Artificial Intelligence Based Framework for Student Engagement Assessment in Classroom Environments

Abstract Student distraction remains a critical barrier to effective learning in classroom environments, yet existing detection methods relying on manual observation or post-session feedback are inherently subjective, time-intensive, and ill-suited for real-time adaptive teaching. This study presents the Artificially Intelligent Technology for Student Engagement Assessment in Classroom Environments (AI-TEACH), a novel co-teacher paradigm that automatically identifies and quantifies student distraction through multimodal fusion of asynchronized audio and video streams captured via classroom surveillance cameras. AI-TEACH integrates YOLO-NAS and ByteTrack for real-time student detection and tracking, MediaPipe for behavioral cue extraction, and Silero-VAD with Wav2Vec2-SVM for audio-based emotional distraction classification; a BiLSTM network fuses these multimodal features to generate per-incident severity scores, and a session-wide engagement report accessible through an interactive instructor dashboard. The framework was validated on a curated dataset of classroom recordings and further evaluated through a controlled experiment involving 200 students divided into experimental and control groups, achieving 90% accuracy and 92% recall in distraction detection. Compared with existing unimodal and non-real-time approaches, AI-TEACH offers a more accurate, scalable, and immediately actionable solution for objective student engagement monitoring in modern classroom practice, with direct implications for data-informed pedagogy and inclusive education. Introduction Student engagement is a cornerstone of effective learning, with research consistently demonstrating that actively engaged students achieve stronger academic outcomes, higher-order thinking skills, and deeper conceptual understanding [12, 14, 36]. Yet sustaining engagement across an entire classroom session remains one of the most persistent challenges educators face. Disengagement arises from a complex interplay of external distractions, emotional fatigue, and cognitive overload [18]. Among its many forms, distraction is perhaps the most insidious: it emerges gradually, erodes attentional focus, and quietly disconnects students from the instructional flow before either teacher or learner is fully aware [6]. Even fleeting lapses in attention can translate into lost learning opportunities and long-term academic setbacks if left unaddressed [5, 47]. As classroom sizes grow and instructional demands become more complex, the capacity to detect and respond to distraction in real time has shifted from an aspirational goal to a practical necessity. Distraction-aware systems that deliver timely, data-driven feedback hold considerable promise for augmenting educators’ capacity to re-engage learners and sustain instructional efficacy [28, 43]. The limitations of conventional distraction monitoring methods are well documented. Teacher observation, student self-reports, and post-lesson assessments share a common weakness: they are inherently subjective, retrospective, and susceptible to cognitive bias [17]. Self-reports demand that students accurately recall attentional states during a lesson a metacognitively demanding task that yields unreliable data [29]. Teacher-led assessment, though more direct, is impractical at scale: in large or diverse classrooms, educators cannot simultaneously deliver instruction and monitor the transient behavioral signals that distinguish genuine engagement from its absence [30]. These approaches also lack the temporal granularity needed to capture brief distraction events, and they provide no mechanism for real-time pedagogical response. The absence of scalable, automated monitoring tools therefore limits not only individual instructional effectiveness but also institutional capacity to ensure equitable, high-quality learning environments [24]. When distraction goes unaddressed at a systemic level, the consequences extend beyond missed lesson content, poor academic performance, reduced motivation, and elevated dropout risk can follow [30]. Existing automated approaches to student engagement monitoring have made notable progress, yet critical gaps remain. Computer vision-based systems have demonstrated the feasibility of detecting behavioral distraction cues such as head pose, gaze direction, and body posture from video alone [19, 27, 37]. Audio-based systems have explored speech activity and emotion recognition as indicators of cognitive-affective states [16, 25]. However, most existing frameworks are unimodal, address only one type of distraction signal, and are designed for controlled or online environments rather than the dynamic, multi-student conditions of a physical classroom [5, 40]. Multimodal fusion combining visual, postural, and audio cues into a unified real-time assessment remains underexplored in deployment-ready classroom systems. Moreover, decision-support paradigms in which AI acts as a co-teacher, continuously informing educator decision-making through actionable engagement metrics, have not been systematically developed or validated. The integration of multi-source data is further complicated by challenges in temporal synchronization, modality weighting, and low-latency inference, challenges that current literature has addressed only in part. Recent advances in fuzzy-set-based aggregation and multi-criteria decision-support frameworks offer promising directions for fusing heterogeneous, uncertainty-laden data in complex real-world environments (Naeem et al., 2022; Rehman et al., 2023; Khan et al., 2023; Ayub et al., 2024; Mahmood et al., 2023), and these principles have informed the design of AI-TEACH’s severity scoring and engagement aggregation modules. To address these gaps, this study introduces the Artificially Intelligent Technology for Student Engagement Assessment in Classroom Environments (AI-TEACH), a novel co-teacher paradigm for real-time, multimodal student distraction detection and engagement quantification. AI-TEACH is designed to operate continuously within a physical classroom using strategically positioned surveillance cameras with synchronized audio capture. Its architecture integrates three tightly coupled processing layers: a Data Acquisition Layer that captures raw audiovisual streams; a Fog Layer that performs low-latency, on-edge analysis; and a Cloud Layer that aggregates, stores, and visualizes session-level engagement data for educators. Within the Fog Layer, video frames are preprocessed using OpenCV and passed through YOLO-NAS for real-time student detection, with ByteTrack maintaining consistent student identity across frames. MediaPipe Pose extracts skeletal landmarks for behavioral distraction cue identification, including gaze aversion, slumping, head-on-desk posture, and object throwing. In parallel, classroom audio is segmented using Silero Voice Activity Detection (VAD), and acoustic embeddings extracted via Wav2Vec2 are classified by a Support Vector Machine (SVM) to detect verbal distraction events such as off-task chatter, loud sounds, and abusive language. A Bidirectional Long Short-Term Memory (BiLSTM) network fuses the multimodal feature streams, assigning severity-weighted distraction scores to individual incidents. Aggregated into a session-wide Class Distraction Score and visualized through an interactive educator dashboard, these outputs support timely, evidence-informed pedagogical interventions. The novelty of AI-TEACH lies in four key contributions. First, it presents a unified, deployment-ready multimodal framework that simultaneously processes visual behavioral cues and audio-based verbal cues within a single real-time inference pipeline, an integration that prior systems have not achieved in live classroom settings. Second, it introduces a structured severity scoring system that distinguishes distraction events by their disruptive impact, enabling educators to prioritize responses rather than react uniformly to all off-task behavior. Third, AI-TEACH operationalizes the co-teacher paradigm: rather than replacing educator judgment, the system augments it by surfacing objective, timestamped distraction data that supports responsive, inclusive instruction. Fourth, the system was validated not only on a curated dataset of over 5,350 labeled classroom images but also through a controlled experiment with 200 students, providing empirical evidence of real-world utility beyond benchmark performance. The primary objectives of this study are: - I. To propose and implement an AI-enabled multimodal framework for real-time distraction detection in physical classroom environments using synchronized audiovisual surveillance data; and. - II. To develop a severity-weighted engagement scoring system that delivers actionable, session-level insights to educators to enhance instructional effectiveness and promote equitable learning outcomes. The remainder of this paper is organized as follows. The Literature Review synthesizes existing research on engagement monitoring, multimodal emotion recognition, and distraction detection across sensing modalities, deployment settings, and task granularities. The Methodology section details the AI-TEACH architecture, including its data acquisition, fog-layer processing pipeline, distraction scoring mechanism, and cloud-based reporting engine. The Results and Analysis section evaluates system performance across video-based and audio-based detection modules using standard classification metrics. The Discussion situates findings within the broader literature, addresses pedagogical implications, and considers limitations and ethical dimensions. The Conclusion and Future Recommendations summarizes contributions and outlines directions for extending AI-TEACH’s scalability and adaptability. Literature Review Research on automated student engagement monitoring has grown substantially over the past decade, driven by advances in computer vision, audio signal processing, and deep learning. Despite this progress, the field remains fragmented across sensing modalities, task definitions, and deployment contexts. The following review synthesizes existing work along three organizing dimensions: (i) sensing modality: RGB-only, audio-only, and multimodal approaches; (ii) task granularity: ranging from broad attention and engagement estimation to fine-grained distraction event detection and behavioral taxonomy; and (iii) deployment setting: in-person classrooms, e-learning platforms, and immersive environments such as virtual reality. This synthesis reveals not only what individual methods achieve, but where collective gaps persist and how AI-TEACH addresses them. RGB-Only and Vision-Based Approaches The most extensively explored modality for classroom distraction monitoring is visual input, primarily in the form of RGB video or webcam imagery. These systems vary considerably in the granularity of behaviors they target and the degree to which they are designed for real classroom deployment versus controlled or online settings. At the engagement estimation level, Rao and Janet [27] developed a lightweight Teacher Assistance System that inferred student attention from head pose estimation in video recordings from Google Meet and WebEx sessions. Using the OpenCV Caffe face detection algorithm combined with Perspective-n-Point (PnP) pose estimation, the system derived head orientation from nose-to-chin landmark geometry without requiring calibrated