Artificial intelligence in student management systems to enhance academic performance monitoring and intervention

SMS have significantly developed over the past decades. Early systems in the 1980s and 1990s primarily focused on administrative tasks such as enrollment, attendance, and grade recording¹³. These systems operated mainly as digital versions of paper records, offering little support for learning analytics or academic monitoring. With the rise of internet technologies, web-based SMS emerged in the early 2000s. They introduced online portals for students and teachers, enabling access to grades, schedules, and course materials remotely¹⁴. However, these platforms lacked predictive capabilities and did not support early academic interventions. Recent advancements have shifted SMS towards intelligent, data-driven platforms. Integration of LMS and SMS has facilitated the collection of fine-grained student social interaction data, such as submission of assignments, participation in forums, or even online activity¹⁵. It is this data from which learning analytics can identify at-risk students and provide support measures to help. This progress has not yet changed many existing systems to reactive instead of proactive. Missing the opportunity to intervene during the learning process, they report academic performance only after assessments are completed¹⁶. Additionally, the adoption of AI to boost SMS functionalities is still limited, mostly in experimental or isolated applications¹⁷. The interest in learning analytics and AI in education has always been rising, from administrative support to student success support. Previous research has shown that these days it is considered a crucial next step to integrate predictive analytics in SMS to create more effective, responsive, and personalized education environments¹⁸. Recent studies have also highlighted the broader implications of generative AI on socioeconomic disparities in education, emphasizing the importance of equitable and responsible AI integration in learning environments¹⁹. Over the last few years, AIs have increasingly been used to improve learning experiences and outcomes in educational settings. The first set of applications of early AI was intelligent tutoring systems (ITS) in which instructional materials were adapted to student performance²⁰. The goal was to make these learning systems as similar in behavior as possible to one-on-one human tutoring, characterized by personalized feedback and learning paths. AI in education extends beyond tutoring, as machine learning and big data have revealed many other ways to apply AI to education. Now that predictive analytics models are used to predict student performance, institutions can discover at-risk learners early and offer targeted interventions²¹. For instance, submission of early coursework and online activities can be used to predict final grades²². It has also greatly contributed to natural language processing (NLP). Real-time support that includes answering administrative and academic queries efficiently is given to students by AI-powered chatbots²³. Similarly, sentiment analysis tools also help assess student feedback to improve the course content and teaching strategies²⁴. Additionally, recommendation systems recommend personalized learning resources to the students to assist them in navigating an abundance of educational content according to their interests and needs²⁵. The focus is on adaptive systems that would make the learning experiences more effective and personalized. Despite these advancements, challenges remain. Many AI models are developed and tested in limited or isolated environments. Thus, it isn’t easy to generalize the findings in different institutions or populations²⁶. Finally, there are concerns regarding fairness, transparency, and ethical use of student data in deploying AI solutions²⁷. Education is an excellent area for AI to take over the market by becoming a highly responsive, personalized, and proactive learning environment. However, it must be carefully designed, validated, and ethically overseen to be successfully integrated into large-scale SMS. Retrospective evaluation methods for monitoring academic performance have included exams, assignments, and attendance records. Nevertheless, in recent years, data mining and machine learning have brought some proactive techniques that can predict the outcome of the students before the final assessments²⁸. Performance monitoring techniques, such as classification models, are among the most used techniques. For example, Decision Trees like the C4.5 algorithm can use demographic and academic behavior of at-risk students to identify them with clear decision rules²⁹. Unlike other ensemble learning methods, Random Forests only combine multiple Decision Trees to further increase the prediction accuracy³⁰. Continued learning data has also been successfully used to classify students using the support vector machines (SVMs). They are instrumental in cases where the data has many academic and behavioral features³¹. Student behaviors have been predicted to be at risk of dropout by learning hidden relationships between them and academic outcomes using deep learning models³². While powerful, these models depend on large datasets and careful tuning to avoid overfitting. K-means clustering techniques are used often and tend to group students in different groups depending on their learning behaviors or performance patterns. Such segmentation makes it possible to provide differentiated interventions for each group¹². However, regression models are still relevant, as they are invaluable at predicting continuous outcomes such as final grades. Linear Regression and its variants can predict a numeric academic result based on early-term data⁶. In addition to model selection, feature engineering is important for successfully monitoring academic performance. Often, the features that provide significant predictive power include features like login frequency, assignment submission patterns, forum participation, etc., which are often demographic variables³³. Now, there are recent studies that show that the best predictive accuracy often results from using several different models or even the hybrid use of multiple and different systems². Thus, to develop a robust performance monitoring framework that would apply to new educational data, careful model comparison, validation, and continuous update are necessary. It has enabled timely and personalized intervention strategies as well as academic monitoring that is better than before. Machine learning powered early warning systems can identify at-risk students during the course, providing the institutions with early warning signs of failure before it occurs³⁴. One such intervention model is the personalized feedback system. The tailored feedback that comes out is generated by AI algorithms that analyze the student submissions and engagement patterns to present feedback that speaks to each individual’s learning needs³⁵. Using such systems increases motivation and academic performance by giving relevant, actionable advice. Another widely adopted strategy is to predict early alerts. Purdue University’s ‘Course Signals’ is an example of a system that sends automated alerts to at-risk students based on predictive analytics models³⁶. These alerts encourage students to seek help and adopt better study habits. Recommendation engines powered by AI suggest remedial resources, such as additional reading materials, tutorials, or peer mentoring, based on a student’s weak areas³⁷. These systems customize learning pathways, ensuring that interventions are adaptive and responsive. Adaptive tutoring systems also leverage AI to adjust the difficulty of exercises and instructional content in real time. Platforms such as Carnegie Learning’s ITS adapt lesson content based on a student’s performance on earlier tasks³⁸. Some institutions have implemented chatbots and virtual coaches that guide students through administrative and academic challenges. These AI agents can answer FAQs, recommend study plans, and even refer students to human advisors when necessary²³. Although AI-driven interventions show promise, they also raise challenges. Issues such as student privacy, data security, and the risk of algorithmic bias must be carefully managed to ensure fair and ethical interventions³⁹. Integrating AI into intervention strategies can significantly enhance student success, provided that systems are designed with transparency, accuracy, and sensitivity to diverse learner needs. Although the integration of AI into educational systems has progressed, several important research gaps remain. A critical limitation is that many predictive models are built and validated using a single dataset, which limits their generalizability across diverse educational environments⁶. Studies often focus on secondary or higher education, rarely combining multiple contexts for a holistic view⁴. Another gap lies in the intervention strategies themselves. Most current systems send generic alerts or feedback without dynamically adapting interventions based on individual student profiles³⁶. Furthermore, while academic performance monitoring is well-explored, few studies address the ethical dimensions of AI deployment, such as algorithmic fairness and student privacy³⁹. Furthermore, comparability in several aspects of different AI models, datasets, and educational settings is lacking. However, most works validate only one machine learning technique that misses opportunities to identify optimal models for different student populations²¹. These gaps demonstrate the necessity of a study built from several datasets and using multiple AI models that also looks at the relationship between academic and ethical results that are driven by the use of AI to manage students. This research seeks to fill the gaps by providing a holistic, multi contexts understanding of AI enhanced SMS, where the technical performance meets with ethical responsibility.

