Innovative AI-Driven Approaches to Mitigate Math Anxiety and Enhance Resilience Among Students with Persistently Low Performance in Mathematics

Abstract

This study explored innovative methods for teaching mathematics to seventh-grade students with persistently low performance by using an AI-driven neural network approach, specifically focusing on solving first-degree inequalities. Guided by the Response to Intervention (RTI) framework, the intervention aimed to reduce math anxiety and build academic resilience through the development of cognitive and metacognitive strategies. A rigorous pre- and post-test design was employed to evaluate changes in performance, anxiety levels, and resilience. Fifty-six students participated in the 12-week program, receiving personalized instruction tailored to their individual needs. The AI tool provided real-time feedback and adaptive problem-solving tasks, ensuring students worked at an appropriate level of challenge. Results indicated a marked decrease in math anxiety alongside significant gains in cognitive skills such as problem-solving and numerical reasoning. Students also demonstrated enhanced metacognitive abilities, including self-monitoring and goal setting. These improvements translated into higher academic performance, particularly in the area of inequalities, and greater resilience, highlighting the effectiveness of AI-based strategies in supporting learners who struggle persistently in mathematics. Overall, the findings underscore how AI-driven teaching approaches can address both the cognitive and emotional dimensions of mathematics learning. By offering targeted, adaptive support, educators can foster a learning environment that reduces stress, promotes engagement, and facilitates long-term academic success for students with persistently low performance in mathematics.

1. Introduction

Mathematics anxiety is a well-documented barrier to academic success. Students with difficulties and low scores in mathematics often face considerable challenges when engaging with mathematical concepts, which are further exacerbated by intense feelings of fear and tension when faced with math tasks. This condition, often referred to as “math anxiety”, can significantly hinder their academic performance and motivation (Drigas & Pappas, 2015; Pappas et al., 2019). The traditional methods of teaching math often fail to accommodate these students’ unique learning needs, leading to a self-reinforcing cycle of frustration and avoidance, which further perpetuates their struggles (Lee & Paul, 2023). As such, addressing math anxiety is a critical area of focus for educators and researchers. Polydoros (2024b) highlights the pressing need to address math anxiety (MA) in educational contexts, particularly its detrimental effects on students’ learning, and advocates for tailored interventions that go beyond traditional teaching methods.

Innovative educational tools, such as artificial intelligence (AI), offer promising solutions to these persistent challenges. AI has the potential to create personalized learning environments that are adaptive to the needs of individual students. Some studies have explored the interplay of MA with cognitive and executive functions, such as working memory and calculation fluency, highlighting the potential of tailored interventions to improve outcomes for these groups (Gökçe & Güner, 2024). Furthermore, integrating AI in educational strategies shows promise for fostering academic resilience and reducing inequities in math education (Yadav & Bhardwaj, 2024; S. Wang et al., 2024).

AI-driven systems can tailor instruction to students’ specific learning paces and provide real-time feedback, thereby reducing the pressure and anxiety often associated with math tasks. These personalized interventions, which often include features such as gamification and immediate error correction, have shown promise in mitigating math anxiety and fostering a growth mindset among students (Inchamnan & Chomsuan, 2020; Santos-Guevara et al., 2024). By creating supportive, low-pressure environments, AI tools can not only help students manage their anxiety but also enhance their resilience in the face of academic challenges.

Considering that a significant portion of teaching time is still devoted to tasks that could easily be automated, this highlights the enormous potential for AI in Education (AIEd) to grow and make an even greater impact (Lampou, 2023; Okonkwo & Ade-Ibijola, 2020; Zaugg, 2024). AI has already proven its ability to improve many aspects of education, such as personalized learning, adaptive assessments, intelligent tutoring systems, automated grading, virtual and augmented reality in classrooms, performance prediction through data analysis, language learning, and promoting accessibility and inclusion (Shetye, 2024).

However, despite these advancements, there is still a notable gap in research when it comes to how AI can help reduce math anxiety in students. While plenty of studies have looked at AI’s impact on academic performance and engagement, very few have focused on its potential to address the specific challenges of math anxiety, particularly in learners with special educational needs (Hwang & Chien, 2022).

Although there is a wealth of research on AI tools and their benefits for academic success, more work is needed to explore how these tools can truly foster engagement, motivation, and resilience in diverse learning environments. These environments often come with unique challenges and barriers that can make it harder for students to thrive (Owan et al., 2023).

While AI has already been shown to boost academic performance and engagement, its role in helping students manage their emotions and build resilience, particularly those struggling with math-related anxiety, remains an area that requires much deeper investigation (X. Wang et al., 2024).

The purpose of this study is to investigate how AI-based technologies can reduce math-related anxiety and bolster academic resilience among seventh-grade students who demonstrate persistently low performance in mathematics. By exploring how AI can personalize learning and provide targeted emotional support, this research aims to fill the gap in the literature regarding the intersection of AI, math anxiety, and learning difficulties/low-performing students. Through this investigation, we aim to elucidate how AI technologies can be leveraged to foster more inclusive, supportive, and effective learning environments for students confronting persistent underachievement in mathematics, particularly in addressing first-degree inequalities.