hardware. While effective for coarse attention estimation in online environments, the approach is inherently limited to a single behavioral signal e.g. head direction and provides no sensitivity to facial affect, gesture, or auditory distraction. It also cannot scale to multi-student physical classrooms where occlusion and distance are significant challenges. Su et al., [37] advanced the task granularity by developing a video analytics framework for in-person classroom settings that used deep learning-based visual feature segmentation to recognize a broader behavioral taxonomy, including both positive behaviors such as hand-raising and negative behaviors such as head-on-desk and mobile phone use. The system employed Dlib and OpenPose for facial and skeletal analysis, drawing on the COCO pose dataset, and achieved real-time performance on edge-compatible hardware evaluated on the ICSAD dataset. While this represented a meaningful step toward multi-behavior detection from RGB input, the system remained unimodal and lacked any mechanism for detecting verbal or audio-based distraction, an important class of classroom disruptions that vision alone cannot capture. Trabelsi et al., [40] addressed a practical deployment challenge by extending YOLOv5 for face-mask-robust attention classification across nine behavior categories, specifically targeting post-COVID classroom conditions where facial occlusion would otherwise degrade performance. The system provided real-time visualization for instructors and demonstrated acceptable accuracy across YOLOv5 variants, though it lacked any engagement scoring mechanism and offered no audio or affective processing. Similarly, Wang, [43] modified the YOLOv5 architecture for the specific context of art education, introducing a bidirectional feature pyramid and model pruning to improve both classification accuracy and deployment efficiency. While the resulting system demonstrated strong behavior recognition in domain-specific classroom conditions, it remained focused on static behavioral labels and did not incorporate temporal dynamics, audio cues, or session-level engagement metrics. Le et al., [19] pursued a hardware-integrated vision approach by embedding cameras within classroom desks to capture close-range facial signals of inattention, including yawning, gaze aversion, and eye closure, using the MediaPipe library. Deployed on Jetson Nano devices, the system maintained high accuracy under variable lighting and proved robust across environmental conditions. However, the design relied entirely on facial signals and was not extensible to postural or gestural distraction cues, nor to any form of audio processing. Betto et al., [5] similarly focused on facial and postural features for e-learning settings, combining OpenFace and GAST-Net to extract gaze direction, neck tilt, and shoulder angles from webcam video, then classifying distraction using Random Forest and XGBoost models. The approach achieved good distraction recall in online sessions but was not designed for the multi-student dynamics, ambient noise, and physical complexity of live in-person classrooms. Taken together, RGB-only systems have demonstrated the feasibility of automated behavioral distraction detection, but they share a structural limitation: visual signals alone cannot distinguish cognitively disengaged students who remain physically still from genuinely attentive ones, nor can they detect verbal disruptions that may occur without any accompanying visible behavior. This ceiling motivates the integration of audio modalities. Audio-Only and Speech-Based Approaches Audio-based engagement monitoring has received comparatively less attention in the classroom domain than vision-based methods, but offers complementary sensitivity to verbal distraction events that visual systems miss entirely. Research in affective computing provides important foundations for this work. Zhao et al., [49] demonstrated that per-sample optimization during multimodal training as implemented in their TEMPO framework significantly improves emotion recognition by preventing dominant modalities from suppressing weaker signals, a finding with direct implications for fusion architectures that combine audio and video streams of unequal reliability. Zhu et al., [50] introduced the RMER-DT framework for multimodal emotion recognition in conversational contexts, leveraging diffusion models and transformer architectures to improve robustness in emotionally ambiguous or acoustically degraded conditions. Xiang et al., [45] further advanced multimodal emotion analysis through contrastive learning with audio-visual text generation, demonstrating that cross-modal consistency objectives improve representation quality when modalities are asynchronously captured. These findings are directly relevant to classroom settings, where audio and video streams are inherently asynchronous and subject to environmental noise. In the specific context of classroom monitoring, Kaewkaisorn et al., [16] developed the Quality Indicators for Remote Students (QIRS) framework for online learning environments, using behavioral and temporal features derived from audio-visual signals to train a Long Short-Term Memory (LSTM) classifier that categorized learners as focused or not focused with 90.2% accuracy. While effective for e-learning, the framework was not designed to handle the multi-speaker, ambient-noise conditions of a physical classroom, and its binary focused/not-focused categorization lacks the behavioral specificity required for actionable instructional feedback. Liao and Wu, [20] examined digital distraction from a learning analytics perspective, combining problem-based learning (PBL) methodology with machine learning analysis of Facebook group communications and survey data from graduate students. Their findings confirmed a negative association between reported digital distraction and academic achievement, while also demonstrating that objectively measured learning engagement was a stronger predictor of academic outcomes than self-reported distraction. Although not a real-time monitoring system, this work underscores the importance of objective, behavioral rather than self-reported measures of engagement, a principle foundational to the AI-TEACH design. Multimodal Approaches Multimodal systems that integrate visual, postural, and audio signals represent the most promising direction for comprehensive classroom distraction detection, yet deployment-ready implementations remain rare. Wang et al., [41] demonstrated in their RAFT framework that adversarial training during multimodal fusion can significantly improve robustness against modality-specific noise, a challenge of direct relevance to classroom systems where video may be partially occluded and audio corrupted by ambient sound. Zhu et al. [51] proposed a client-server architecture for non-contact single and multiple behavioral and emotional state assessment, demonstrating the viability of distributed real-time inference for multi-person monitoring, an architectural pattern analogous to AI-TEACH’s fog-cloud design. The integration of audio-visual multimodal signals with contrastive learning objectives (Zhang et al., 2024) further highlights the potential of cross-modal training strategies to improve classification accuracy beyond what each modality achieves independently. In the educational domain, Asish et al., [3] explored distraction detection in virtual reality (VR) learning environments using eye gaze data, classifying distraction into low, medium, and high levels using Random Forest, k-Nearest Neighbor, and XGBoost models. While the gaze-based approach provided fine-grained distraction levels and benefited from individualized model calibration, the VR-specific setup limits generalizability to conventional classroom contexts where gaze tracking hardware is unavailable. Importantly, even in highly immersive VR environments, external distractions and cognitive fatigue continued to induce disengagement, reinforcing that no sensing modality or immersion level fully eliminates the need for active engagement monitoring [33]. Ives et al., [15] approached engagement from a pedagogical intervention perspective, using observational methods to compare distraction levels under different polling technologies and instructional formats in introductory physics lectures. Their finding that distraction was substantially lower during active learning formats than passive delivery is consistent with existing cognitive load theory and has practical implications for how engagement monitoring outputs should be used by educators: not merely to identify problems, but to inform instructional mode decisions. However, the study relied entirely on non-digital observational methods with no automated detection or AI component, representing the continued need for scalable, objective monitoring tools. Nabung, [24] similarly relied on observational and literature synthesis methods to document the motivational drivers of digital multitasking e.g. habitual smartphone use, perceived productivity, and social bonding and to identify self-regulated learning strategies as a mitigation pathway. While illuminating the behavioral ecology of classroom distraction, this work also highlights the limits of non-automated approaches in capturing the temporal dynamics of attention lapses. Synthesis and Identified Gaps Across these three dimensions, sensing modality, task granularity, and deployment setting, several convergent gaps emerge. First, no existing deployment-ready system integrates RGB video, skeletal pose, facial landmark analysis, and audio-based speech emotion recognition into a single real-time inference pipeline specifically designed for multi-student physical classrooms. Second, behavioral taxonomies in existing systems are typically limited to two to nine coarsely defined categories, with no structured severity weighting that would allow educators to prioritize responses. Third, most systems are evaluated in online, single-student, or controlled settings and lack empirical validation in live multi-student classroom environments. Fourth, the co-teacher paradigm, in which AI continuously augments educator decision-making through actionable, session-level engagement metrics has not been operationalized or empirically validated in the classroom engagement literature. Table 1 provides a comparative summary of representative distraction detection approaches, highlighting the key methods, deployment settings, and limitations of each in relation to the design objectives of AI-TEACH. AI-TEACH directly addresses each of these identified gaps: it fuses video-based behavioral cues and audio-based verbal cues through a unified BiLSTM classification pipeline; it implements a structured ten-indicator severity scoring system calibrated to classroom disruption impact; it was validated in a live multi-student classroom experiment with 200 participants; and it operationalizes the co-teacher paradigm through a cloud-based educator dashboard that translates real-time distraction data into actionable pedagogical insights. Impact on Sustainable Development The emergence of AI-driven smart classroom technologies reflects a broader transformation in how educational institutions leverage data to improve learning outcomes, address inequity, and create more responsive instructional environments [2, 18, 30]. Intelligent learning environments that integrate real-time multimodal analytics have the potential to move education beyond one-size-fits-all approaches, enabling