Simulation-Based Evaluation of Artificial Intelligence-Integrated Visual Teaching Design for Mental Health Promotion in College Music Courses

Summary

This study proposes a graph-based emotional curriculum intelligence framework for AI-driven visual teaching in college music education. Using a simulation-based dataset, the framework models knowledge acquisition, emotional dynamics, and engagement. Results indicate improved learning efficiency and stress reduction compared to conventional models, demonstrating potential for adaptive, mental health-aware educational systems.

Abstract

Visual teaching design in college music education has attracted increasing attention because of its potential to enhance student engagement and promote mental well-being. However, conventional instructional approaches often lack the adaptability needed to respond to learners' evolving cognitive and emotional states, limiting their effectiveness in supporting both academic achievement and psychological balance. The main aim of this simulation-based study is to develop and evaluate an artificial intelligence-driven framework that integrates visual teaching strategies with graph-based emotional modeling to optimize learning outcomes in college music courses. A novel graph-based emotional curriculum intelligence framework (GECIF) is proposed, representing the learning environment as an interconnected graph of knowledge, activity, and emotional nodes. The framework incorporates graph-based representation learning, dynamic state modeling, and adaptive learning mechanisms to simulate the evolution of knowledge acquisition, engagement, emotional stability, and stress over time. A synthetic dataset comprising interaction logs, emotional indicators, and learning performance variables was generated through a controlled MATLAB-based simulation pipeline to ensure reproducibility. GECIF was evaluated against a conventional non-adaptive baseline model under identical experimental conditions across normalized simulation steps. Performance was assessed using normalized simulation metrics, including knowledge gain, emotional stability, stress reduction, engagement level, and robustness. The proposed framework consistently outperformed the baseline, demonstrating higher simulated knowledge acquisition, improved emotional regulation, greater engagement, reduced simulated stress, and enhanced robustness. Statistical analysis based on mean values, standard deviations, and 95% confidence intervals confirmed consistent improvements across all evaluation metrics. These findings demonstrate the potential of AI-driven, emotion-aware visual teaching strategies for adaptive music education while providing a scalable and reproducible computational framework to support future research in intelligent educational systems.

Introduction

The integration of artificial intelligence (AI) into educational systems has significantly transformed traditional teaching paradigms, enabling more adaptive, personalized, and data-driven learning experiences^1,2,3. In recent years, visual teaching design has emerged as an effective pedagogical approach, particularly in disciplines such as music education, where abstract concepts, emotional expression, and perceptual understanding play a central role^4,5,6,7. Visual representations, including graphical interfaces, interactive simulations, and multimedia elements, facilitate deeper cognitive processing and enhance learners’ engagement^8,9,10,11. However, despite these advancements, existing visual teaching methods often remain static and lack the capability to dynamically respond to learners’ evolving cognitive and emotional states^12,13,14. Music education, especially at the college level, presents unique challenges due to its inherently multidimensional nature^15,16,17. Learning outcomes are influenced not only by cognitive understanding of musical theory and performance techniques but also by emotional factors such as motivation, stress, engagement, and psychological well-being¹⁸. Previous studies have highlighted the strong relationship between emotional states and learning effectiveness, indicating that students experiencing high levels of stress or low engagement tend to exhibit reduced academic performance^19,20,21. While several AI-based educational systems have been proposed to enhance learning personalization, most of them primarily focus on cognitive performance metrics and overlook the critical role of emotional dynamics in shaping the learning process²².

Recent developments in intelligent learning systems have introduced graph-based representations and deep learning techniques to model complex relationships among educational components^23,24,25. Graph-based learning frameworks, including graph neural networks (GNNs), have demonstrated strong capabilities in capturing structured dependencies between knowledge elements, learner interactions, and contextual factors²⁶. Additionally, emotion-aware computing has gained attention for its ability to incorporate psychological signals into adaptive learning environments²⁷. Nevertheless, the integration of these approaches within a unified framework for music education remains limited^28,29,30. Existing works often treat emotional variables as auxiliary features rather than core components, and they rarely provide mechanisms for continuous adaptation based on the interplay between cognition and emotion³¹.

Furthermore, the application of AI in mental health-aware education has predominantly relied on empirical or sensor-based data collected from real participants³². However, such approaches require extensive ethical approvals, standardized measurement instruments, and controlled experimental conditions, which may not always be feasible in early-stage research³³. As a result, simulation-based methodologies have emerged as an alternative for exploring complex educational dynamics in a controlled and reproducible manner³⁴. By generating synthetic datasets that mimic interaction logs, emotional indicators, and learning performance metrics, researchers can evaluate the behavior of proposed models under various scenarios while maintaining full control over system parameters and assumptions³⁵.

Despite these advancements, several research gaps remain³⁶. First, there is a lack of unified frameworks that simultaneously model knowledge acquisition, emotional evolution, and learning engagement within a structured representation³⁷. Second, current visual teaching systems do not effectively leverage graph-based intelligence to capture the interdependencies between musical concepts, learning activities, and psychological states³⁸. Third, limited attention has been given to the development of scalable and reproducible simulation pipelines that allow systematic evaluation of AI-driven educational models^39,40,41. Finally, existing approaches often fail to provide a holistic view of learning that integrates both cognitive and emotional dimensions in a dynamically adaptive manner⁴².