Research Questions

• How does the AI-based intervention affect mathematics performance among seventh-grade students experiencing persistent academic underachievement?
• What is the impact of the AI-based intervention on math anxiety levels in seventh-grade students?
• How does the utilization of cognitive strategies evolve following the AI-based intervention in seventh-grade students?
• How does the AI-based intervention influence academic resilience in seventh-grade students experiencing persistent academic underachievement?
• What are the interrelationships among cognitive strategies, math anxiety, and academic resilience before and after the intervention in seventh-grade students?

2. Literature Review

2.1. Impact of AI-Based Interventions on Math Performance

The introduction of artificial intelligence (AI) in mathematics education has been transformative in addressing challenges faced by students with mathematical difficulties. AI-driven platforms enhance learning by offering personalized, adaptive instruction tailored to each student’s abilities, thus improving math performance (Holstein et al., 2020; Zheng & Tse, 2023). For instance, intelligent tutoring systems provide real-time feedback and targeted support, enabling students to overcome conceptual difficulties and build confidence (Hwang & Chien, 2022).

AI systems analyze extensive data to create individualized learning paths, ensuring content is engaging and appropriately challenging (Bhutoria, 2022). These tools help bridge learning gaps and promote understanding by facilitating deeper comprehension and enhanced problem-solving skills among students. Consequently, AI-based interventions contribute significantly to improved academic outcomes (Hillmayr et al., 2020).

2.2. Effect of AI-Based Interventions on Math Anxiety

Math anxiety, a debilitating fear of engaging with mathematical tasks, has been shown to negatively affect students’ academic achievement and emotional well-being (Pizzie & Kraemer, 2019). AI-based interventions offer unique opportunities to address this issue by creating supportive, low-pressure learning environments that reduce stress and foster confidence (Johnston-Wilder & Lee, 2024).

Gamified elements in AI systems, such as rewards and progress tracking, encourage positive engagement and diminish anxiety, reframing mistakes as opportunities for growth (Drigas et al., 2022a; Johnston-Wilder & Lee, 2024) Furthermore, personalized feedback from AI tutors helps students manage their emotions, fostering a growth mindset and reducing the fear of failure (Soler-Dominguez et al., 2024). The adaptive nature of AI ensures that students feel supported at their individual learning levels, minimizing feelings of frustration and apprehension (Hamari et al., 2016).

2.3. Changes in Cognitive Strategies Following AI Interventions

The integration of cognitive strategies into AI-driven learning platforms has been instrumental in enhancing students’ academic resilience and critical thinking abilities. AI systems encourage the use of metacognitive skills, such as self-monitoring and reflection, to improve problem-solving and decision-making processes (Holstein et al., 2019).

Studies have highlighted how AI tools support the development of working memory and processing speed, which are critical for mathematical success (Petronzi et al., 2021). By tailoring content delivery to individual cognitive profiles, AI helps students adopt effective learning strategies, such as chunking and error analysis, to improve comprehension and retention.

2.4. Influence of AI-Based Interventions on Academic Resilience

Academic resilience refers to a student’s ability to recover from setbacks and persist in the face of challenges, an essential trait for overcoming math anxiety and achieving success (Johnston-Wilder & Lee, 2024). AI applications contribute to building resilience by fostering a supportive and engaging learning atmosphere.

Interactive AI platforms promote perseverance by encouraging students to view challenges as opportunities for growth (Ng et al., 2024). Real-time feedback and adaptive pacing ensure that students remain motivated and confident, even when faced with complex tasks (Jaiswal & Arun, 2021).

2.5. Relationships Between Cognitive Strategies, Math Anxiety, and Academic Resilience

Research shows that students with effective metacognitive strategies are better equipped to manage anxiety and maintain resilience, leading to improved academic performance (Anthonysamy, 2023; Drigas et al., 2022b; Swanson et al., 2024).

AI systems provide a platform for integrating these elements by teaching students to regulate emotions, adapt learning strategies, and persevere in challenging situations (Akavova et al., 2023; Akintayo et al., 2024). For instance, students using AI-based tools report reduced anxiety and increased resilience due to personalized feedback and supportive learning environments (Abed & Salha, 2014). The holistic integration of cognitive, emotional, and motivational support through AI fosters a comprehensive approach to overcoming learning difficulties and achieving success.

3. Materials and Methods

The study was grounded in theories of cognitive and metacognitive development, focusing on enhancing academic resilience among seventh-grade students with learning disabilities in mathematics. It incorporated the Response to Intervention (RTI) model (Safari et al., 2020), emphasizing the role of AI-driven learning tools in promoting effective strategies and reducing math anxiety. Drawing on Self-Regulated Learning (SRL) theory (Zimmerman, 2002), the study also explored how AI fostered independent problem-solving skills and supported cognitive growth through personalized feedback.

3.1. Sample and Sampling Method

The study involved 56 seventh-grade students (average age: 13 years), all identified as low-performing in mathematics based on consistently scoring below 50% in mathematics assessments throughout the academic year prior to the intervention. The sample consisted of 48% boys and 52% girls, with an average mathematics score of 48 out of 100. Participants were selected using purposive sampling from six schools in a mid-sized city to ensure diverse socio-economic backgrounds and varied learning profiles.