more personalized, inclusive, and evidence-informed teaching practice. Within this context, the AI-TEACH framework is positioned not merely as a technical instrument for distraction detection, but as a contribution to the global agenda for sustainable and equitable education articulated in the United Nations Sustainable Development Goals (SDGs). AI-TEACH is most directly aligned with SDG 4: Quality Education, which calls for inclusive, equitable, and high-quality learning for all. Specifically, the framework is designed to support Target 4.1, which seeks to ensure that all students benefit from effective learning outcomes in primary and secondary education, by providing educators with objective, real-time data on student engagement that can inform timely instructional adjustments. Target 4.5, which addresses the elimination of educational disparities for disadvantaged and marginalized groups, is relevant insofar as automated engagement monitoring can reduce reliance on subjective teacher observation, a process prone to unconscious bias, and can surface patterns of disengagement that might otherwise go unnoticed in large or diverse classrooms. Target 4.a, which calls for the construction of safe and inclusive school environments, aligns with AI-TEACH’s capacity to detect disruptive verbal behaviors, including abusive language and hostile interactions, in real time. Finally, Target 4.c, which aims to increase the supply of qualified, ICT-capable teachers, is supported by the co-teacher paradigm: by systematically surfacing behavioral data and engagement patterns, AI-TEACH can serve as a professional development resource, helping educators reflect on the relationship between instructional choices and student attention over time. Beyond SDG 4, the framework touches on several related goals. Its capacity to detect postural and behavioral indicators that may reflect fatigue, distress, or withdrawal connects to SDG 3: Good Health and Well-being (Target 3.4), by equipping educators with data that may prompt empathetic responses and appropriate referrals. The system’s potential to surface gender-differentiated engagement patterns across large student groups is relevant to SDG 5: Gender Equality (Targets 5.1 and 5.5), while its role in reducing disengagement-driven dropout risk has implications for SDG 8: Decent Work and Economic Growth (Target 8.6). The framework’s use of AI and cloud infrastructure for educational innovation connects to SDG 9: Industry, Innovation and Infrastructure (Target 9.5), and its promotion of inclusive pedagogical practices for marginalized learners relates to SDG 10: Reduced Inequalities (Target 10.2). The monitoring of hostile or aggressive classroom interactions aligns with SDG 16: Peace, Justice and Strong Institutions (Target 16.1), while the framework’s design as a shareable, institution-scalable system supports SDG 17: Partnerships for the Goals (Targets 17.6 and 17.18) by enabling collaborative, evidence-based educational policy development. It is important to note that the SDG connections articulated here represent the intended design rationale and theoretical alignment of AI-TEACH, not empirical claims about educational impact. The extent to which AI-TEACH demonstrably advances these goals in practice based on the results of dataset evaluation and the controlled classroom experiment conducted with 200 students is addressed in the Discussion section, where findings are interpreted against these stated objectives (Fig. 1). Methodology Nonverbal cues such as body movements, body orientation, and facial expressions can communicate students’ attentional condition and cognitive engagement [29]. Observe behaviors such as head turning, slouching, looking away, or otherwise physical removal of attention, fidgeting, and misdirected attention can all signal distraction in the form of disengaged behavior [31]. The proposed AI-TEACH system utilizes AI-based multimodal recall analysis of audiovisual data scraping to automatically code those behaviors. This provides marker-based real-time assessment at scale to monitor and promote cognitive thinking and attention in classrooms. The framework of the proposed Artificially Intelligent Technology for Student Engagement Assessment in Classroom Habitats (AI-TEACH) system with real engagement assessment in classroom habitats is illustrated in Fig. 2. The system has three integrated layers: Data Acquisition Layer, Fog Layer, and Cloud Layer. The data acquisition layer is the continuous audio and video recording from a surveillance camera in the classroom, which streams and saves recordings in real-time to the Fog Layer, where all the data processing and analysis takes place. In our video preprocessing pipeline, the first step is to extract frames and enhance them with OpenCV [7]. YOLO-NAS [1] is then used to identify and separate individual students in the classroom setting in real time. ByteTrack [48] is used to continuously track detected individuals across frames to maintain spatial-temporal coherence. Skeletal keypoints obtained using MediaPipe Pose are evaluated for the motion characteristics and posture adjustment for behavioral indications of distraction [21, 46]. If the movements are quick than the skeletal keypoints will catch gaze aversion and slouching. There will also be an audio stream for speech activity using Silero Voice Activity Detection (Silero-VAD) from the Silero Team [35]. Once the speech segments have been detected, we will infer emotional states based on embeddings from Wav2Vec2 [4]. To identify and interpret the emotional content contained in the audio input, the extracted embeddings are then processed and categorized using a Support Vector Machine (SVM) model [10]. These inputs from both modalities are fused and classified through a Bi-directional Long Short-Term Memory (BiLSTM) network [32] that results in a distraction score and measures attentiveness. The cloud layer then transmits the processed data and analyzed engagement reports for educators. These reports allow educators to implement timely interventions and inform pedagogical decisions based on evidence. The process flow of the proposed AI-TEACH framework is illustrated in Fig. 3. During the Training Phase, a customized training dataset based on classroom video footage is created that incorporates videos of each student member in the classroom exhibiting various distractions, such as looking away, excessive movement, yawning, or elevated vocal expressions. Both video and audio will be segmented by frames and pre-processed to enhance feature extraction. Deep learning and machine learning models will be trained in ways to encounter models, such as YOLO-NAS, MediaPipe Face and Pose, Silero-VAD, Wav2Vec2, and BiLSTM should then be trained separately or together to be able to detect and classify distractions. The Validation Phase represents live processing of audio and video streams from the classroom. The video processed will be segmented by frame, while each audio stream collected is also pre-processed. Once the videos are segmented, they can be processed for feature extraction with a focus on physical posture (sitting, standing, etc.); facial movements (eyes, mouth, eyebrows, etc.); and auditory (excitement, crying, disharmonic). These features are investigated to recognize indicators of distraction. If distraction is detected among students, the system calculates a distraction score from a distraction classifier via BiLSTM that uses the recorded behaviors’ intensity and frequency. The report generation phase begins with the system recording educator metadata and the duration of the session. After a session is completed, the framework combines all distraction indicators that were detected and calculates a final distraction score. A report is generated according to what is seen as engagement dynamics during the session. The report identifies potential actions that would support the educator where possible. This report is sent to the concerned instructor to help facilitate timely and data-informed pedagogical interventions. Data Acquisition Layer The data acquisition layer of the proposed AI-TEACH framework consists of surveillance cameras that have audio capturing capability that have been positioned so that all students stay in the cameras’ field of view, and through the course of the activities in the classroom. The cameras record video footage that is synchronized with the audio portion at a resolution of 640 × 640 pixels and 25–30 frames per second, so that the fidelity of the input is preserved, enabling real-time processing. The remote sensing ability of the camera enables the ongoing broadcasting of video and audio data to the processing units, allowing for an uninterrupted monitoring of student behavior and engagement. This robust architecture comprehensively records audio and video data in the classroom, allowing AI-TEACH to identify distraction cues, calculate distraction scores, and generate real-time engagement measurement, establishing the stage for what will happen in the fog layer and cloud layer. Fog Layer The fog layer in the AI-TEACH architecture is the real-time processing center that evaluates the video and audio streams from the classroom. This layer is designed for edge processing to provide low-latency analysis and decision-making without reliance on cloud-enabled inference. By combining deep learning, computer vision, and voice signal processing, it can integrate raw sensor data with higher-order instructional concepts. They can be synchronized and evaluated by parallel pipelines fusion with multiple modalities, such as audio input from classroom microphones and visual input from cameras. The video pipeline recognizes nonverbal behavioral indicators, including looking away, yawning, laying head on desk, and stillness, and uses classification techniques, including YOLOv8 for identifying students, MediaPipe for identifying landmarks, and rule-based classifiers for attention distraction detection with prior research. The audio pipeline evaluates cognitive-affective signals in parallel, focusing on voice activity detection (VAD), speech-emotion recognition (SER), and patterns of turn-taking or conversational dynamics. Every modality produces a set of distraction indicators, assigned to levels of severity. These are aggregated with respect to the reliability and importance of the respective modality using a weighted average. This provides an aggregated Distraction Score that represents the level of distraction throughout the class. This score, along with timestamped behavioral metadata, is securely sent to the cloud layer for storage and processing. It adds to the detailed session reports available to educators, offering actionable feedback on student attention levels/re-engagement strategies so that they may determine effective pedagogical interventions. The fog layer ensures that AI-TEACH is both responsive and scalable, while maintaining the promise of socially inclusive, evidence-based education that advances the social and economic equity objectives outlined in the SDG 4 Agenda. The computational requirements of the AI-TEACH architecture are distributed across its two processing tiers to balance latency, throughput, and energy efficiency. The fog layer requires hardware equivalent to or exceeding a NVIDIA Jetson Xavier NX (≥ 6 TFLOPS FP16, 8 GB RAM) to sustain frame-level video analytics and concurrent audio processing at the target frame rate. The cloud layer requires a server-class GPU equivalent to a NVIDIA Tesla T4 (16 GB