To address these challenges, this study proposes a graph-based emotional curriculum intelligence framework (GECIF) for AI-driven visual teaching design in college music education. Table 1 provides essential definitions for graph components, node states, and model parameters. This table ensures the methodology is both clear and reproducible. The proposed framework models the educational environment as an interconnected graph consisting of knowledge nodes, activity nodes, and emotional nodes, enabling a comprehensive representation of learning dynamics. A simulation-based dataset is generated using a controlled computational pipeline that includes interaction logs, emotional indicators, and performance metrics, ensuring reproducibility and scalability. The framework incorporates adaptive learning mechanisms, emotional state modeling, and graph-based optimization to simulate the evolution of learning outcomes and psychological conditions over time. Comparative evaluation against a conventional learning model is conducted using multiple performance metrics, including knowledge gain, emotional stability, stress reduction, engagement, and robustness.

The main aim of this study is to develop and evaluate a simulation-based, AI-driven visual teaching framework that integrates graph-based modeling and emotional intelligence to enhance learning efficiency and support mental health-aware education in college music courses.

Protocol

This study is a computational and simulation-based investigation that does not involve human participants, clinical data, or personal information. All data used in this research were synthetically generated through controlled simulation processes. Therefore, ethical approval from an institutional review board and informed consent were not required. The study adheres to standard guidelines for computational research and ensures full reproducibility and transparency of the methodology.

Theoretical framework and rationale

The proposed graph-based emotional curriculum intelligence framework (GECIF) is grounded in the integration of artificial intelligence, graph-based learning, emotional computing, and visual pedagogy, aiming to model learning as a dynamic and multi-dimensional process that simultaneously captures cognitive development, emotional evolution, and behavioral engagement. Traditional instructional approaches often assume a linear progression of knowledge acquisition and do not adequately reflect the complex interactions between cognitive performance and psychological states. This limitation is particularly critical in college music education, where emotional expression, motivation, and creativity are inseparable from technical and theoretical learning. Within this context, the framework conceptualizes the educational process as an interconnected system in which knowledge elements, learning activities, and emotional states continuously influence one another. Graph-based modeling enables a structured and scalable representation of these relationships, allowing both direct and indirect dependencies to be captured effectively. Emotional dynamics are incorporated as intrinsic components rather than auxiliary variables, enabling the simulation of how psychological conditions evolve and propagate during the learning process. The integration of visual teaching design further enhances interpretability by transforming abstract computational interactions into intuitive graphical representations.

Methodology and simulation study design

This study adopts a computational simulation-based design in which no human participants are involved, and all data are synthetically generated through a controlled and reproducible pipeline. The outcomes reported, therefore, represent model-based evaluations rather than empirical observations derived from real learners. Improvements in learning performance, engagement, emotional stability, and stress reduction are interpreted as simulated or potential effects of the proposed framework. The simulation environment approximates realistic learning conditions by incorporating temporal dependencies, variability in interaction patterns, and evolving emotional states. A comparative experimental configuration is implemented in which the proposed GECIF framework is evaluated against a conventional baseline model that represents a non-adaptive learning system where knowledge evolves solely based on interaction intensity without graph-based representation, emotional modeling, or curriculum optimization. Both models are executed under identical simulation conditions to ensure a consistent and unbiased comparison.

Dataset generation and structure

A large-scale dataset was constructed to simulate learning dynamics in college music education by integrating interaction logs, emotional indicators, and learning performance metrics within a unified temporal framework. The dataset was derived and extended from the publicly available student performance dataset (UCI Machine Learning Repository), which provides academic performance data along with demographic and behavioral attributes. This dataset served as a foundational reference for modeling realistic learning patterns. Interaction logs represent simulated engagement with learning activities such as practice sessions, listening exercises, and visual learning tasks. Emotional indicators capture psychological states, including stress, engagement, and motivation, while learning performance reflects knowledge acquisition and progression over time. To enrich the dataset and incorporate temporal dependencies, a MATLAB-based simulation pipeline was employed with fixed random seed initialization to ensure reproducibility. Temporal continuity was introduced through autoregressive processes, enabling interaction behavior and emotional states to evolve gradually over time. The final dataset includes one thousand simulated learners observed over five hundred time steps, with multiple interaction features representing diverse learning behaviors. All variables are normalized within the range [0,1] to ensure consistency and comparability. The dataset is partitioned into training, validation, and testing subsets using a 70%–15%–15% split to ensure unbiased model evaluation.

Graph construction and representation

The learning environment is modeled as a weighted graph defined as:

G = (V,E,W)  (1)

where denotes the set of nodes, represents the set of edges, and is the weighted adjacency matrix describing the strength of relationships between nodes. The node set is decomposed as:

V = V_k∪V_a∪V_e  (2)

where V_k, V_a, and V_e correspond to knowledge nodes, activity nodes, and emotional nodes, respectively. The edge weights are computed as:

W_ij = α S_c(i,j) + βS_e(i,j)  (3)

where S_c(i,j) represents cognitive similarity between nodes, S_e(i,j) represents emotional correlation, and α and β are weighting coefficients controlling their relative influence. Equation (3) defines how structural and emotional relationships jointly determine the connectivity of the graph.