3.2. Data Collection Tools and Data Analysis

The data for this study were collected using validated assessment scales, with the evaluation and analysis being facilitated by the AI-based neural network system.

3.3. Data Collection and Pre-Processing

• The Math Anxiety Scale for Children (MASC), a 5-point Likert scale, was used to assess students’ anxiety related to mathematics (Chiu & Henry, 1990).
• The Cognitive and Metacognitive Strategy Use Questionnaire, scored out of 10, measured the frequency of students’ use of cognitive strategies, such as planning, monitoring, and problem-solving (Askell-Williams et al., 2012).
• The Academic Resilience Scale, with scores ranging from 0 to 1, assessed students’ ability to adapt and persist in the face of academic challenges (Cassidy, 2016).

3.4. Neural Network Analysis

○

The neural network system processed the raw survey responses, performing necessary data cleaning (e.g., handling missing data and outliers) and normalization to ensure uniformity across variables. The neural network approach aligns with methodologies employed in previous studies utilizing AI-driven adaptive learning systems (X. Wang et al., 2024; Holstein et al., 2020).

○

The system generated key statistical outputs, such as means, standard deviations, and changes in variables (pre-test vs. post-test), which were essential for subsequent analysis.

○

The neural network was utilized to predict outcomes based on the relationship between metacognitive and cognitive strategies, math anxiety, and academic resilience, providing preliminary insights into the effects of the AI-based intervention.

3.5. Statistical Analysis

○

After the neural network analysis, traditional statistical methods such as Mixed ANOVA, correlation, and regression analyses were conducted to further assess the effects of the intervention and examine the relationships between the variables.

○

Mixed ANOVA was employed to assess the intervention’s effects on math anxiety, metacognitive and cognitive strategies, and academic resilience.

○

Correlation analysis explored the relationships between metacognitive and cognitive strategies, math anxiety, and academic resilience both before and after the intervention.

○

Multiple regression analysis was conducted to examine the predictive relationships between cognitive strategies, metacognitive strategies, and math anxiety on academic resilience.

These combined methods allow for a comprehensive understanding of how the AI intervention impacts, examining both the direct effects (via ANOVA) and the predictive and relational effects (via regression and correlation analysis).

3.6. The Process

• Step 1: Data Collection and Neural Network Processing

• Pre- and Post-Test Data

  ○

  Data were gathered from 56 students on five key variables: math anxiety, cognitive strategy use, metacognitive strategy use, and academic resilience.

  ○

  Measurements were taken at two points in time: before (pre-test) and after (post-test) the intervention.

• Neural Network Processing

  ○

  A neural network was trained on the pre-test data to detect patterns and predict how the AI-based intervention would affect the students.

  ○

  The network’s input consisted of individual pre-test scores for each of the five measures; the output was the predicted post-test scores for these same measures.

  ○

  Using supervised learning, the model was trained on the pre-test data and then tested on the post-test data to infer how changes in one variable (e.g., cognitive strategies) might impact others (e.g., math performance, resilience).

• Step 2: Pre- and Post-Test Data Preparation for Statistical Analysis

• The neural network’s predicted post-test scores were combined with the raw pre-test scores to create the dataset for further statistical evaluation.

• Step 3: Statistical Analysis with Traditional Methods

• Descriptive Statistics

  ○

  Means and standard deviations were calculated for each variable (e.g., math anxiety, cognitive strategies) at both pre-test and post-test, providing an initial overview of how scores changed.

• Mixed ANOVA

  ○

  A Mixed ANOVA was conducted to examine within-subject changes in scores from the pre-test to the post-test.

  ○

  The time factor (pre-test vs. post-test) served as the primary independent variable, and the dependent variables were the measures of math anxiety, cognitive strategies, etc.

• Correlation Analysis

  ○

  Correlations were assessed to identify relationships among the variables before and after the intervention (e.g., how an increase in cognitive strategy use might relate to reduced math anxiety or higher resilience).

• Multiple Regression Analysis

  ○

  A multiple regression analysis was performed to determine the extent to which predictors such as cognitive strategies, metacognitive strategies, and math anxiety contributed to academic resilience.

  ○

  This allowed for an in-depth exploration of how much each factor influenced changes in academic resilience before and after the intervention.

In summary, the process involved collecting relevant data, using a neural network to predict post-test outcomes, and then applying traditional statistical methods—such as mixed ANOVA, correlation, and regression—to determine the effectiveness of the intervention and explore the relationships among the key variables.

3.7. Method of Instruction

Before the intervention began, students were assessed to establish baseline levels of key variables: math anxiety, cognitive strategy use, metacognitive strategy use, and academic resilience. These assessments helped measure the initial status of each student before the AI-based intervention.

• 12-Week AI-Based Intervention

Over the course of 12 weeks, the intervention included:

• Personalized Quizzes and Real-Time Feedback: Students were provided with quizzes that adapted to their performance, ensuring that each task was suitable to their level of understanding.
• Gamified Problem-Solving Tasks: These tasks engaged students in a fun and interactive way to solve math problems, motivating them to improve their skills while managing math anxiety.
• Embedded Cognitive Strategy Prompts: Hints and scaffolding were included in the tasks to help students apply cognitive strategies like problem-solving and diagram drawing.