VRAM) with a minimum upstream bandwidth of ≥ 10 Mbps to receive fog-layer metadata without bottlenecking the inference pipeline. Under these configurations, the system achieves an average end-to-end latency of approximately 350 ms from frame capture to distraction event logging, with a fog-node power consumption of approximately 15 W during continuous operation. Pre-Processing The incoming video stream is continually broken up into individual frames, primarily at a rate of 25 to 30 frames per second. To unify input dimensions for downstream processing modules, each frame is scaled to 640 × 640 pixels (as shown in Eq. 1). To improve model resilience across classrooms, color normalization and histogram equalization approaches are used to lessen variability in skin tones and lighting conditions. where \(\:{F}_{t}\) is the processed frame at time \(\:t\) and \(\:{V}_{t}\)is the raw frame. Video Analytics for Distraction Detection in AI-TEACH After the frame has been pre-processed, it is run through YOLOv8, a fast object identification model that separates and recognizes each pupil. To ensure temporal consistency between frames, each student is enclosed in a bounding box and given a unique ID using ByteTrack (as shown in Eq. 2). Each \(\:{b}_{i}\), or the ith bounding box, represents the pixel locations of the \(\:{i}^{th}\) student in the video. ByteTrack maintains a consistent student ID across frames as the student moves, thereby enabling the analysis of temporal behavioural patterns. The skeletal (pose) landmarks and facial landmarks are computed on each student’s pose and are the basis for understanding body language, gaze direction, and micro-expressions that could represent cognitive states. Pose key points \(\:{P}_{i}\) for each student are anatomical joints including the nose, eyes, shoulders, and wrists. Facial landmarks \(\:{L}_{i}\) has a total of 468 points made up from the mouth, eyes, and jawline. The facial and pose landmarks are computed frame-wise and saved to a temporal buffer for movement and posture analysis. After the pose is computed and landmarks are extracted, there are several characteristic distraction cues that may be produced as a result of combinations, geometric relationships, and temporal variation (as shown in Algorithm 1). For mobile phone usage, persistent hand-to-lap or hand-to-head motions were monitored as well as gaze estimation to determine if the student’s gaze was directed toward their lap or head area (as shown in Eq. 3). Here, \(\:{\overrightarrow{H}}_{wrist}\) and \(\:{\overrightarrow{H}}_{hip}\) denote wrist and hip key-points, respectively. Similarly, Facial landmark deviations are tracked to capture exaggerated expressions (as shown in Eq. 4). Here, \(\:{\sigma\:}_{exp}\:\) represents the standard deviation of facial landmark displacements (in 2D image pixel coordinates) over k consecutive frames, where \(\:k\) is the temporal window size (number of frames in the sliding window, empirically set to 10 frames at 25–30 fps). \(\:{L}_{face}\) is the temporal mean. A high \(\:{\sigma\:}_{exp}\) exceeding threshold \(\:{{\uptau\:}}_{exp}\) flags possible face-making or mimicry behaviour. Similarly, pretending to sleep is detected by analysing prolonged eye closure, chin tilt angle, and motionlessness (as shown in Eq. 5). Here, \(\:{\alpha\:}_{neck}\) is the neck-tilt angle (in radians) between the head axis and the torso axis, computed in the 3D skeletal coordinate space (x, y, z) derived from MediaPipe Pose landmarks, where the z-axis represents depth. Here, \(\:{v}_{neck}\in\:\mathbb{R^3}\) is the unit direction vector of the neck axis (from mid-shoulder to nose landmark) and \(\:{v}_{torso}\in\:\mathbb{R^3}\mathbb{\:}\)is the unit direction vector of the torso axis (from mid-hip to mid-shoulder landmark); both vectors are normalized prior to computing the dot product. The formula therefore computes \(\:arccos({v}_{neck}\:\cdot\:{v}_{torso}\:/\:\Vert\:{v}_{neck}\Vert\:\:\Vert\:{v}_{torso}\Vert\:)\), yielding a value in [0, π] radians. If the eye-aspect ratio remains below threshold \(\:{\tau\:}_{ear}\) for more than t seconds, and \(\:{\alpha\:}_{neck}\) indicates downward head tilt, the behaviour is classified as sleep simulation. Sudden arm extension combined with projectile motion near the detected student can indicate an object thrown in the class. Trajectories are extracted using Farnebäck optical flow across frames. In Eq. (6), if the arm motion vector \(\:{\overrightarrow{v}}_{throw}\) exhibits rapid displacement, and there is an associated fast-moving contour within the throw angle \(\:{\theta\:}_{throw}\)The event is flagged. The action of running out of class is detected through abrupt full-body displacement towards the classroom boundary. Here, \(\:{P}_{hip}\) denote the hip key-point at frame \(\:t\). If net displacement over a window of \(\:k\). If \(\:{\varDelta\:}_{exit}\) exceeds spatial threshold \(\:{\tau\:}_{exit}\), and the trajectory aligns with the door location map, an escape or exit is confirmed. Audio Analytics for Distraction Detection in AI-TEACH In the AI-TEACH fog layer, verbal distraction detection is enabled through real-time audio signal processing (as shown in Algorithm 2). Incoming audio streams are first segmented using the Silero Voice Activity Detection (VAD) model, which identifies speech and non-speech intervals with low latency. First, the detected speech segments are passed into a fine-tuned Wav2Vec2 model for acoustic embedding, which is then passed to an SVM-based classifier for cue-specific classification. This two-part pipeline preserves temporal accuracy and semantic distinction. Localized speech events are identified using beamforming or by cross-correlating microphones. In single-channel setups, overlapping low-amplitude utterances during ongoing instruction are identified. For each utterance segment \(\:{s}_{i}\), we compute \(\:{P}_{back-talk}\) (as shown in Eq. 8). Here \(\:SV{M}_{context}\) classifies intent based on spectral-temporal features. Repeated low-confidence speech outside the teacher’s directional audio profile suggests off-focus chatter. Similarly, using amplitude envelope tracking and spectral flatness measure (SFM), unusually high-energy or chaotic sound events (e.g., whistling, shouting, singing) are flagged. To check for abusive language in class, each transcribed utterance is checked for toxic or offensive content using an abuse classifier trained on profanity datasets. \(\:\text{T}\text{o}\text{x}\text{i}\text{c}\text{C}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{e}\text{r}\) is a binary classifier that is applied to the transciption, \(\:\text{T}\left({\text{s}}_{\text{i}}\right)\), of the \(\:{i}^{th}\) utterance. The decision rule of the binary classifier prescribes that indicator variable \(\:{I}_{abuse}\left({s}_{i}\right)\) is 1 if the \(\:{i}^{th}\) utterance is classified as abusive or \(\:{I}_{abuse}\left({s}_{i}\right)\) is 0 if the utterance is not classified as abusive (as shown in Eq. 9). Wav2Vec2 embeddings guide both the sentiment polarity and lexical analysis, and a high toxicity score triggers the abuse flag. Impulse-like acoustic signals with periodic temporal spacing and consistent spectral envelope are detected using autocorrelation and short-time energy (STE). Here, \(\:E\left(t\right)\) is the STE at time \(\:t\). High-energy bursts with quasi-regular periodicity (detected via autocorrelation peaks) are marked as tapping events. Their recurrence over a threshold rate \(\:{R}_{t}\) activates the cue. To detect offensive or sarcastic comments, speech segments are classified using emotion labels and linguistic features. The SVM classifier takes in Wav2Vec2 embeddings and parses affective tone (anger, sarcasm, hostility). Where the probability exceeds the threshold \(\:{\tau\:}_{off}\) marks the utterance as inappropriate. For each speech segment Si, a multimodal feature vector is created by concatenating the outputs of SVM-based classifiers, such as likelihood for back-talk, abuse, and offensive cues, with Wav2Vec2 embeddings along with low-level signal properties \(\:{\text{s}}_{\text{i}}\) (as shown in Eq. 12). Here \(\:{a}_{i}\) stands for manually created acoustic descriptors, \(\:{p}_{i}\) contains cue probabilities determined from SVM, and \(\:{w}_{i}\) is the Wav2Vec2 embedding. The input is constructed across a sequence of sliding windows of T segments \(\:X=[{x}_{1},\:\dots\:,\:{x}_{t}]\). This sequence is fed into a Bidirectional Long Short-Term Memory (BiLSTM) network to model temporal dependencies \(\:{\text{h}}_{\text{t}}\) (refer to Eq. 14). A SoftMax classifier y is applied to a pooled representation of hidden states using Eq. (15). Here \(\:W\) and b are learning parameters and \(\:\widehat{y}\in\:{\text{R}}^{\text{C}}\) is the predicted distraction category. Thus, BiLSTM enhances the system’s contextual awareness, enabling the identification of complex temporal patterns, such as escalating vocal distractions and recurrent distraction behavior. Distraction Score Generation Once distraction incidents are identified using audio and video analysis in the fog layer, they are then associated with a specific behavioural or verbal indicator, denoted with an Indicator ID (see Table 2). The indicators, which range from physical gestures like “using mobile phones” to verbal distractions like “abusive language”, each have a descriptive metadata profile associated with it, along with an empirically designed severity score depending on the level of classroom disruption. As the system processes the data stream in real time, it maintains a timeline of each indicator that it detects for future retrieval, consisting of indicator type, timestamp, frequency, and the source modality (i.e., audio or video) for each indicator. The fog layer will count the number of detected occurrences for each Indicator ID through the session and multiply it by the related severity score to produce a summary distraction score through calculating Eq. (16) (described in Algorithm 3). In this case, \(\:Occurrence{s}_{I{D}_{i}}\) is the amount of Indicator \(\:I{D}_{i}\) occurred and \(\:Scor{e}_{I{D}_{i}}\) is the fixed severity score associated with Indicator \(\:I{D}_{i}\), while \(\:N\) is the total number of unique distraction indicators detected in the session. Over all possible observations, this single score may provide a numerical measure of the level of classroom distraction severity. Along with the score all relevant metadata – such as Indicator IDs, timestamps, frequency counts and source modality – is also formatted into a standard metadata format. The composite score metadata is then securely sent to the cloud layer, archived, and later used for report generation related to the entire session. The cloud layer allows values to be presented on dashboards or to be archived in various analytics modules which provide the opportunity for educators to examine behavioural trends, correlate patterns of distraction and make pedagogical decisions to increase engagement and classroom behaviour. Cloud Layer In the AI-TEACH framework, the Cloud Layer offers data storage, reporting, and long-term analysis. The Cloud Layer operates off-site; this allows it to perform large-scale or retrospective evaluations. The Fog Layer enables the real-time processing of the participants’ actions in the classroom. It detects and scores indicators of distractions based on many forms of input, like audio and video. The architecture of the cloud layer is also designed to continue to persistently store