Research methodology

This study proposes a hybrid AI framework with a multilayer machine learning system to reduce scholarly performance and risk prediction. The research leverages the strengths of Decision Trees, Random Forests, SVM, and Neural Networks and combines them in a way that does not rely on a single model. The hybrid model combines multiple algorithms to enhance the algorithms’ accuracy, robustness, and generalizability on different educational datasets. For this purpose, it uses three real-world datasets: the UCI Student Performance dataset, the OULAD, and the NELS:88. The hybrid model is developed, trained, and evaluated in a structured and ethical manner using the methodology.

Research design and approach

The research employs a quantitative experimental approach. First, the data from the three selected datasets are collected, and then preprocessed, i.e., missing data is handled, categorical variables are encoded, and features are normalized. Once preprocessed, each machine learning model: Decision Tree, Random Forest, SVM, and Neural Network- is trained separately on each dataset. These models then provide their outputs using voting and stacking methods of hybrid ensemble strategies. Once the model is integrated with the hybrid system, it is evaluated based on accuracy, precision, recall, F1 score, and AUC-ROC. The hybrid model thus has the advantage of the strengths of individual algorithms and the disadvantages eliminated. This focuses on data privacy, fairness in predictive analytics, and ethical use throughout the process. The overall workflow of the proposed research methodology is shown in Fig. 1. Data is collected from the UCI Student Performance data, the OULAD, and the NELS:88 for the process to start. Data is preprocessed, feature engineered, and four independent machine learning models, namely Decision Tree, Random Forest, SVM, and Neural Network, are trained. To fully exploit the strengths of diverse classifiers, these models are combined through a stacking ensemble strategy, where the outputs of base learners serve as meta-level inputs, thereby improving overall generalization across the dataset.