3.8. AI Technology in the Study: Neural Network

The AI tool used in this study was based on a neural network that processed and interpreted complex data related to each student’s performance. This technology provided individualized learning experiences, aimed at improving math performance, reducing math anxiety, and enhancing resilience.

• Inputs:

The neural network began with collecting and processing the following data:

Demographic Information:

○

Average age: 13 years old.

○

Gender distribution: 48% boys, 52% girls.

○

Average math performance: 48 out of 100.

Math Anxiety Data:

○

Average anxiety score (Math Anxiety Scale): 8.2 out of 10, indicating generally high levels of anxiety.

○

Behavioral observations: 70% of students exhibited anxiety-related behaviors (e.g., avoidance of math tasks). Behavioral observations were systematically conducted by trained classroom teachers during the entire duration of the 12-week intervention program.

Cognitive and Metacognitive Strategies:

○

Cognitive strategies (e.g., “talking through problems”, “drawing diagrams”): average score of 2.3 on a scale of 1 to 5.

○

Metacognitive strategies (e.g., “self-monitoring”, “evaluating comprehension”): average score of 2.2 on a scale of 1 to 5.

• Hidden Layers

Hidden Layer 1:

○

Contained 64 neurons with a ReLU activation function. This layer processed input data to identify patterns in students’ demographics, anxiety levels, and cognitive strategy use. For instance, it helped detect if students with lower cognitive strategy scores had higher anxiety levels.

Hidden Layer 2:

○

Contained 32 neurons. It refined patterns detected in the first layer, specifically focusing on the relationship between anxiety and academic resilience. This layer helped the model understand how cognitive and metacognitive strategies contribute to reducing anxiety and improving resilience.

• Output Layer:

The final output of the neural network consisted of two key predictions:

• Anxiety Level Prediction: A continuous output indicating each student’s predicted math anxiety level after the intervention.
• Academic Resilience Prediction: A score from 1 to 5, representing the predicted level of academic resilience for each student.

3.9. Explanation of Model Training

This section provides a concrete numerical example of the model’s operation. It illustrates how the two hidden layers (64 ReLU neurons in the first layer and 32 ReLU neurons in the second layer) process normalized student data—such as age, gender, math performance, anxiety, and strategy scores—to generate a shared feature representation. From this representation, two output neurons compute predictions for math anxiety and academic resilience using Sigmoid activation functions. The section further demonstrates, step by step, how the loss is calculated and how weight updates are performed via backpropagation and gradient descent. This detailed example offers scientific evidence of the training process and serves as a foundation for interpreting the overall results.

• Data and Input Encoding

As outlined, each student’s data included:

• Demographic Information: e.g., age, gender, and average math performance.
• Math Anxiety Indicators: Math Anxiety Scale (MAS) score (0–10), behavioral observations of avoidance.
• Cognitive & Metacognitive Strategy Scores: e.g., 2.0/5 for cognitive, 2.1/5 for metacognitive strategies.

• Example Input Vector

For demonstration, we had the following normalized/encoded features for one student:

x = [13, 0, 42, 8.5, 2.0, 2.1], where 13 represents age, 0 indicates male, 42 is the math score, 8.5 is the anxiety score, 2.0 is the cognitive strategy score, and 2.1 is the metacognitive strategy score.

• Network Architecture

• Hidden Layer 1: 64 neurons, ReLU activation.
• Hidden Layer 2: 32 neurons, ReLU activation.
• Output Layer: 2 neurons—one for math anxiety (o_anxiety), one for academic resilience (o_resilience), each using Sigmoid.

• Forward Pass (Illustrative Example)

We illustrate a single forward pass for the student’s input x.

• Hidden Layer 1

Each of the 64 neurons computes:

$$ne{t}_j^1={\displaystyle \sum }_i\left(w_{ji}^1\times x_i\right)+b_j^1,h_j^1=ReLU\left(ne{t}_j^1\right)$$

Representative Neuron: Suppose for neuron j = 1:

$$ne{t}_1^1=1.20,ReLU\left(1.20\right)=1.20$$

All 64 neurons produce $h^1=\left[h_1^1,h_2^1,\dots ,h_{64}^1\right]$

• Hidden Layer 2

Each of the 32 neurons processes h¹:

$$ne{t}_{\kappa }^2={\displaystyle \sum }_j\left(w_{kj}^2\times h_j^1\right)+b_k^2,h_k^2=ReLU\left(ne{t}_k^2\right)$$

• Representative Neuron: $n\mathrm{e}{t}_1^2=0.75,because0.75>0$ thus $ReLU\left(0.75\right)=0.75$
• Collectively, we get $h^2=\left[h_1^2,h_2^2,\dots ,h_{32}^2\right]$

• Output Layer

Two neurons, each applying Sigmoid:

• Math Anxiety Neuron:

  $$ne{t}_{anx}={\displaystyle \sum }_j\left(w_{anx,k}\times h_k^2\right),o_{anxiety}=Sigmoid\left(ne{t}_{anx}\right)$$

For $ne{t}_{anx}=0.54$, $o_{anxiety}={\displaystyle \frac{1}{1+e^{-0.54}}}\approx 0.63$