metadata on a session basis, produce structured reporting, and scale longitudinal participant engagement tracking. Upon the completion of a classroom session, the fog layer transmits structured metadata to the cloud infrastructure, encapsulating multimodal distraction-related features. This package includes distraction indicator IDs, timestamps associated with them, frequency of occurrences, severity-weighted scores, and contextual tags based on behavioral and verbal analysis happening in the moment. All communication is sent over secure channels, and all messages are encrypted end-to-end to protect the client’s privacy and data integrity. Once the messages are received in the cloud service infrastructure, the platform initiates a pipeline for the aggregation of data and solves for overlapping events, aligned timestamped records, while normalizing input from multiple sources. The distraction patterns detected throughout the session are identified by parsing the resultant data. Each item will have a specific severity score and frequency of occurrence, as determined by a different indicator instance type. Because they are unique, quantifiable indicators of typical classroom attentiveness and behavioral regulation, these values are computationally weighted and summed to create the Class Distraction Score and the session-specific metric. This processed data is then utilized by the cloud-based reporting engine to automatically generate an interactive graphical dashboard (see Algorithm 4). This user interface starts with a clean, displayed header showing the date of the session and the associated unique session ID, followed by a distraction score dial presenting the computed distraction score visually, following the defined severity thresholds in color gradients. This visual representation of distraction levels supports leadership and teaching staff in the classroom to quickly assess distraction levels. The dashboard also includes the overall Total Distraction Score, which is located at the bottom of the indicator table, and is a summary metric for instructional planning. This feature allows educators to quickly visualize what is happening within their classroom, helps to promote informed pedagogical decision-making, and provides real-time feedback on how distracted students are overall. Each graphical report and raw metadata file is kept in a secure, scalable cloud format that allows for archival searches and periodic reading of classroom interactions. Sensitive behavioral data is available to authorized personnel, including teachers, school leaders, and approved researchers, via role-based access control (RBAC). In addition, the dashboard facilitates teachers’ adding contextual notes for each session, outlining the qualitative narratives within the data to enhance continuous model refinement and adaptive learning analytics. Having an end-to-end system supports a high-resolution behavioral archive, ongoing assessments of student engagement, and the ability to make evidence-based pedagogical decisions. Finally, the cloud-based aggregation allows for greater educational intelligence at multiple levels, supporting institutional audits, behavior analytics, and policy that aids progress of Sustainable Development Goal 4 (SDG 4) through inclusive, equitable high-quality education. In addition to developing the teacher’s greater awareness, the AI-TEACH model contributes to institutional efforts to provide inclusive, focused, and adaptable learning environments. Results and Analysis To determine the integrity of the proposed AI-TEACH framework, a complete series of experimental evaluations was carried out on real streaming data of a classroom surveillance video composed with psychological distraction and verbal distraction markers for audio. The model training base was constructed by gathering images from the data collected from open-source public datasets [13, 34, 39], all of which were made accessible through the Roboflow platform [11]. The images were categorized according to a predetermined set of class labels that were designed during the time the study was established. The final dataset contained over 5350 labeled images, and each image was curated with the intent of increasing the range, variety, and modes of distraction, which included mobile phone use, throwing objects, speaking out of turn, and the posture in which a learner slouched back, yawned, looked away, or rested their head. This basis for annotations was used to quantitatively evaluate the performance of the system, as well as qualitatively assess the classification performance of the model. A comprehensive collection of machine learning and computer vision libraries was used in the Python implementation of AI-TEACH. YOLO-NAS was utilized for high-precision, real-time human recognition and monitoring, and OpenCV [7] for frame-level video segmentation along with preprocessing. MediaPipe Pose [46] was used for skeletal key point extraction, allowing for fine-grained detection of motion-based distraction behaviors. The input video was gathered at 25–30 frames per second, and every single frame was examined to note and then categorize distracting events with inclusive time specificity. On the audio processing side, Silero Voice Activity Detection (VAD) was used to distinguish speech from noise in the background, and afterwards, feature extraction of contextual audio embeddings was done using a pre-trained Wav2Vec2 [4]. A Support Vector Machine (SVM) [10] implemented through scikit-learn [26] was used to classify the embeddings for identifying verbal distractions utilizing cues like loud talking, driving obnoxious comments, and depreciating language. Performance of the Video-Based Distraction Detection Module To score distractions and provide reports, the system produced annotated outputs with time-stamped information and visual overlays with bounding boxes and distraction class labels (shown in Fig. 4). In real-world classrooms, AI-TEACH, an explainable pipeline for distraction detection and scoring, shows great promise for scale implementation, allowing teachers to track student involvement in real-time and modify their teaching methods correspondingly. The detection results of the proposed AI-TEACH framework for video-based behavioral distraction detection are shown in Fig. 5. Figure 5 (a–c) displays the input video frames, which include scenes from a classroom with noticeable distraction signals. The corresponding position landmarks and annotated bounding boxes that show identified distracting behaviors are displayed in the processed outputs in Figs. 5 (d–f). These visual markers are superimposed on the frames by the algorithm to draw attention to postural cues linked to disengagement. The confusion matrix as a heatmap of the video-based distraction detection module in the AI TEACH system is shown in Fig. 6. Predicted distraction labels are along the column, and true labels are along the rows. It shows more than six separate distraction behaviors including mobile phone distractions, changes in facial expression, simulated sleep, an object being thrown, and student unreasonable leaving the classroom, to assess the model classification performance. The dark cells going on the diagonal indicates the location of the correctly classified observations where they contribute to the heatmap values, which overall gives us an impression of where the model is making salient predictions across various distracting behaviors. Overall, the heatmap nicely depicts the model’s ability to differentiate across the distractive behavior indices based upon the skeletal key points from the MediaPipe Pose, which captures the posture and gestures of distraction. Table 3 displays the performance measures for the proposed AI-TEACH architecture and indicates superior classification tendencies across several modelling iterations. The model proved its effectiveness at detecting and distinguishing between distracted behavior versus normal behavior; the model achieved a weighted average recall of 92%, which means that it captured most, if not all, of the distraction signals in the real-world. Moreover, it also attained a macro-average F1 score of 92%, which speaks to its consistency, across all the distraction categories, regardless of class frequency and variation in class sizes, and put basic classification bias out of commission. The model’s performance accuracy was consistent at 90% to 94% accuracy at detecting complex behavioral indicators and immediate postures, and arguably at its optimum outcome yielded a weighted average F1score of 0.98. Overall, the system executed well in real-world situations, achieving an F1 score of 0.93 and recall of 0.92. The support values and number of labeled distraction instances for each iteration provided support ranging from 11 to 41, which also represents enough statistical validity. Macros and weighted averages were employed to evaluate the generalizability of the model and class-based reliability. The weighted average gives emphasis to the model’s consistency under class imbalance situations, which is common in real classroom data; whereas the macro average affirms that all classes were fairly represented. All this evidence supports the conclusion that AI-TEACH can identify distractions accurately and in real time, suggesting the potential for broad acceptance and implementation in learning contexts. Audio-Based Distraction Detection The audio-based distraction detection module of the proposed AI-TEACH system uses a combination of vocal activity detection and contextual speech categorization to process classroom speech recordings to detect verbal distractions and generate score reports. For every recognized speech cue, annotated outputs with time-stamped metadata and associated distraction indicator labels are generated. Figure 7 depicts a heatmap of audio distraction events for five major verbal indicators: chatting behind the teacher’s back, humorous or loud sounds, abusive language, tapping or drumming, and inappropriate comments. The heatmap shows the temporal “hot zones” during a 40-minute class period where distraction indications are most noticeable. The color scales represent the volume of detected audio events, revealing periods of potential high verbal distraction to the instructors. Figure 8 illustrates the confusion matrix showing the classification performance of the audio-based module. The columns show the predicted labels, while the rows show the actual labels pertaining to the distraction. The heatmap accurately classifies speech distraction cues, giving priority to high-intensity values along the diagonal. The distribution of values in the matrix shows that the system was able to classify multiple forms of audio-based distractions, specifically those with definite acoustic features (tapping, volume, etc.). The overlapping speech categories can be expected to be misclassified and are a common issue in noisy classroom settings. In general, the performance of the entire system remained stable and reliable. Table 4 presents a summary of the classification metrics for the audio-based module. The overall accurate results for audio recognition system ranged from 88% to 94%. In one round, the highest accuracy achieved was 94% and an f1 score of 0.95. The balanced use of verbal distraction classes by the model is substantiated by the macro average F1 score of 92%, where the wayward average F1 score of 92% further supports the reliable performance of the model in the case of class imbalance. To achieve statistical