Node state representation and graph learning

Each node is associated with a dynamic state vector defined as:

x_i(t) = [k_i(t),e_i(t),a_i(t)]  (4)

where k_i(t), e_i(t), and a_i(t) denote knowledge level, emotional state, and engagement level at the time t, respectively. Node representations are updated using graph-based aggregation:

   (5)

where is the embedding of the node i at layer l, σ is a nonlinear activation function, and N(i) denotes the set of neighboring nodes. Equation (5) describes how each node integrates information from its neighbors. An attention mechanism is further introduced:

     (6)

where α_ij is the attention coefficient and ∅(·) is a similarity function. Equation (6) enables adaptive weighting of neighboring node contributions.

Emotional dynamics modeling

The temporal evolution of emotional states is modeled as:

     (7)

where γ is the emotional retention coefficient. Equation (7) captures both persistence and propagation of emotional influence. Emotional stability is quantified as:

    (8)

where E⁺(t) and E^-(t) denote positive and negative emotional components, and  is a small constant preventing division by zero. Stress accumulation is defined as:

     (9)

where λ(τ) represents instantaneous stress intensity. Equation (9) models cumulative stress over time.

Learning optimization and adaptive curriculum design

The learning process is governed by a multi-objective optimization function:

J = λ₁K + λ₂K - λ₃S  (10)

where K denotes knowledge gain, E denotes emotional stability, and S denotes stress level, while λ₁, λ₂, λ₃ are weighting coefficients. Equation (10) balances cognitive and emotional objectives. Adaptive curriculum rewiring is defined as:

W_ij(t + 1) = W_ij(t) + ηΔ_ij(t)  (11)

where η is the learning rate and Δ_ij(t) represents update adjustments driven by learning dynamics. Equation (11) allows the graph structure to evolve over time.

Temporal scale and simulation duration

All simulations are conducted using normalized time steps rather than real-time units. Short intervals correspond to scaled simulation steps representing system evolution, while extended simulations over five hundred steps are used to evaluate long-term behavior, convergence, and robustness.

Evaluation metrics and statistical analysis

The framework is evaluated using normalized metrics, including knowledge gain, emotional stability, stress level, engagement level, and robustness. Statistical analysis is performed using MATLAB statistical functions, with mean values, standard deviations, and 95% confidence intervals computed for each metric. Normality assumptions are verified prior to applying paired t-tests. Comparative evaluation between the proposed framework and the baseline model is conducted across n = 1000 simulated learners, with statistical significance defined as p < 0.05.

Reproducibility and implementation details

All simulations, data processing pipelines, and evaluation procedures are fully reproducible. The dataset generation process, graph construction parameters, model configurations, and optimization settings are explicitly defined. Random seed initialization is fixed to ensure consistent outputs across repeated runs. All hyperparameters, scripts, and dataset descriptions are provided in the supplementary materials, enabling full replication of the study using MATLAB or equivalent computational tools.

All simulations, data processing pipelines, and evaluation scripts are provided in Supplementary File 1. The dataset includes raw interaction logs, emotional indicators, and learning performance metrics. Graph construction parameters and optimization settings are fully documented. The system can be reproduced using standard machine learning frameworks. Random seed initialization is fixed to ensure consistency. All hyperparameters are reported explicitly. The protocol ensures full transparency and reproducibility.

Results

The performance of the proposed Graph-based emotional curriculum intelligence framework (GECIF) was evaluated using a simulation-based dataset consisting of n = 1000 learners over T = 500 normalized time steps. Figure 1 illustrates the conceptual architecture of the GECIF for AI-driven visual music education and mental health promotion. All metrics are reported on normalized scales within the range [0,1]. Statistical evaluation includes mean values, standard deviations (SD), and 95% confidence intervals (CI), with significance assessed using paired t-tests (p < 0.05). All comparisons are computed across the full simulated population. Figure 2 illustrates the training dynamics and knowledge acquisition behavior of the proposed framework compared to the baseline model. The GECIF framework achieves a knowledge gain of 0.88 ± 0.04 (95% CI: 0.84–0.92), while the baseline model achieves 0.65 ± 0.06 (95% CI: 0.59–0.71), with a statistically significant difference (p < 0.001). The learning curve shows a steeper growth rate and stable convergence without oscillatory behavior.