Academic Resilience Neuron:

$$ne{t}_{res}={\displaystyle \sum }_j\left(w_{res,k}\times h_k^2\right),o_{res}=Sigmoid\left(ne{t}_{res}\right)$$

For $ne{t}_{res}=-0.32,o_{resilience}={\displaystyle \frac{1}{1+e^{-0.32}}}\approx 0.42$

Hence, for this single forward pass, the network outputs:

$$o_{anxiety}\approx 0.63,o_{resilience}\approx 0.42$$

• Loss Function

We combine the Mean Squared Errors (MSE) of both outputs. For $t_{anxiety}=0.80$, and $t_{resilience}=0.25$, then:

$${\Delta }_{\mathrm{anx}}={\mathrm{o}}_{\mathrm{anxiety}}-{\mathrm{t}}_{\mathrm{anxiety}}=0.63-0.80=-0.17.$$

$${\Delta }_{\mathrm{res}}={\mathrm{o}}_{\mathrm{resilience}}-{\mathrm{t}}_{\mathrm{resilience}}=0.42-0.25=0.17.$$

$$L={\displaystyle \frac{1}{2}}[{({\Delta }_{anx})}^2+\left({\Delta }_{res}{)}^2\right]={\displaystyle \frac{1}{2}}\left(0.0289+0.0289\right)=0.0289.$$

• Backpropagation Example

We illustrate how the error propagates and updates a few representative weights.

1.

Output Neuron for Math Anxiety

• ${\displaystyle \frac{\partial \mathrm{L}}{{\partial }_{{\mathrm{o}}_{\mathrm{anxiety}}}}}=\left({\mathrm{o}}_{\mathrm{anxiety}}-{\mathrm{t}}_{\mathrm{anxiety}}\right)=0.63-0.80=-0.17.$
• Sigmoid derivative:${\displaystyle \frac{{\partial }_{o_{anxiety}}}{{\partial }_{ne{t}_{anx}}}}=0.63\times \left(1-0.63\right)=0.63\times 0.37\approx 0.2331$

By the chain rule, for a weight w_anx,k:

$${\displaystyle \frac{\partial L}{{\partial }_{w_{anx,k}}}}={\displaystyle \frac{\partial L}{{\partial }_{o_{anxiety}}}}\times {\displaystyle \frac{{\partial }_{o_{anxiety}}}{{\partial }_{ne{t}_{anx}}}}\times {\displaystyle \frac{{\partial }_{ne{t}_{anx}}}{{\partial }_{w_{anx,k}}}},\mathrm{where}{\displaystyle \frac{{\partial }_{ne{t}_{anx}}}{{\partial }_{w_{anx,k}}}}=h_k^2.$$

2.

Hidden Layer 2

ReLU derivative: $ReL{U}^{\prime }\left(ne{t}_k^2\right)=1$, if $ne{t}_k^2>0$, else 0.

We sum the backpropagated gradients from both output neurons o_anxiety and o_resilience for each hidden neuron.

3.

Hidden Layer 1

• Similar process: each neuron receives combined gradients from the second hidden layer and uses ReLU’s derivative if ${\mathrm{net}}_j^1>0.$

4.

Gradient Descent Update

$$\Delta w=-\eta {\displaystyle \frac{\partial L}{\partial w}},\mathrm{for}\eta =0.01,\mathrm{and}{\displaystyle \frac{\partial L}{{\partial }_{w_{anx,k}}}}=-0.026,\mathrm{then}:$$

$$\Delta w_{anx,k}=-0.01\times \left(-0.026\right)=0.00026,\mathrm{thus}$$

$$w_{anx,k}^{new}=w_{anx,k}^{old}+0.00026=0.5+0.00026=0.50026.$$

It means that the math anxiety output weight was updated by a very small increment—specifically, increased by 0.00026. In practical terms, if the previous weight was 0.5, the new weight becomes 0.50026. This slight change is part of the gradient descent process, where weights are adjusted incrementally to reduce the prediction error, thus gradually improving the model’s accuracy.

Similarly, applying the same procedure for academic resilience, the updated weight is decreased by 0.00026. For the original weight was −0.2, then the new weight would be: $w_{res,k}^{new}=w_{res,k}^{old}-0.00026=-0.2-0.00026=-0.20026$.

• Interpretation

• Math Anxiety Weight Update

A representative math anxiety output weight increases by 0.00026 (e.g., from 0.5 to 0.50026).

• Academic Resilience Weight Update

In contrast, a representative academic resilience output weight decreases by 0.00026 (e.g., from −0.2 to −0.20026).

These weight adjustments, though small in a single update, accumulate over many training iterations (epochs) and samples, enabling the network to fine-tune its predictions for both math anxiety and academic resilience. Ultimately, this process helps the model better understand the shared and unique factors influencing each output, allowing it to deliver personalized, adaptive interventions in real-world educational settings.

3.10. Model Architecture Visualization and Validation

Figure 1 illustrates the neural network architecture employed in this study. The diagram clearly depicts the input layer, which includes demographic factors (age, gender, math performance), math anxiety indicators, and cognitive and metacognitive strategy scores. Data flow sequentially into two hidden layers comprising 64 and 32 neurons, respectively, each employing the ReLU activation function to effectively extract and refine relevant patterns. The final output layer consists of two neurons predicting post-intervention math anxiety and academic resilience scores, utilizing the Sigmoid activation function.