reliability across the assessment rounds, there needed to be a minimum number of audio distraction events that were tagged, ranging from 17 to 36 each iteration. Figure 9 provides an example of a distraction score report produced by the AI-TEACH platform. The report integrates vital information about distraction in the classroom, including the session ID, the date of the session, and the distraction score, which was calculated from the video and audio analytics collected in real-time. The Class Distracting Score (normalized from 0 to 100) is a standardized measure of classroom attention derived from the frequency and severity of the distracting behaviors that were observed. The report itemizes specific distraction indicators (cell phone use, vocal disruptions, physical disengagement) and includes a severity score for each indicator, as well as the frequency of each indicator. This report can help teachers identify the patterns of disengagement that are most prevalent in their classrooms and in what ways they affect engagement in this classroom. By transforming raw behavioral data into meaningful measures, the AI-TEACH platform offers another scalable system for ongoing engagement monitoring and useful data-informed decision-making for academic administrators and teachers. To evaluate the effectiveness of the proposed AI-TEACH framework, a comparative study was conducted against different state-of-the-art distraction and behavior detection models commonly used in affective computing and classroom analytics. A multi-modal dataset, uniformly annotated, allows for the comparative support of deep learning architectures and to compare several traditional machine learning approaches in the same experimental context. To assess performance as a comprehensive evaluation of model performance, precision, recall, and F1 score (commonly accepted metrics in a classification procedure) were used. The AI-TEACH framework outperformed conventional models such as convolutional neural networks (CNN), long short-term memory (LSTM), and feature-based approaches like HOG + SVM, and see Table 4 for a summary of this performance. The system supports real-time processing of audio and video inputs through the combination of YOLO-NAS object detection, MediaPipe Pose for posture tracking, Wav2Vec2 embeddings for extracting features from speech, and SVM for final classification. AI-TEACH demonstrated a high level of accuracy at identifying the indications of distraction in behavioral and verbal modalities, achieving an overall F1 score of 93.50% (Fig. 10). The results reflected on the architecture, by demonstrating the validity of the modular, multi-staged inference architecture that allowed for timely scoring, accurate identification, and reporting. It provides evidence of the framework’s scalability for applications in real education contexts while providing educators with realistic advice on fostering inclusive, engaged, data-rich learning environments. The current study provides a comprehensive experimentation to assess the AI TEACH framework’s capabilities to detect in real-time classroom distractions by processing audio-visual data. The findings demonstrate that the framework is sufficiently accurate and flexible to detect many types of distraction. All audio and video modules consistently achieved high accuracy and recall scores across several tests. The final system outputs are practical and easily accessible in the form of annotated video frames, confusion matrices, heatmaps, and distraction score distributions at the session level, all of which provide unambiguous, actionable insight for researchers and educators. The audio-visual and/or audio-only modules performed well during this evaluation and exhibit numerous advantages over conventional baselines; for example, the extent to which the method is reliable in unpredictable, dynamic, and noisy classrooms. In addition, the structure of report production and cloud-based visualization allows for reports to be actionable and scalable for the long-term monitoring of academic achievement. These results support the potential of AI-TEACH as a viable, deployable, and available option to assist teachers with data-informed decision-making, while supporting the goal of improving student engagement, social-emotional-disruptive behaviors, and promoting equitable and successful learning, as reflected in the Sustainable Development Goals for education (SDGs). To assess the effectiveness of the Artificial Intelligence Technology for Engaging Students in Classroom Habitats (AI TEACH) framework in a real-world context, 200 students were randomly assigned to either the experimental group or control group for the study of student engagement. Academic achievement levels were defined at baseline in both groups with a pretest prior to the intervention. The average scores are reported as \(\:pr{e}_{cntrl}\:\)and \(\:pr{e}_{exp}\) for the control and experimental groups, respectively. During the post-intervention stage, the control group experienced traditional instruction while the experimental group received instruction by teachers using AI-TEACH, which also offered real-time feedback during instruction. After the instruction period, a post-test was given with mean scores designated as \(\:pos{t}_{cntrl}\) and \(\:{post}_{exp}\). A two-sample t-test was executed using the equation to determine whether the score increase from the pre-test to the post-test was statistically significant between the two groups (as shown in Eq. 17). Here, \(\:{i}_{exp}^{2}\) and \(\:{i}_{cntrl}^{2}\) represent, respectively, the differences in score improvements between the experimental and control groups. Using IBM SPSS, the t-value was compared to the critical value at significance level α = 0.05 with degrees of freedom \(\:df={n}_{exp}+{n}_{ctrl}-2=198\) [23]. Cohen’s d [9] was calculated to measure the intervention’s magnitude using Eq. (18). The experimental and control groups achieved gains of 9 and 5, respectively, with variances \(\:{i}_{exp}^{2}=20\) and \(\:{i}_{cntrl}^{2}=24\). The effect size can be calculated as: Using Cohen’s standards, this finding indicates a large effect size (small: d ≈ 0.2, medium: d ≈ 0.5, and large: d > 0.8). The findings suggest that AI-TEACH has a large effect on the dissemination of education because it has allowed educators to make positive and dynamic adjustments in their teaching practice in real-time. Therefore, the educational advantages of employing AI-TEACH in the classroom settings were further confirmed as students in the experimental group substantially outperformed students in the control group regarding their academic results. Discussion Distraction is recognized as one of the most common and unsolved problems in the contemporary classroom, resulting in reduced student engagement, reduced cognitive processing, and obstruction of instruction [24, 38]. Past studies have identified distraction as being problematic for student retention and student academic success. Real-time identification of distraction depends on instantaneous monitoring, procrastinated action evaluation, or retrospective evaluation via self-reporting [8, 22, 40]. The proposed AI-TEACH framework proposes to satisfy this need through the utilization of multi-modal AI to recognize and rate classroom distractions and report findings through real-time audio and video analysis. The system utilizes a fog-computing functionality to provide edge-level inference while the system processes ongoing ambient audio streams and video surveillance streams in a subtle manner. Providing the platform with the ability to perceive a wider array of behavioral distractions and verbal distractions with very little impact on the natural flow of the classroom. The experimental results demonstrate that AI-TEACH performs robustly and consistently across both its video-based and audio-based distraction detection modules. The video module achieved a weighted average recall of 92% and overall accuracy of 90% across nine evaluation iterations, with a macro-average F1 score of 92% confirming balanced performance across both frequent and infrequent distraction categories. The audio module achieved comparable results, with accuracy ranging from 88% to 94% and a macro-average F1 score of 92%, indicating reliable classification of verbal distraction cues even in the presence of overlapping speech and ambient classroom noise. These results compare favorably against established baseline architectures: AI-TEACH’s overall F1 score of 93.5% substantially outperformed CNN (74.14%), LSTM (76.04%), BiLSTM alone (85.95%), HOG + SVM (80.52%), and OpenPose + SVM (83.26%), as shown in Table 4. The performance advantage of AI-TEACH over these baselines stems from three architectural features that individually existing systems lack: the temporal modeling capacity of the BiLSTM fusion layer, the complementarity of behavioral and verbal distraction signals captured across modalities, and the structured severity weighting that allows the system to distinguish high-impact events from low-level disruptions rather than treating all distraction categories equally. These findings are consistent with the broader multimodal emotion and behavior recognition literature. Prior work has shown that per-sample optimization during multimodal training prevents dominant modalities from suppressing weaker signals (Sun et al., 2023), and that adversarial fusion strategies improve robustness under modality-specific degradation (Liu et al., 2024) patterns reflected in AI-TEACH’s relatively stable performance across audio and video modules despite the greater environmental variability of audio signals in classroom settings. The confusion matrix analysis reveals that visually salient behaviors such as mobile phone use and head-on-desk posture were classified with high spatial precision, while audio-based categories such as back-talk and tapping exhibited strong diagonal dominance in classification indicating that the Wav2Vec2-SVM pipeline effectively distinguished these acoustically distinct events even under noisy conditions. The controlled classroom experiment further supports these findings at the pedagogical level. Among the 200 participants divided into experimental and control groups, students in the AI-TEACH-supported experimental group achieved a mean score gain of 9 points compared to 5 points in the control group, yielding a Cohen’s d of approximately 0.90, a large effect size by conventional standards [9]. This outcome is consistent with research demonstrating that timely, data-informed instructional feedback improves pedagogical responsiveness and, consequently, learning outcomes [42, 44]. It also aligns with Liao and Wu’s (2022) finding that objectively measured engagement is a stronger predictor of academic performance than self-reported distraction, reinforcing the value of an automated, objective monitoring system over traditional observational approaches. Traditional distraction monitoring relies on teacher observation, student self-report, or post-lesson assessment, all of which share fundamental limitations of subjectivity, temporal delay, and scalability [17]. AI-TEACH addresses these limitations by operating as a co-teacher: a continuous, objective, and non-intrusive partner that augments rather than replaces educator judgment. Unlike systems that simply flag distraction events without context, AI-TEACH translates raw behavioral and verbal signals into structured, severity-weighted engagement metrics that educators can act on immediately or review