reported needs and concerns regarding AI adoption in higher education contexts (Khlaif et al., 2024). At the institutional level, governance frameworks should require prior validation, minimisation of sensitive data capture, auditability, and pedagogical oversight; adoption should ensure traceability (signal → interpretation → adaptation → evaluation), especially in multimodal settings where privacy risks are non-trivial (Prinsloo et al., 2023). Finally, at the policy level, minimum conditions for responsible deployment include transparency, independent validity assessment, stakeholder participation, and robust privacy and non-discrimination safeguards (Al Daraai et al., 2024).

In this sense, “integrating attention and emotion” is not merely measuring more variables; it means closing a full pedagogical loop: signal inference with known quality, construct-based interpretation, justifiable adaptation, and impact evaluation on academic outcomes and student well-being.

The implications for future research are particularly salient. Beyond mapping the field, this review diagnoses its main challenge: conceptual and methodological fragmentation. The most visible gap—reflected in the bibliometric maps and narrative synthesis—is the absence of a unified theoretical–practical framework that integrates simultaneous measurement of attention and emotion to guide pedagogical adaptation. Therefore, research should prioritize developing and empirically validating such integrated frameworks. Rigorous experimental studies are needed that not only propose models but also evaluate their causal impact on learning, motivation, and student well-being. Future work should move beyond predicting “academic success” and investigate underlying mechanisms, including how emotion regulation mediates the relationship between AI interventions and learning outcomes.

4.6. Confirmation of the Conceptual and Terminological Gap

To examine whether a formalized framework addressing this gap existed, we conducted a targeted post hoc search in Scopus, WoS, ERIC, and IEEE Xplore. This search, independent from the main corpus selection, used term combinations aimed at identifying an integrated neuroeducational–technological framework, such as “NeuroAdaptive Artificial Intelligence Learning Flow”, “Neuro-Adaptive Learning Flow”, and expressions combining “Neuroeducation”, “Artificial Intelligence”, and “Learning Flow”. The search yielded no relevant documents. This absence suggests that, beyond the scarcity of empirical studies integrating emotion and attention, the literature also lacks a formalized and named theoretical construct that unifies both dimensions. This reinforces the need to develop and validate integrative frameworks capable of supporting a new generation of truly neuroadaptive learning systems.

4.7. Conclusions

In conclusion, despite the growth of technically feasible research on AI for measuring emotional and attentional states in online education, the field remains fragmented. The lack of systematic integration between affective and cognitive dimensions, anchored in an established neuroeducational framework, appears to be the main obstacle preventing adaptive learning systems from moving beyond content personalisation towards genuinely learner-centred adaptation. This gap reflects not only limited empirical integration, but also the absence of a shared conceptual paradigm and unified nomenclature that formalize emotion–attention integration, calling for sustained efforts in the design and validation of integrative frameworks.

In sum, this study contributes a triangulated diagnosis—combining narrative synthesis, bibliometric structure, and appraisal-informed interpretation—showing that the primary barrier is not a lack of algorithms, but a lack of operational integration and educational validity. The identified gap—absence of integrative frameworks and shared nomenclature for emotion–attention—helps explain why many proposals remain difficult to transfer, interpret, or justify in authentic teaching practice. Priorities therefore include developing and validating theory-informed frameworks that connect signals, constructs, and instructional decisions, supported by replicable and ecologically valid studies that evaluate not only performance but also learner experience and well-being.

4.8. Responses to the Research Questions

Regarding RQ1, the literature uses a broad range of signals to infer or measure emotion in digital education, including multimodal approaches (e.g., text, voice, facial expression, physiological signals, and interaction traces). Machine learning and deep learning methods dominate, while psychometrically validated instruments integrated into learning analytics systems appear less frequently.

Regarding RQ2, attention/engagement is operationalised mainly through behavioural and interaction indicators (e.g., logs, clickstream, time-on-task, participation patterns) and/or observable signals (e.g., eye tracking/EEG in specific studies). However, explicit modelling of the emotion–attention relationship is exceptional: most studies treat both constructs separately, keeping the field organized in silos.

Finally, for RQ3, levels of evidence are heterogeneous and tend to centre on technical validation (model performance) with frequent limitations in construct validity, methodological transparency, and generalizability to real-world contexts. Priority gaps include: (i) integrative emotion–attention frameworks grounded in neuroeducation; (ii) metrics with stronger psychometric validity and multimodal triangulation; and (iii) replicable studies with complete reporting of data and procedures, alongside ethical and privacy considerations.