To validate the neural network model’s predictive capability, a 5-fold cross-validation approach was employed. This rigorous validation yielded an average accuracy of 92% and a root mean square error (RMSE) of 0.65. These metrics indicate robust predictive performance, affirming the reliability and applicability of the neural network model for personalized educational interventions aimed at reducing math anxiety and enhancing academic resilience.

4. Results

The descriptive statistics in

Table 1. Descriptive statistics of the data from the neural network.

Measure	Pre-Test M (SD)	Post-Test M (SD)
Math Test Scores	52.30 (8.45)	64.15 (10.20)
Math Anxiety Score	6.13 (1.28)	4.08 (1.13)
Cognitive Strategy Score	2.40 (0.75)	3.20 (0.60)
Metacognitive Strategy Score	2.20 (0.85)	3.10 (0.70)
Academic Resilience Score	2.15 (0.70)	2.90 (0.90)
N	56

provide a detailed overview of the changes observed across all measured variables from the pre-test to the post-test phases. The consistent improvements across all metrics underscore the effectiveness of the neural network-facilitated intervention in enhancing students’ learning outcomes.

The marked advancements in math test scores reduced math anxiety, and increased use of cognitive and metacognitive strategies reflect the intervention’s positive impact, particularly on academic resilience. These findings emphasize the potential of AI-powered educational tools to address both cognitive and emotional barriers to learning, thereby fostering a more supportive and effective educational environment (Polydoros, 2024a; Ventista et al., 2024b).

Overall, the pre- and post-test comparisons presented in Table 1 highlight substantial progress in academic performance, emotional well-being, and strategic learning behaviors (Polydoros, 2024b). These results validate the role of neural networks in facilitating not only academic achievements but also personal growth and resilience among students.

• Mixed ANOVA Results

A mixed ANOVA was conducted to analyze the differences in math test scores, math anxiety, cognitive strategy use, and academic resilience scores from the pre-test to the post-test.

• Math Test Scores

There was a significant main effect of time on math test scores, F(1, 54) = 54.00, p < 0.0001, η² = 0.49 (

Table 2. ANOVA results for math test scores.

Source	SS	df	MS	F	p	η2	Cohen’s d
Time (Pre-Test vs. Post-Test)	2800.00	1	2800.00	54.00	0.00 *	0.49	1.26
Error	2900.00	54	53.70

Note. SS (Sum of Squares) measures the total variability in the data, df (Degrees of Freedom) indicates the number of independent values that can vary, and MS (Mean Square) is the average variance per degree of freedom. The F-value is the ratio of explained variance to unexplained variance. The p-value shows the likelihood that the result is due to chance, with values less than 0.05 considered significant, η² (Eta Squared) indicates the proportion of variance explained by the independent variable. The effect size (Cohen’s d) was 1.26 (large effect), indicating a substantial improvement in math test scores following the intervention. * p < 0.01.

). This indicates that the intervention significantly improved math test scores from the pre-test (M = 52.30, SD = 8.45) to the post-test (M = 64.15, SD = 10.20).

• Results for Math Anxiety

A significant reduction in math anxiety scores was observed, F(1, 54) = 35.00, p < 0.0001, η² = 0.39, indicating that students experienced lower levels of anxiety post-intervention (M = 6.13, SD = 1.28) compared to pre-intervention (M = 4.08, SD = 1.13). The effect size (Cohen’s d) was 1.70 (large effect), reflecting a strong reduction in math anxiety post-intervention (

Table 3. ANOVA results for math anxiety scores.

Source	SS	df	MS	F	p	η2	Cohen’s d
Time (Pre-Test vs. Post-Test)	600.00	1	600.00	35.00	0.001 *	0.39	1.70
Error	1000.00	54	18.52

* p < 0.01.

).

• Results for Cognitive Strategies

The results indicate a significant improvement in the use of cognitive strategies from the pre-test to the post-test, as evidenced by the ANOVA analysis (F(1, 54) = 48.00, p < 0.0001, η² = 0.36). The effect size (Cohen’s d) was 1.17, indicating a strong improvement in cognitive strategy use. This improvement highlights the effectiveness of the intervention in enhancing students’ application of targeted learning techniques (

Table 4. ANOVA results for cognitive strategy use.

Source	SS	df	MS	F	p	η2	Cohen’s d
Time (Pre-Test vs. Post-Test)	120.00	1	120.00	48.00	0.00 *	0.36	1.17
Error	130.00	54	2.41

* p < 0.01.

).

The increase in mean scores from the pre-test (M = 2.40, SD = 0.75) to the post-test (M = 3.20, SD = 0.60) underscores the intervention’s impact on fostering better problem-solving and learning processes. The observed effect size (η² = 0.36) suggests a substantial impact of the intervention, affirming its success in promoting strategic cognitive engagement.

• Results for Metacognitive Strategies

The results from

Table 5. ANOVA results for metacognitive strategy use.