longitudinally. The session-wide Class Distraction Score, delivered through the cloud-based dashboard after each session, enables educators to identify which distraction categories were most prevalent, at which points in the session they clustered, and how overall engagement compared across sessions. This supports reflective teaching practice and targeted instructional adjustment, for example, recognizing that distraction peaks during particular lesson segments may indicate the need for a change in instructional format or pacing [15]. In terms of how educators interact with the system, AI-TEACH is designed to present outputs through an intuitive visual dashboard that does not require technical expertise to interpret. The dashboard displays the total distraction score, a breakdown of specific distraction indicators with severity scores and frequencies, and a temporal heatmap of distraction events across the session. Teachers are not expected to act on individual incident alerts in real time; rather, the system is designed to support post-session reflection and planning. Educators can also add contextual notes to each session within the dashboard, allowing them to annotate the quantitative data with qualitative observations for example, noting that a particular student’s restlessness was related to a known external circumstance, which is important for appropriate interpretation of the system’s outputs. This contextual override capacity is discussed further in Sect. 5.4. With respect to teacher professional development, AI-TEACH contributes to the kind of evidence-based reflective practice that research identifies as a driver of instructional improvement [42]. By providing structured behavioral data over time, the system enables educators to examine how changes in lesson design, classroom arrangement, or instructional strategy correlate with changes in engagement patterns a feedback loop that can make professional development more concrete, data-grounded, and self-directed. Training requirements for teachers are minimal: the system requires no configuration during a session and delivers its outputs in plain visual form. Initial onboarding, covering dashboard navigation and interpretation of distraction categories and severity scores, is estimated to require approximately one to two hours. A critical limitation that educators and system designers must be aware of concerns the behavioral diversity of student populations. Several of the behavioral indicators monitored by AI-TEACH, including fidgeting, repetitive movement, gaze aversion, and physical restlessness may be associated with neurodevelopmental conditions such as autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), and other forms of neurodiversity. For students with these conditions, such behaviors may represent natural self-regulatory strategies rather than indicators of disengagement from learning and flagging them as distraction events risks misrepresentation of student engagement and potential stigma. AI-TEACH is explicitly designed as a class-level monitoring tool, generating aggregate session metrics rather than individualized diagnostic assessments. The severity scoring system is calibrated to overall classroom disruption impact behaviors that affect the learning environment as a whole rather than to individual student cognitive states. Nevertheless, educators must be trained to interpret system outputs with this distinction in mind. A student who consistently generates distraction flags due to fidgeting, for example, should not have those flags treated as evidence of disengagement if other indicators suggest active participation. The dashboard’s contextual note feature provides a mechanism for educators to record such nuances and to manually disregard specific indicators for identified students when warranted. The system does not provide diagnostic labels, clinical recommendations, or individualized behavioral profiles; it is a classroom management support tool, and its outputs must be interpreted by educators with appropriate professional judgment. Future iterations of AI-TEACH should incorporate user-configurable behavioral thresholds that allow educators to suppress or downweight specific distraction indicators for individual students, for example, disabling the fidgeting/tapping cues for a student with a documented neurodevelopmental condition. This would substantially improve the system’s inclusivity and reduce the risk of inappropriate interpretation in heterogeneous classrooms. The deployment of AI-based surveillance in educational settings raises substantive ethical concerns that must be addressed explicitly. AI-TEACH collects continuous audio and video data from classrooms, including data about minors in many deployment contexts. All data transmission is encrypted end-to-end, and access to behavioral data and session reports is governed by role-based access control (RBAC), restricting visibility to authorized personnel classroom teachers, school administrators, and approved researchers only. Raw audio and video recordings are processed at the fog layer and are not transmitted to the cloud; only structured metadata, distraction indicator IDs, timestamps, and severity scores are stored remotely, minimizing the volume of sensitive data in long-term storage. Algorithmic bias is a significant concern in any automated system that makes inferences about human behavior. AI-TEACH’s training dataset, while curated to span a range of distraction behaviors and classroom contexts, may not adequately represent the full diversity of student populations in terms of skin tone, body type, cultural behavioral norms, and linguistic variation. Postural and gestural norms vary across cultural contexts, and behaviors that are considered disruptive in one cultural setting may be normative in another. The audio-based classifiers, trained on English-language profanity and tonal classifiers, may underperform or generate false positives in multilingual or non-English classroom environments. These are recognized limitations, and they imply that AI-TEACH should not be deployed in culturally or linguistically novel contexts without retraining or recalibration of the relevant classification modules. Transparency in governance is essential. Schools and institutions adopting AI-TEACH should establish clear policies regarding data retention periods, access rights, the purposes for which distraction data may be used, and the rights of students and parents to be informed about monitoring practices. Where participants include minors, informed consent procedures must include parental or guardian consent in addition to student assent, and the specific nature of what is being monitored, how data is stored, and who has access must be communicated clearly. Ethical approval for research deployments should be obtained from relevant institutional review boards, and ongoing audits of system outputs for evidence of differential performance across student demographic groups should be conducted as part of responsible deployment practice. The experimental findings provide empirical support for several of the SDG connections introduced in Sect. 2. With respect to SDG 4 (Quality Education), the controlled classroom experiment demonstrated a statistically significant and practically meaningful improvement in academic achievement for students in the AI-TEACH-supported group (Cohen’s d ≈ 0.90), supporting the claim that real-time engagement monitoring can contribute to improved learning outcomes aligned with Target 4.1. The system’s capacity to surface differential engagement patterns across student groups without relying on subjective teacher observation is relevant to Target 4.5, which addresses equitable access to quality education for disadvantaged groups, by reducing the potential for bias in identifying disengaged learners. The co-teacher paradigm, in which AI-TEACH generates session-level behavioral data that feeds into teacher reflection and instructional adjustment, supports Target 4.c by contributing to the continuous professional development of educators through evidence-based feedback. With respect to SDG 3 (Good Health and Well-being), the system’s detection of behavioral indicators such as prolonged head-on-desk posture and signs of fatigue can alert educators to students who may be experiencing physical or emotional distress, enabling empathetic responses. However, it must be emphasized that AI-TEACH does not diagnose health or psychological conditions and should not be used as a substitute for professional health screening or referral processes. With respect to SDG 16 (Peace, Justice and Strong Institutions), the real-time detection of abusive language and hostile verbal interactions supports proactive behavioral management and contributes to safer, more inclusive classroom environments provided that such detections are responded to with appropriate educator judgment rather than automated disciplinary action. Several limitations of the current study warrant explicit discussion. First, the training dataset, comprising approximately 5,350 labeled images drawn from three open-source repositories, may not fully represent the behavioral diversity of global classroom populations. The prevalence of specific distraction behaviors in the training data including the relative frequency of mobile phone use, object throwing, and verbal disruptions should be examined in relation to the distraction patterns observed in the 200-student classroom experiment, and discrepancies between dataset-derived and real-world behavioral distributions should be investigated in future work. Second, the 200-student controlled experiment was conducted in a single institutional context; the educational level, age range, cultural background, and socioeconomic characteristics of participants may limit the generalizability of the academic achievement findings to other settings. Third, the current system does not capture physiological distraction signals such as heart rate variability or galvanic skin response, which would provide access to internal cognitive and emotional states not visible through behavioral observation alone. Fourth, the validity of claims about broader educational impact including long-term retention, dropout reduction, and equity outcomes requires longitudinal empirical support beyond what the current study provides and should be treated as directional hypotheses for future investigation rather than established findings. Fifth, performance under conditions of severe occlusion, highly overlapping speech, and extreme ambient noise remains a challenge, and the system’s robustness in such conditions should be evaluated more systematically in future deployments across varied classroom environments, lighting conditions, and room layouts. Conclusion Student distraction continues to be a longstanding problem in many modern classrooms and as described throughout this paper, the AI-TEACH architecture represents an innovative and pragmatic approach to resolving this issue. AI-TEACH can utilize advanced computer vision and speech analytics available in a modular fog-cloud architecture to facilitate the real-time detection, categorization, and scoring of both verbal and behavioral distraction indicators. AI-TEACH allows for reliable, data-informed decisions to minimize distractions and promote instructional interventions that enhance classroom focus, with an overall accuracy of 90% and 92% verified recall. More importantly, AI TEACH promotes an inclusive, equitable, and high-quality