Figure 3 presents the temporal evolution of emotional stability. The proposed framework achieves an Emotional Stability Index of 0.92 ± 0.03, compared to 0.70 ± 0.05 in the baseline model (p < 0.01). Variability is reduced, with lower standard deviation values observed across time. Figure 4 shows stress dynamics over time. The proposed model maintains a lower stress level of 0.15 ± 0.05, compared to 0.45 ± 0.07 for the baseline (p < 0.001). The accumulation rate remains consistently lower across the simulation duration. Figure 5 presents the distribution of learning performance across learners. The proposed framework shows a distribution centered at higher values with reduced spread compared to the baseline, indicating lower variance in outcomes. Figure 6 illustrates the evolution of node embeddings within the graph. The feature representations exhibit continuous transitions across iterations, with no abrupt changes observed, indicating stable updates. Figure 7 shows the interaction strength between knowledge, activity, and emotional nodes. The heatmap reveals statistically significant correlations (p < 0.05) across node categories, with varying magnitudes of interaction.

Figure 8 presents the relationship between normalized knowledge gain and emotional stability throughout the simulated learning process. The proposed GECIF framework demonstrates a strong positive correlation between the two variables (r = 0.82), compared with a moderate positive correlation observed for the conventional baseline model (r = 0.65). Both correlations were statistically significant (p < 0.05), indicating that improvements in knowledge acquisition are accompanied by enhanced emotional stability, thereby reflecting the balanced integration of cognitive development and emotional regulation achieved by the proposed framework. Figure 9 shows the combined contribution of knowledge and emotional components over time. Knowledge values increase progressively, while emotional variability decreases, with consistent trends observed across all simulation steps. Figure 10 compares the distributions of knowledge gain achieved by the proposed GECIF framework and the conventional baseline using box-and-whisker plots. The proposed framework achieved a significantly higher mean knowledge gain (9.87 ± 3.46) than the conventional model (6.90 ± 2.88), with an independent-samples t-test confirming a statistically significant difference (p < 0.0001; Cohen's d = 1.43). Although the GECIF framework exhibited greater variability, as reflected by its larger standard deviation and interquartile range, the entire distribution was shifted toward higher knowledge-gain values, indicating that the increased dispersion accompanies improved overall performance rather than reduced stability. These results demonstrate that the proposed framework maintains robust learning performance while effectively accommodating diverse simulated learner responses. Figure 11 presents engagement dynamics across time. The proposed framework achieves an engagement level of 0.90 ± 0.04, compared to 0.68 ± 0.06 for the baseline model (p < 0.01), with reduced variability. Figure 12 shows engagement behavior in a cyclic form. The amplitude of fluctuations is lower in the proposed framework, indicating reduced variability compared to the baseline.

Figure 13 illustrates spatial feature intensity across the graph. The feature distribution shows continuous variation with identifiable peaks and troughs, indicating heterogeneous activation patterns. Figure 14 presents the directional flow of knowledge transfer. The vector field shows organized propagation patterns, with convergence regions corresponding to higher accumulation of values. Figure 15 shows the evolution of graph topology over time. The structural updates occur gradually, with smooth transitions in feature values across iterations. Figure 16 presents feature interdependency analysis. The correlation heatmap indicates moderate to strong relationships between features, with no dominant redundancy observed. Figure 17 illustrates the contribution of multi-dimensional features. The contribution levels vary across nodes, indicating heterogeneous influence patterns. Figure 18 shows the joint evolution of knowledge and emotional states. Knowledge values increase steadily while emotional variability decreases over time, with consistent trends across the full simulation period. A quantitative summary of these results is provided in Table 2, where all metrics, including knowledge gain, emotional stability, stress level, engagement, and robustness, are reported with statistical measures. The results show consistent differences between the proposed framework and the baseline model across all evaluated metrics. Statistical analysis confirms that these differences are significant across the simulated population. Overall, the simulation results consistently demonstrate that the proposed GECIF framework outperforms the conventional baseline across all evaluated normalized performance metrics. These findings support the study objective and hypothesis that integrating graph-based emotional intelligence with AI-driven visual teaching can enhance simulated learning performance, emotional stability, engagement, and stress management within a reproducible computational environment.

DATA AVAILABILITY:

All data and computational procedures utilized in this study are fully reproducible and openly available. The complete code for dataset generation, along with structured data representations, is provided in Supplementary File 1.

Figure 1: Conceptual architecture of the graph-based emotional curriculum intelligence framework (GECIF) for AI-driven visual music education and mental health promotion. This figure presents the complete system structure, including knowledge–emotion graph construction, graph neural network analysis, emotional pathway optimization, adaptive curriculum rewiring, visual learning interface, and multi-objective optimization. It illustrates how cognitive and emotional components are integrated into a unified adaptive learning system. Please click here to view a larger version of this figure.

Figure 2: Evolution of knowledge acquisition under adaptive graph-based learning compared with the conventional baseline model. The figure illustrates the progression of normalized knowledge gain over normalized simulation steps, demonstrating faster convergence and higher simulated learning efficiency for the proposed GECIF framework. Statistical comparisons between the proposed and baseline models were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized knowledge gain. Statistically significant differences were considered at p < 0.05. Please click here to view a larger version of this figure.