Source	SS	df	MS	F	p	η2	Cohen’s d
Time (Pre-Test vs. Post-Test)	120.00	1	120.00	48.00	0.00 *	0.36	1.14
Error	130.00	54	2.41

Note. * p < 0.01.

demonstrate a significant improvement in metacognitive strategy use over time (F = 48.00, p < 0.0001), with a large effect size (η² = 0.36). This indicates that the intervention had a strong impact on participants’ ability to apply self-monitoring, planning, and problem-solving strategies. The calculated effect size (Cohen’s d) was 1.14, indicating substantial gains in metacognitive abilities. The substantial improvement highlights the effectiveness of teaching metacognitive skills in fostering academic resilience and better performance. The scores rose from the pre-test (M = 2.20, SD = 0.85) to the post-test (M = 3.10, SD = 0.70).

• Results for Academic Resilience

A significant improvement in academic resilience was identified (F(1, 54) = 40.00, p < 0.01, η² = 0.42), reflecting an increase in students’ ability to adapt and persevere in academic challenges (

Table 6. ANOVA results for academic resilience scores.

Variable	SS	df	MS	F	p	η2
Time (Pre-Test vs. Post-Test)	1800.00	1	1800.00	40.00	0.00 *	0.42
Error	2200.00	54	40.74

Note. * p < 0.01.

). The scores rose from the pre-test (M = 2.15, SD = 0.70) to the post-test (M = 2.90, SD = 0.90).

• Regression Comparison for the Post-Test and Pre-Test Models

Regression comparison for both the post-test and pre-test models showed how different variables (math anxiety, cognitive strategy, and metacognitive strategy) influence academic resilience over time (post-test and pre-test). The post-test values suggested that the intervention had a stronger positive effect on academic resilience compared to the pre-test, with cognitive and metacognitive strategies showing a notable increase in their positive influence (

Table 7. Regression comparison table for academic resilience.

Variable	Post-Test	Pre-Test	Interpretation
Intercept	21.75	18.50	The baseline (constant) is higher for the post-test, indicating a higher starting point for resilience after the intervention.
Math Anxiety	−0.25	−0.35	Math anxiety has a weaker negative effect on academic resilience post-test compared to the pre-test.
Cognitive Strategy	3.15	2.80	Cognitive strategy use has a stronger positive impact on resilience after the intervention (post-test).
Metacognitive Strategy	2.10	1.90	Metacognitive strategies also show a stronger positive impact post-test.

).

• Correlation Comparison

The intervention significantly improved the relationships between cognitive/metacognitive strategies, math anxiety, and academic resilience (

Table 8. Comparison of variable relationships for pre- and post-test.

Variable Pair	Pre-Test	Post-Test	Change
Cognitive Strategies and Math Anxiety	−0.55 **	−0.62 **	Strengthened negative correlation
Cognitive Strategies and Academic Resilience	0.40 **	0.54 **	Strengthened positive correlation
Metacognitive Strategies and Math Anxiety	−0.28 *	−0.50 **	Strengthened negative correlation
Metacognitive Strategies and Academic Resilience	0.35 *	0.48 **	Strengthened positive correlation
Math Anxiety and Academic Resilience	−0.30 *	−0.45 **	Stronger negative correlation

Note. ** p < 0.01, * p < 0.05.

). This suggests that both cognitive and metacognitive strategies became more effective in managing anxiety and fostering resilience following the intervention, underscoring their critical role in improving educational outcomes for students.

5. Discussion

The incorporation of Information and Communication Technology (ICT) within the educational sphere has been validated by a multitude of empirical studies as a tool to improve understanding and boost student achievement (Drigas et al., 2020; Galitskaya & Drigas, 2019, 2020; Polydoros, 2021; Polydoros, 2024a; Ventista et al., 2024a). In our investigation, we implemented an intervention aimed at students facing academic difficulties by employing artificial intelligence—specifically, a neural network. Our findings corroborate that integrating AI into traditional pedagogical methodologies yields customized learning experiences and improves academic outcomes (Drigas & Petrova, 2014).

Notably, the pre-intervention test scores were significantly lower (M = 52.30, SD = 8.45) compared to the post-intervention scores (M = 64.15, SD = 10.20), indicating that AI-based tools provide superior educational benefits relative to conventional teaching strategies. This finding aligns with previous studies demonstrating enhanced learning outcomes and inclusive learning environments for all students, particularly those with learning disabilities (Drigas et al., 2021; Galitskaya & Drigas, 2021; Thomas et al., 2024; Polydoros & Antoniou, 2025).

Our work builds on earlier research by Polydoros (2021), which examined educational software for teaching fractions and its effect on students’ learning styles, and Polydoros (2024b), which investigated math anxiety in virtual classrooms during the COVID-19 pandemic and its relationship to academic achievement. These studies provide a strong foundation for our use of AI in addressing math anxiety. Additionally, Polydoros et al. (2025) have demonstrated that integrating AI technology in elementary geometry significantly impacts performance, anxiety levels, learning styles, cognitive styles, and executive functions, further supporting the potential of AI-based interventions.

In our study, we focused specifically on mathematics anxiety among seventh graders. Our results indicate that post-test anxiety scores (M = 4.08, SD = 1.13) were significantly lower than pre-test scores (M = 6.13, SD = 1.28), echoing findings from recent literature (Bressane et al., 2024; Jayaraman et al., 2024). Furthermore, AI-driven interventions have been shown to positively influence students’ perceptions of mathematics by enhancing academic performance and self-esteem, ultimately reducing anxiety (Atoyebi & Atoyebi, 2022; Ersozlu, 2024).