Datasets description

We employ three publicly available benchmark datasets: UCI Student Performance, Open University Learning Analytics Dataset (OULAD), and NELS:88, to ensure robust validation and generalizability of the proposed model across varied academic contexts. This study trains and validates the proposed hybrid model using three publicly available educational datasets. Each dataset has a different number of attributes, different levels of academia, and different sizes, giving a wide variety of environments to evaluate in the learning environment.

UCI student performance dataset

The UCI Student Performance dataset⁷ comes from two secondary schools in Portugal and deals with the students’ academic performance in the mathematics and Portuguese language courses. The UCI Student Performance dataset is publicly available and can be accessed at:

https://archive.ics.uci.edu/ml/datasets/Student+Performance. The data set has 649 instances, and 33 attributes describe each instance. Such attributes include demographic (e.g., age, gender, parental education), academic (e.g., study time, number of failures), and social (e.g., family relationships, free time) indicators. The students’ grades at three stages, G1 (first period), G2 (second period), and G3 (final grade), are the primary target variables. The dataset includes around 62% of the students being female, and the average final grade (G3) is around 10.4 out of 20. This dataset is widely used to perform the prediction of performances, and it can be used for regression and classification modeling.

Open university learning analytics dataset

The OULAD⁸ was collected from The Open University, a large online higher education institution in the United Kingdom. The OULAD dataset is publicly accessible at:

https://analyse.kmi.open.ac.uk/open_dataset. It includes detailed information about 32,593 students, 22 courses (modules), and their interactions with the virtual learning environment (VLE). The dataset comprises seven interconnected tables covering student demographics, assessment scores, course registration, activity logs, and withdrawal status. Key attributes include clickstream activity (such as the number of online resources accessed), assessment results, final grades, and withdrawal date, if applicable. Approximately 61% of the students in the dataset are female, and 14% withdrew before completing their courses. OULAD supports predictive modeling for outcomes such as course completion, academic success, and dropout prediction.

National educational longitudinal study

The NELS:88⁹ was initiated by the U.S. Department of Education to track students’ educational progress over time. The NELS:88 dataset is hosted by the U.S. National Center for Education Statistics and can be accessed via:

https://nces.ed.gov/surveys/nels88/. It began with a nationally representative sample of approximately 24,599 eighth-grade students from 1,052 schools across the United States. The dataset includes variables covering student demographics, family background, academic achievement, attitudes toward education, and postsecondary outcomes. Follow-up surveys were conducted in 1990, 1992, and 1994, allowing researchers to analyze longitudinal trends. Important measured outcomes include standardized test scores, high school graduation status, college enrollment, and employment information. NELS:88 enables the exploration of long-term academic performance patterns and the factors influencing educational attainment.

Data preprocessing and feature engineering

Effective data preprocessing and feature engineering are essential to optimize predictive performance and ensure model robustness. In this study, the dataset consisted of diverse feature types, including demographic attributes, academic performance records, and behavioral interaction logs. This multi-dimensional feature set was carefully prepared through a structured preprocessing pipeline to enhance model accuracy and generalization. Preprocessing began with missing value imputation. For numerical features, missing entries were replaced with the feature mean, calculated as:

where \(\:{x}_{i}\)​ represents the observed values, and \(\:n\) is the total number of non-missing entries. For categorical variables, the mode of the feature was used for imputation.

Categorical features were encoded using one-hot encoding to transform non-numeric variables into binary vectors. Each categorical variable with \(\:k\) possible categories were transformed into \(\:k\) binary features, allowing machine learning models to interpret them correctly without assuming any ordinal relationship.

Feature scaling was applied to ensure that numerical attributes operated on a similar range, particularly for algorithms sensitive to feature magnitudes, such as SVM and Neural Networks. Standardization was used, transforming each feature \(\:x\) according to:

where \(\:\mu\:\) is the mean and \(\:\sigma\:\) is the standard deviation of the feature. This rescaling ensures that each feature has zero mean and unit variance.

Feature selection was performed using a two-step approach. First, correlation analysis was conducted to remove redundant features that are highly correlated with each other (\(\:r>0.85\)). Pearson’s correlation coefficient \(\:r\) between two features \(\:x\) and \(\:y\) was calculated by:

where \(\:\stackrel{-}{x}\) and \(\:\stackrel{-}{y}\)​ are the means of features \(\:x\) and \(\:y\) respectively. Second, feature importance was evaluated using Random Forests, where attributes contributing significantly to model performance were retained. Features with low importance scores, below a threshold \(\:\theta\:=0.01\), were discarded.The integration of heterogeneous feature types—demographic, behavioral, and academic—proved vital for model accuracy. Their complementary nature enriched the decision-making process within the hybrid ensemble, enabling it to identify complex academic risk patterns more effectively across all three datasets.