Figure 3: Temporal improvement of emotional stability through intelligent curriculum adaptation. The figure illustrates the evolution of the normalized Emotional Stability Index over normalized simulation steps, demonstrating smoother stabilization and reduced emotional fluctuations for the proposed GECIF framework compared with the conventional baseline model. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized Emotional Stability Index. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 4: Reduction of stress levels through emotion-aware learning pathways. The figure illustrates the temporal evolution of normalized stress levels over normalized simulation steps, showing that the proposed GECIF framework maintains lower and more stable simulated stress than the conventional baseline model, indicating improved emotional regulation throughout the learning process. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized stress level. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 5: Distribution of learning outcomes across simulated learners. The figure illustrates the distribution of normalized knowledge gain for the proposed GECIF framework and the conventional baseline model. The proposed framework demonstrates a distribution shifted toward higher normalized knowledge gain with reduced variability, whereas the baseline model is concentrated at lower performance levels. Statistical comparisons between the two distributions were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. No error bars are presented because the figure displays frequency distributions rather than summary statistics. Please click here to view a larger version of this figure.

Figure 6: Evolution of graph-based knowledge representations across learning iterations. The figure illustrates the temporal evolution of graph node embeddings over normalized simulation steps. Peaks represent stronger feature activations, whereas valleys indicate weaker feature representations, demonstrating the continuous adaptation and refinement of the knowledge graph during the simulation process. This visualization presents the evolution of learned feature representations and does not include error bars. Statistical comparisons reported in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 7: Strength of relationships between knowledge and emotional nodes in the learning graph. The figure visualizes the normalized interaction strengths between cognitive and emotional nodes within the graph-based learning framework. Higher values indicate stronger cognitive–emotional dependencies, whereas lower values represent weaker relationships, demonstrating the complex interconnected structure and adaptive evolution of the learning graph throughout the simulation. This visualization presents graph connectivity patterns and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 8: Relationship between learning progress and emotional well-being in adaptive curriculum environments. The figure illustrates the positive relationship between normalized knowledge gain and emotional stability throughout the simulated learning process. As knowledge acquisition increases, emotional stability also improves, indicating effective integration of cognitive development and emotional regulation within the proposed GECIF framework. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized performance values. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 9: Contribution of knowledge acquisition and emotional dynamics to overall learning performance over normalized simulation steps. The figure illustrates the temporal contributions of normalized knowledge gain and emotional dynamics to the overall simulated learning performance. Knowledge acquisition exhibits a dominant and sustained increase throughout the simulation, while the influence of emotional dynamics gradually stabilizes, demonstrating the effective integration of cognitive and emotional processes within the proposed GECIF framework. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized performance values. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 10: Variability and robustness of learning performance across adaptive and conventional models. The figure presents box-and-whisker plots comparing the distributions of normalized knowledge gain for the proposed GECIF framework and the conventional baseline model. The proposed framework exhibits a higher median knowledge gain and an overall distribution shifted toward greater performance values, while also displaying a wider interquartile range and larger overall dispersion, reflecting greater variability among simulated learners. Despite this increased variability, the consistently higher distribution of knowledge gain demonstrates that the proposed framework maintains robust learning performance across diverse learner conditions. Statistical comparisons were performed using an independent-samples t-test across n = 1000 simulated learners per model. The boxes represent the interquartile range (IQR), the horizontal line within each box indicates the median, and the whiskers extend to 1.5 × IQR. The proposed framework achieved a significantly higher mean knowledge gain than the conventional model (p < 0.0001; Cohen's d = 1.43). No error bars are presented because the box-and-whisker plot summarizes data distributions using quartiles and whiskers rather than confidence intervals. Please click here to view a larger version of this figure.

Figure 11: Dynamic changes in student engagement during the learning process. The figure illustrates the temporal evolution of the normalized Student Engagement Index over normalized simulation steps, demonstrating that the proposed GECIF framework achieves higher and more adaptive engagement than the conventional baseline model throughout the simulated learning process. Statistical comparisons between the proposed and baseline models were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized Student Engagement Index. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 12: Cyclic patterns of student engagement in adaptive learning environments. The figure illustrates the temporal evolution of the normalized Student Engagement Index over normalized simulation steps, showing periodic yet stable engagement fluctuations that indicate the proposed GECIF framework maintains a balanced and adaptive rhythm of learner interaction throughout the simulated learning process. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized Student Engagement Index. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Figure 13: Spatial variation of feature intensity across the learning graph structure. The figure visualizes the normalized spatial distribution of feature activation across the graph-based learning framework. Regions with higher intensity indicate stronger feature activations, whereas lower-intensity regions represent weaker activations, revealing the structured, non-linear organization of knowledge representations and their adaptive relationships within the learning graph during the simulation. This visualization presents spatial feature patterns and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 14: Directional flow of knowledge transfer within the curriculum graph. The figure illustrates the normalized propagation of knowledge across interconnected nodes within the graph-based curriculum framework. The directional pathways represent the organized flow of information through the learning graph, highlighting efficient knowledge dissemination and adaptive information exchange among curriculum components during the simulation. This visualization depicts knowledge propagation patterns and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 15: Structural evolution of the learning graph over time through adaptive feature transformation. The figure illustrates the progressive evolution of the graph topology during normalized simulation steps as adaptive feature transformation continuously refines the relationships among knowledge nodes. The dynamic structural changes reflect the framework's ability to reorganize and optimize learning pathways, resulting in increasingly coherent and adaptive graph representations throughout the simulated learning process. This visualization depicts graph topology evolution and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 16: Interdependency patterns among extracted learning features in the adaptive framework. The figure visualizes the normalized interdependency patterns among the extracted learning features within the proposed GECIF framework. Stronger relationships indicate higher correlations and more effective interactions between feature dimensions, whereas weaker relationships represent lower dependencies, demonstrating the complex multidimensional organization of cognitive and emotional representations learned by the adaptive graph-based framework. This visualization presents feature interaction patterns and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 17: Contribution of multi-dimensional features to knowledge acquisition performance. The figure illustrates the normalized contributions of multiple feature dimensions to knowledge acquisition across the graph-based learning framework. Variations in feature importance among graph nodes highlight the heterogeneous nature of cognitive and emotional learning factors, demonstrating how different features interact and contribute to adaptive knowledge acquisition within the proposed GECIF framework. This visualization presents multidimensional feature contribution patterns and does not include error bars. Statistical comparisons reported elsewhere in the study were performed using a paired t-test across n = 1000 simulated learners, with statistical significance defined as p < 0.05. Please click here to view a larger version of this figure.