educational experience, thereby contributing significantly to the advancement of several Sustainable Development Goals (SDGs), particularly SDG 4. AI-TEACH democratizes educational input and modernizes classroom surveillance in large or resource-poor learning environments. The following avenues for further development are suggested to increase the AI-TEACH framework’s potential and social impact: - To obtain more complex markers of cognitive load and emotional engagement, use physiological data, such as heart rate variability and galvanic skin response. Include feedback loops that will inform teachers and students, in real time, when distracted patterns occur over time. - Develop multilingual emotion and voice recognition modules to address potential cultural bias across language classrooms. You should be able to increase the resilience of the model using transformer-based spatiotemporal learning as well in visually dynamic or occluded environments. - Provide teachers with adaptable dashboards, showing trend predictions, longitudinal data, and ways to prompt educational responses. Use federated educational infrastructures to allow for decentralized training from a variety of institutions, and to retain data privacy. - Develop large trial deployments in diverse socioeconomic and geographic settings to evaluate educational value, equity, and scalability. These future paths can help AI-TEACH develop into a comprehensive educational intelligence system that can change traditional classrooms into data-rich, emotionally intelligent, and adaptable learning environments. Data Availability All the datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. References Aharon S, Louis-Dupont,., Baratz O, Masad O, Yurkova K, Fridman L, lkdci ., Khvedchenya E, Rubin R, Bagrov N, shairoz-deci ., avideci, Tymchenko ., Shamir B, Bochman J, Keren D, Zhilko T, Perl A, Baruch S…B. S. (2023). Deci-AI/super-gradients: 3.0 .8 (3.0.8) [Software]. Zenodo. https://doi.org/10.5281/zenodo.7789328 Allison BR, Dickay SA. School environment and the Sustainable Development Goals (SDG) for greater performance among staff and students. IOSR-JHSS. 2020;25:33–9. Asish SM, Kulshreshth AK, Borst CW. Detecting distracted students in educational VR environments using machine learning on eye gaze data. Computers Graphics. 2022;109:75–87. Baevski A, Zhou Y, Mohamed A, Auli M. wav2vec 2.0: A framework for self-supervised learning of speech representations. Adv Neural Inf Process Syst. 2020;33:12449–60. Betto I, Hatano R, Nishiyama H. Distraction detection of lectures in e-learning using machine learning based on human facial features and postural information. Artif Life Robot. 2023;28(1):166–74. Bian Y, Zhou C, Liu J, Geng W, Shi Y. The effect of reducing distraction on the flow-performance link in virtual experiential learning environment. Virt Real. 2022;26(4):1277–90. Bradski G. The opencv library. Dr Dobb’s Journal: Softw Tools Prof Program. 2000;25(11):120–3. Chen MH, Agrawal S, Krishna SM, Lan PY. (2025). How does innovative teaching influence learning outcome via community of inquiry model? A time-lagged study. Behav Inform Technol, 1–18. Cohen J. Statistical power analysis for the behavioral sciences. routledge; 2013. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273–97. Dwyer B, Nelson J, Hansen T. & others. (2024). Roboflow (Version 1.0) [Software]. https://roboflow.com Er-radi H, Touis B, Aammou S. Machine learning in adaptive online learning for enhanced learner engagement. Technological Tools for Innovative Teaching. IGI Global Scientific Publishing; 2024. pp. 43–63. Harish T. (2024). Students behaviors detection [Open source dataset]. Roboflow Universe. https://universe.roboflow.com/harish-t-rimzy/students-behaviors-detection-wkavr-koe3g Ialuna F, Civitillo S, Jugert P. Culturally responsive teaching, teacher-student relationship and school belongingness: a multi-informant study in ethnically diverse classrooms. Learn Cult Soc Interact. 2024;47:100839. Ives J, Rieger G, Renani FR. Requiring mobile devices in the classroom: the use of web-based polling does not lead to increased levels of distraction. J STEM Educ Res. 2024;7(2):307–23. Kaewkaisorn K, Pintong K, Bunyang S, Tansawat T, Siriborvornratanakul T. Student attentiveness analysis in virtual classroom using distraction, drowsiness and emotion detection. Discover Educ. 2024;3(1):56. Kafaee M, Kheirkhah MT, Balali R, Gharibzadeh S. Conflict of interest as a cognitive bias. Account Res. 2022;29(6):379–96. Kaushik MB, Doshi BM, Ranjan SK. (2024). Unravelling the Attention Crisis: Investigating Student Engagement in Contemporary Classrooms. Le MH, Doan TM, Nguyen DD, Nguyen MS. (2022). Smart Desk in Hybrid Classroom: Detecting student’s lack of concentration when studying. In 2022 9th NAFOSTED Conference on Information and Computer Science (NICS) (pp. 13–18). IEEE. Liao CH, Wu JY. Deploying multimodal learning analytics models to explore the impact of digital distraction and peer learning on student performance. Comput Educ. 2022;190:104599. Lugaresi C, Tang J, Nash H, McClanahan C, Uboweja E, Hays M, Grundmann M. (2019). Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172. Meng L, Tang C, Yang X. (2025). Evaluating the emotional impact of virtual expressions: ERPs analysis of smart agent appearance. Behav Inform Technol, 1–16. Meyers LS, Gamst GC, Guarino AJ. Performing data analysis using IBM SPSS. John Wiley & Sons; 2013. Nabung A. The impact of multitasking with digital devices on classroom learning: A critical review on the future of digital distraction in education. US-China Educ Rev. 2024;14(6):369–83. Nanda Kishore MS, Jayakumar SKV, Reddy S, Dhavachelvan G, Chandramohan P, D., Reddy S. N. P. (2011). Web service suitability assessment for cloud computing. In International Conference on Web and Semantic Technology (pp. 622–632). Berlin, Heidelberg: Springer Berlin Heidelberg. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Duchesnay É. Scikit-learn: Machine learning in Python. J Mach Learn Res. 2011;12:2825–30. Rao KP, Janet B. (2022). Teacher assistance system to detect distracted students in online classroom environment. In 2022 4th International Conference on Smart Systems and Inventive Technology (ICSSIT) (pp. 1743–1749). IEEE. Rojas JC, Faure J, Barril JP, Jara RF, Zambrano VV, Almuna J, Salter P. Informal learning experiences in the construction and development of professional teacher identity. Learn Cult Social Interact. 2025;51:100893. Saini M, Sengupta E, Sharma N. (2025). AI-assisted hand gestures based smart feedback system for educators. Educ Inform Technol, 1–28. Saini M, Sengupta E, Singh M, Singh H, Singh J. Sustainable Development Goal for Quality Education (SDG 4): A study on SDG 4 to extract the pattern of association among the indicators of SDG 4 employing a genetic algorithm. Educ Inform Technol. 2023;28(2):2031–69. Scheiman M, Rouse MW. Optometric management of learning-related vision problems. Elsevier Health Sciences; 2006. Schuster M, Paliwal KK. Bidirectional recurrent neural networks. IEEE Trans Signal Process. 1997;45(11):2673–81. Sengupta E, Saini M, Singh M, Singh J. (2022, April). An Exploration into Artificial intelligence based advancement in education field. In 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE) (pp. 1250–1255). IEEE. Shafqat A. (2025). Classroom behavior detection [Open source dataset]. Roboflow Universe. https://universe.roboflow.com/asad-shafqat/classroom-behavior-detection Silero Team. (2024). Silero VAD: Pre-trained enterprise-grade Voice Activity Detector (VAD), Number Detector and Language Classifier [Computer software]. GitHub. https://github.com/snakers4/silero-vad Strachan R, Liyanage L. (2015). Active student engagement: The heart of effective learning. Global Innov Teach Learn High education: transgressing boundaries, 255–74. Su MC, Cheng CT, Chang MC, Hsieh YZ. A video analytic in-class student concentration monitoring system. IEEE Trans Consum Electron. 2021;67(4):294–304. Tatlı C, Karadağ M. Reducing public speaking anxiety through the use of virtual reality assisted distraction: a systematic pilot study. Behav Inform Technol. 2024;43(13):3061–70. Topicos II. (2023). Classroom dataset [Open source dataset]. Roboflow Universe. https://universe.roboflow.com/topicos-ii-cm0i3/classroom-dataset-2pqda Trabelsi Z, Alnajjar F, Parambil MMA, Gochoo M, Ali L. Real-time attention monitoring system for classroom: A deep learning approach for student’s behavior recognition. Big Data Cogn Comput. 2023;7(1):48. Wang R, Xu D, Cascone L, Wang Y, Chen H, Zheng J, Zhu X. (2025). Raft: robust adversarial fusion transformer for multimodal sentiment analysis. Array, 100445. Wang SK, Hsu HY, Reeves TC, Coster DC. Professional development to enhance teachers’ practices in using information and communication technologies (ICTs) as cognitive tools: Lessons learned from a design-based research study. Comput Educ. 2014;79:101–15. Wang W. Application of deep learning algorithm in detecting and analyzing classroom behavior of art teaching. Systems and Soft Computing. 2024;6:200082. Woolverton GA, Pollastri AR. An exploration and critical examination of how intelligent classroom technologies can improve specific uses of direct student behavior observation methods. Educational Measurement: Issues Pract. 2021;40(3):7–17. Xiang J, Zhu X, Cambria E. (2025). Integrating audio–visual text generation with contrastive learning for enhanced multimodal emotion analysis. Inform Fusion, 103809. Zhang F, Bazarevsky V, Vakunov A, Tkachenka A, Sung G, Chang CL, Grundmann M. (2020). Mediapipe hands: On-device real-time hand tracking. arXiv preprint arXiv:2006.10214. Zhang X, Zhang Y. Exploring the collective agency of postgraduate students in an academic study group: a social network perspective. Learn Cult Soc Interact. 2025;52:100908. Zhang Y, Sun P, Jiang Y, Yu D, Weng F, Yuan Z, Wang X. (2022). Bytetrack: Multi-object tracking by associating every detection box. In European conference on computer vision (pp. 1–21). Cham: Springer Nature Switzerland. Zhao Y, Cambria E, Zhu X. (2026). TEMPO: Training-time Equilibration of Modalities for Per-sample Optimization in Multimodal Sentiment. IEEE Trans Affect Comput. Zhu X, Liu Z, Cambria E, Yu X, Fan X, Chen H, et al. A client–server based recognition system: non-contact single/multiple emotional and behavioral state assessment methods. Comput Methods Programs Biomed. 2025;260:108564. Zhu X, Wang Y, Cambria E, Rida I, López JS, Cui L, Wang R. RMER-DT: robust multimodal emotion recognition in conversational contexts based on diffusion and transformers. Inform Fusion. 2025;123:103268. Acknowledgements Not Applicable. Funding None. Author information Authors and Affiliations Contributions Munish Saini and Harsh Sharma contributed to the conceptualization, methodology development, and data collection and wrote the main manuscript text. Eshan Sengupta contributed to conceptualization, technical writing, visualization, and manuscript review. All authors reviewed and approved the final manuscript. Corresponding author Ethics declarations Competing interests The authors declare no competing interests. Additional information Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. About this article Cite this article Saini, M., Sharma, H. & Sengupta, E. Artificial Intelligence Based Framework for Student Engagement Assessment in Classroom Environments. Cogn Comput 18, 108 (2026). https://doi.org/10.1007/s12559-026-10629-z Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s12559-026-10629-z