Our intervention also encouraged goal-directed learning strategies, facilitating the shift from critical to creative thinking (Poce, 2024; Roozafzai, 2024). Consistent with this, our participants exhibited significant improvements in cognitive strategy scores—from a pre-test mean of 2.40 (SD = 0.75) to a post-test mean of 3.20 (SD = 0.60)—and in metacognitive skills, with scores rising from 2.20 (SD = 0.85) to 3.10 (SD = 0.70) (Jäger et al., 2024; Walter, 2024). Polydoros (2024a) further underscores the benefits of harnessing AI and digital tools to enhance learning outcomes, especially for students with learning difficulties, by enabling micro-credentialing in mathematics education.

Finally, our results indicate that the AI-facilitated intervention led to improved academic performance and increased academic resilience. Specifically, resilience scores improved from a pre-test mean of 2.15 (SD = 0.70) to a post-test mean of 2.90 (SD = 0.90). AI possesses the potential to develop personalized intervention programs that take into account the distinct needs of each student (learning disabilities, academic performance, and learning styles) (Bressane et al., 2024; Strielkowski et al., 2024; X. Wang et al., 2024). These programs enhance students’ cognitive and meta-linguistic abilities, equipping them to address problems and bolster their mathematical resilience (Fitriani et al., 2023; Kouzalis et al., 2024).

The study specifically focused on first-degree inequalities as a representative mathematical topic suitable for evaluating the effectiveness of AI-driven interventions. The rationale behind choosing inequalities was due to their abstract nature and common difficulty among students experiencing math anxiety, making them an ideal context to test the intervention’s cognitive and emotional impact. Although our results highlight substantial improvements within this domain, generalizations to all mathematical topics should be made cautiously.

Thus, we can deduce that AI-based interventions positively influence student achievement, foster a favorable disposition toward mathematics, enhance self-esteem, alleviate mathematical anxiety, promote cognitive and metacognitive skills, and bolster academic resilience. This comprehensive approach caters to individual learning requirements and fosters a deeper comprehension of mathematical concepts, ultimately preparing students for future academic challenges. By amalgamating technology with personalized learning strategies, educators can cultivate an engaging and supportive environment that empowers students to excel in mathematics and beyond.

These findings suggest practical implications for educational practitioners, highlighting how AI-driven personalized instruction can effectively support students facing persistent academic difficulties. To enhance practical relevance and scalability, future implementations could explore adaptations for different age groups, curricular areas, or educational systems, potentially expanding the intervention’s benefits across broader contexts.

Limitations and Future Work

The findings were affirmative, as elaborated upon in the preceding section; however, it is crucial to acknowledge certain minor limitations that are inherent to the research. Our investigation was confined exclusively to seventh-grade students, thereby constraining the generalizability of the results to alternative age cohorts. Additionally, the intervention was conducted within a singular curricular unit focused on inequalities precluding the ability to ascertain whether similar outcomes would be observed in different units or other cognitive domains. Moreover, the intervention was executed in a specified educational context.

It should be noted that while efforts were made to minimize confounding factors by standardizing the intervention protocol and employing trained classroom teachers, the influence of teacher-specific factors, variability among schools, and contextual elements (e.g., student motivation, home environment) were not fully controlled and could have influenced outcomes. Future research should address these factors explicitly.

The research conducted is promising, and it would be advantageous to expand the scope to encompass other age groups or educational tiers. Additionally, a strategy could be applied across different mathematics teaching modules. Additionally, it would be pertinent to reassess the students after a designated period to investigate whether the knowledge acquired and the reduction in anxiety levels were sustained over the long term.

This follow-up evaluation could yield significant insights into the efficacy and durability of the intervention’s influence on students’ mathematical comprehension and emotional well-being. The implementation of such strategies could facilitate the development of more customized educational methodologies, ensuring that interventions are not only effective in the short term but also engender a lasting positive impact on students’ perceptions of mathematics.

6. Conclusions

This study demonstrates that the AI-based intervention significantly enhances the educational outcomes of seventh-grade students experiencing persistent academic underachievement in mathematics. The results indicate that the intervention not only improved mathematics performance but also effectively reduced math anxiety, fostered the development of cognitive strategies, and increased academic resilience. Specifically, integrating a neural network with two hidden layers—employing ReLU for efficient feature extraction and Sigmoid activations in the output layer—enabled precise predictions for both math anxiety and academic resilience, providing actionable insights for tailored interventions.

Moreover, our analysis reveals that the interrelationships among cognitive strategies, math anxiety, and academic resilience evolve following the intervention. Enhanced cognitive and metacognitive skills, coupled with reduced anxiety, suggest that targeted digital tools can mitigate the negative psychosocial factors affecting academic performance in mathematics.

Overall, these findings support the efficacy of AI-driven approaches in creating personalized, adaptive learning environments. By addressing both affective and cognitive dimensions, the intervention contributes to a holistic educational strategy that prepares students for future challenges in mathematics. Future research should explore the long-term effects and scalability of such interventions across diverse educational settings.