Figure 18: Simultaneous evolution of cognitive and emotional states during adaptive learning. The figure illustrates the temporal evolution of normalized knowledge gain and emotional stability over normalized simulation steps, demonstrating steady cognitive improvement accompanied by progressively reduced emotional fluctuations. The coordinated progression of these two learning dimensions indicates that the proposed GECIF framework effectively maintains a balanced integration of cognitive development and emotional regulation throughout the simulated learning process. Statistical comparisons were performed using a paired t-test across n = 1000 simulated learners. Where error bars are shown, they represent 95% confidence intervals (95% CI) of the mean normalized performance values. Differences were considered statistically significant at p < 0.05. Please click here to view a larger version of this figure.

Symbol	Description	Category
G	Graph representing the entire learning system	Graph Component
V	Set of all nodes	Graph Component
E	Set of edges (relationships)	Graph Component
W	Weight matrix representing relationship strength	Graph Component
V_k	Knowledge nodes (musical concepts)	Node Type
V_a	Activity nodes (learning actions)	Node Type
V_e	Emotion nodes (psychological states)	Node Type
W_ij	Weight between node i and j	Edge Parameter
S_c(i,j)	Cognitive similarity between nodes	Edge Parameter
S_e(i,j)	Emotional correlation between nodes	Edge Parameter
alpha	Weight for cognitive similarity	Model Parameter
beta	Weight for emotional correlation	Model Parameter
x_i(t)	State vector of node i at time t	Node State
k_i(t)	Knowledge level of node i	Node State
e_i(t)	Emotional state of node i	Node State
a_i(t)	Activity engagement of node i	Node State
h_i	Node embedding representation	GNN Parameter
gamma	Emotional decay factor	Model Parameter
eta	Learning/adaptation rate	Model Parameter
J	Multi-objective optimization function	Optimization

Table 1: Summary of variables, parameters, and notations used in the Graph-based emotional curriculum intelligence framework (GECIF). This table provides definitions for graph components, node states, and model parameters, supporting reproducibility and clarity of the proposed methodology.

Metric	Proposed GECIF (Mean)	Proposed SD	Proposed 95% CI	Conventional (Mean)	Conventional SD	Conventional 95% CI	Improvement (%)	p-value
Knowledge Gain	0.88	0.04	0.84–0.92	0.65	0.06	0.59–0.71	35	0.001
Emotional Stability	0.92	0.03	0.89–0.95	0.7	0.05	0.65–0.75	31	0.002
Stress Reduction	0.85	0.05	0.80–0.90	0.55	0.07	0.48–0.62	54	0.0005
Engagement Level	0.9	0.04	0.86–0.94	0.68	0.06	0.62–0.74	32	0.0015
Robustness	0.87	0.03	0.84–0.90	0.66	0.05	0.61–0.71	32	0.002
Learning Efficiency	0.91	0.04	0.87–0.95	0.69	0.06	0.63–0.75	32	0.001
Convergence Speed	0.89	0.03	0.86–0.92	0.67	0.05	0.62–0.72	33	0.001
Prediction Accuracy	0.93	0.02	0.91–0.95	0.72	0.05	0.67–0.77	29	0.0008
Emotional Variance Reduction	0.86	0.04	0.82–0.90	0.6	0.06	0.54–0.66	43	0.0012
Adaptability Score	0.9	0.03	0.87–0.93	0.68	0.05	0.63–0.73	32	0.001

Table 2: Comparative performance metrics between the proposed GECIF framework and the conventional learning model. The table summarizes improvements in knowledge gain, emotional stability, stress reduction, engagement, robustness, and overall learning efficiency, demonstrating the superiority of the proposed approach.