AI models and algorithms employed

This research utilizes a hybrid architecture that combines multiple machine learning algorithms to improve the accuracy and robustness of academic risk prediction. Four primary models were independently developed: Decision Trees, Random Forests, SVM, and ANN. Each algorithm contributes distinct strengths to the overall hybrid model.

Decision Trees classify instances by recursively partitioning the feature space based on feature values that maximize information gain. At each node, the algorithm selects the feature \(\:{X}_{j}\)​ that provides the maximum reduction in entropy, calculated as:

where \(\:H\left(Y\right)\) represents the entropy of the target variable, and \(\:H(Y\mid\:{X}_{j})\) denotes the conditional entropy given the feature \(\:{X}_{j}\)​. This enables the tree to form a series of hierarchical decision rules that are easy to interpret.

Random Forests, an ensemble extension of Decision Trees, improve model stability and accuracy by constructing multiple trees on bootstrap samples of the dataset and aggregating their predictions through majority voting. For regression or probability estimation, the Random Forest output \(\:\widehat{f}\left(x\right)\) is the average of the outputs from individual trees:

where \(\:T\) is the number of trees and \(\:{f}_{t}\left(x\right)\) is the prediction from tree \(\:t\).

SVM finds the optimal hyperplane that separates data points of different classes by maximizing the margin between them. Given training data points \(\:({x}_{i},{y}_{i})\), where \(\:{y}_{i}\in\:\{-\text{1,1}\}\), the SVM solves the following optimization problem:

where \(\:w\) is the normal vector to the hyperplane and \(\:b\) is the bias term. Kernel functions were applied to enable non-linear classification by projecting data into higher-dimensional spaces.

ANN simulates the functioning of biological neurons. Each neuron computes a weighted sum of its inputs, applies an activation function \(\:\sigma\:(\cdot\:)\), and outputs a value:

where \(\:{w}_{i}\)​ are the weights, \(\:{x}_{i}\)​ the input features, and \(\:b\) the bias. This study used a multi-layer perceptron (MLP) architecture with one hidden layer to capture complex, non-linear relationships between input features and academic outcomes.

The outputs of these four models were integrated into a hybrid ensemble using both soft voting and stacking approaches. In soft voting, predicted class probabilities from each base model are averaged, and the final class label is assigned to the class with the highest mean probability. In stacking, a meta-learner is trained to combine the predictions from the base models, learning to correct their weaknesses and enhance overall predictive performance. Algorithm 1 describes the steps of the proposed hybrid machine learning model for academic risk prediction. It processes data from three different educational datasets, trains multiple classifiers, integrates their outputs using soft voting, and evaluates performance using standard metrics.

By combining diverse classifiers in a hybrid framework, the proposed system benefits from the interpretability of Decision Trees, the robustness of Random Forests, the margin maximization of SVMs, and the non-linear learning capacity of Neural Networks, resulting in a powerful academic risk prediction tool.

Evaluation metrics and validation methods

Five widely accepted classification metrics—accuracy, precision, recall, F1-score, and AUC-ROC, were used to assess the proposed hybrid model’s performance. These metrics evaluate the model’s ability to predict academic risk and correctly balance false positives and negatives. Accuracy measures the proportion of correctly classified instances over the total number of predictions. It is calculated as:

where \(\:TP\) is true positives, \(\:TN\) is true negatives, \(\:FP\) is a false positive, and \(\:FN\) is a false negative. Precision reflects how many of the predicted positive cases are truly positive. It is defined as:

Recall, also known as sensitivity, measures the proportion of actual positives that were correctly identified:

The F1-score is the harmonic mean of precision and recall, providing a balanced metric that considers both false positives and false negatives:

An assessment of how well the model discriminates across thresholds was also performed using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). It is the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative one by the model, and hence AUC. The AUC ranges from 0.5 (random guessing) to 1.0 (perfect classification). The stratified 10-fold cross-validation was used to achieve robustness and generalizability. Each fold was partitioned so that the dataset was 90% training and 10% testing, and the class distribution was preserved. The average of all folds’ performance scores was calculated as the final score. These metrics collectively provide a balanced view of both the classification performance (accuracy, precision, recall, F1-score) and model discriminative ability (AUC-ROC), ensuring a comprehensive evaluation of the hybrid model.