Supplementary File 1: Code for dataset generation and structured data representations. Summary of the MATLAB-based dataset generation, processing pipeline, and evaluation procedures used in the Graph-based emotional curriculum intelligence framework (GECIF). This file provides an overview of the simulation stages, including interaction log generation, emotional indicator modeling, learning performance computation, graph construction, feature engineering, statistical analysis, validation checks, and data export procedures, supporting reproducibility, transparency, and clarity of the proposed methodology.Please click here to download this file.

Discussion

The present study proposed a graph-based emotional curriculum intelligence framework (GECIF) that integrates graph-based representation learning, emotional state modeling, adaptive curriculum optimization, and AI-driven visual teaching into a unified computational framework for college music education. Unlike conventional AI-supported educational models that primarily optimize academic performance, the proposed framework simultaneously models interactions among knowledge acquisition, emotional regulation, learner engagement, and adaptive curriculum evolution. The simulation results consistently demonstrated superior performance of the proposed framework over the conventional baseline across multiple evaluation metrics, including knowledge gain, emotional stability, learner engagement, and stress reduction. These findings suggest that integrating relational graph structures with emotion-aware learning provides a more comprehensive computational representation of educational processes than approaches relying exclusively on cognitive adaptation.

The observed improvements in simulated learning performance are consistent with recent developments in AI-driven adaptive education. AI technologies have increasingly been recognized as effective tools for developing personalized and adaptive learning environments capable of responding to learner characteristics and educational needs^18,19,20. In particular, adaptive curriculum design supported by artificial intelligence has been shown to improve instructional flexibility, learner-centered decision-making, and educational effectiveness^19,20. Furthermore, AI-assisted assessment and intelligent feedback systems have demonstrated significant potential for enhancing student achievement through continuous evaluation and adaptive instructional support^21,22,23. Compared with these previous studies, the proposed GECIF framework extends existing methodologies by integrating cognitive knowledge, emotional dynamics, and adaptive learning behaviors within a single graph-based computational architecture. Consequently, the present framework provides a unified representation of educational processes rather than treating cognitive performance and learner adaptation as separate components.

The integration of emotional modeling within the proposed framework is also supported by previous investigations in music education and affective learning. Li¹ demonstrated that emotional visualization combined with deep learning improves music aesthetic teaching by strengthening the interaction between learners and educational content. Similarly, Zhao³ reported that integrating cognitive mapping with audiovisual music therapy effectively reduces anxiety among university students, emphasizing the close relationship between emotional regulation and learning outcomes. Iyendo et al.⁴ further demonstrated that audiovisual art and music interventions significantly improve psychological well-being by reducing symptoms associated with emotional stress. The improvements in emotional stability observed in the present simulation, therefore, agree with previous evidence indicating that emotional awareness constitutes an essential component of successful music education. Unlike previous studies that primarily investigated emotional interventions as independent educational strategies, the proposed framework embeds emotional dynamics directly into graph-based curriculum optimization, allowing emotional states to continuously influence adaptive instructional decisions throughout the learning process.

The simulated reduction in learner stress is similarly consistent with recent literature concerning student mental health in higher education. Academic stress has been identified as a major predictor of reduced psychological well-being and diminished educational performance among university students^28,29,30. Systematic reviews have further shown that structured stress-management interventions significantly improve both academic functioning and psychological health²⁷. More recently, AI-supported digital interventions have demonstrated effectiveness in promoting stress management and adaptive self-regulation among university students²⁹. The present framework differs from these intervention-based approaches because stress reduction emerges through adaptive optimization of cognitive and emotional interactions rather than through separate therapeutic mechanisms. Consequently, the lower stress levels observed in the proposed framework suggest that emotionally adaptive curriculum design itself may contribute to reducing cognitive overload while simultaneously supporting improved learning performance.

The enhanced learner engagement observed in the present study is likewise supported by previous research investigating AI-supported visual education and personalized instructional technologies. Bian et al.¹² demonstrated that AI-generated visual content significantly enhances classroom engagement, learner self-efficacy, and cognitive performance in educational environments. Similarly, Johnson et al.³⁴ reported that digital visual learning environments strengthen student engagement and promote greater participation under stressful educational conditions. Within music education specifically, AI-supported instructional technologies have been shown to facilitate personalized learning pathways and improve educational innovation⁴¹. The proposed GECIF framework complements these findings by combining adaptive visual teaching with graph-based emotional modeling, thereby enabling learner engagement, emotional regulation, and cognitive development to evolve simultaneously throughout the learning process.

The relationship between emotional regulation and learning performance observed in the present study is also supported by broader mental health research. Emotional regulation has been identified as a fundamental determinant of psychological well-being and healthy educational functioning³². Likewise, models of student well-being emphasize that emotional resilience and psychological stability are closely associated with successful academic outcomes^15,16,17. Studies of university students have consistently reported that promoting mental health enhances resilience, engagement, and long-term educational achievement^14,15,16,17. The present simulation demonstrated similar computational behavior, where improvements in emotional stability were accompanied by higher knowledge acquisition and stronger learner engagement. Although the current investigation is entirely simulation-based and therefore cannot establish causal relationships in real learners, the computational results reinforce the growing consensus that emotional regulation should be incorporated directly into intelligent educational systems rather than treated as an independent educational outcome.

Compared with existing computational approaches, the proposed framework offers several conceptual advantages. Contemporary AI-based educational systems increasingly employ adaptive assessment, personalized feedback, and intelligent decision support to improve learning effectiveness^21,22,23. Similarly, AI-driven adaptive curriculum models emphasize learner personalization and dynamic instructional planning^18,19,20. However, these approaches generally evaluate learner performance primarily through cognitive indicators while representing emotional information separately or omitting it altogether. In contrast, the proposed GECIF framework simultaneously models knowledge acquisition, emotional dynamics, learner engagement, and adaptive curriculum evolution within a unified graph representation. This integrated modeling strategy enables continuous interaction among cognitive, behavioral, and emotional variables, thereby providing a more comprehensive computational description of educational processes than conventional AI-driven educational systems.

The framework is particularly relevant to music education because musical learning inherently combines technical skill acquisition, emotional expression, creativity, and learner engagement. Previous investigations have shown that integrating music, visual arts, and creative educational activities contributes to psychological development, emotional well-being, and educational motivation^36,37. Moreover, music students frequently experience substantial academic workload and psychological stress, highlighting the importance of adaptive educational support mechanisms^25,26. The proposed framework extends these educational perspectives by providing a computational mechanism capable of simultaneously representing technical learning progression, emotional adaptation, and adaptive instructional optimization within a single graph-based model. Consequently, the framework provides an interpretable computational foundation for future intelligent music education systems that seek to optimize both educational performance and learner well-being.

Several limitations should nevertheless be acknowledged. First, the framework was evaluated exclusively using synthetic data generated within a controlled computational simulation; therefore, the reported improvements reflect algorithmic performance rather than empirical evidence obtained from real student populations. Second, emotional variables were represented using mathematical abstractions that simplify the complexity of authentic human emotional processes. Third, predefined graph structures and optimization parameters may not fully represent the diversity of educational environments encountered in practice. Finally, graph-based optimization introduces additional computational complexity that may require further optimization before large-scale deployment in intelligent educational systems.

Future investigations should therefore validate the proposed framework using authentic educational datasets and real classroom environments. Incorporating multimodal educational information, including behavioral interaction logs, physiological measurements, audiovisual observations, and digital learning traces, may substantially improve the realism of emotional state estimation and adaptive curriculum optimization. Comparative evaluations against recent AI-driven adaptive learning systems ^{18,19,20,21,22} would further clarify the relative advantages of graph-based emotional modeling. Extending the framework beyond music education to additional educational disciplines would also strengthen evidence for its generalizability and practical applicability. The present simulation-based investigation demonstrates that integrating graph-based relational learning with emotional modeling represents a promising direction for next-generation intelligent educational systems. The findings are consistent with recent advances in AI-supported adaptive education, emotion-aware music teaching, personalized learning, and student mental health research^{25,26,27,28,29,30}. Although validation using real learners remains necessary, the proposed GECIF framework provides a structured computational foundation for future AI-driven educational systems capable of simultaneously optimizing cognitive performance, emotional well-being, learner engagement, and adaptive curriculum design.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
32 GB DDR4 RAM Module	Kingston Technology	NA
CUDA Toolkit for GPU Acceleration (v11.8)	NVIDIA	version 1.18
Curriculum Adaptation Engine	Custom Implementation	CAE003
Dataset: Emotional Indicators	Generated Dataset	DEI006
Dataset: Learning Performance Metrics	Generated Dataset	DLPM007
Dataset: Student Interaction Logs	Generated Dataset	DIL005
Emotional State Simulation Module	Custom Implementation	ESM002
GPU Workstation (NVIDIA RTX 3080)	NVIDIA	NA
Graph Neural Network Implementation Module	Custom Implementation	NA
Intel Core i7 Processor (3.6 GHz)	Intel Corporation	NA
Jupyter Notebook Environment (v6.5)	Project Jupyter	version 6.5
MATLAB Software (R2023a)	MathWorks	R2023a
Matplotlib Visualization Library (v3.7)	Matplotlib	version 3.7
NetworkX Graph Processing Library (v3.1)	NetworkX Developers	version 3.1
NumPy Numerical Computing Library (v1.24)	NumPy Developers	version 1.24
Optimization Module (Multi-objective Solver)	Custom Implementation	MOS008
Pandas Data Processing Library (v2.0)	Pandas Development Team	version 2.0
Python Programming Environment (v3.10)	Python Software Foundation	version 3.10
PyTorch Deep Learning Framework (v2.0)	Meta AI	version 2.0
Random Seed Initialization Script	Custom Implementation	RSI009
Scikit-learn Machine Learning Library (v1.3)	Scikit-learn	version 1.3
Seaborn Statistical Visualization Library (v0.12)	Seaborn	version 0.12
TensorFlow Library (v2.12)	Google	version 2.12
Visual Learning Interface Toolkit	Custom Implementation	VLI004
Windows 11 Operating System	Microsoft	W11