💻 computer science

Machine Learning Prediction of Undergraduate Academic Performance for Early Identification of At-Risk Students: Evidence from Nnamdi Azikiwe University

This study demonstrates that machine learning models, particularly Support Vector Machines, can effectively predict undergraduate academic risk at Nnamdi Azikiwe University using routine records like first-semester GPA and carryover courses, thereby enabling timely interventions for at-risk students.

Original authors: Chinenye Amaonye, Judith Nneka OKAFOR, Samuel Oweh Abang, Ngozi Eucharia Anyikwa, Juliana Uzonna UMEOZOR

Published 2026-10-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chinenye Amaonye, Judith Nneka OKAFOR, Samuel Oweh Abang, Ngozi Eucharia Anyikwa, Juliana Uzonna UMEOZOR

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Universities today face a quiet but growing challenge: they are educating more students than ever before, yet they often lack the time or tools to notice when a specific student begins to struggle until it is too late. In the past, an educator might wait for a failing grade or a missed exam to realize a student was in trouble, but by then, the gap in learning has often become too wide to bridge easily. To solve this, researchers are turning to a field called machine learning, which is essentially a way for computers to learn from patterns in data rather than following a strict set of rules written by a human. By feeding these systems historical records of student performance, attendance, and background, the computer can learn to spot the subtle signs that precede academic failure. The goal is not to replace teachers with algorithms, but to give them an early warning system—a way to see which students might need help before they fall behind, allowing for timely support that can change a student's entire academic trajectory.

At Nnamdi Azikiwe University in Nigeria, a team of researchers decided to test whether this approach could work in their specific context. They gathered the academic records of 400 undergraduate students, looking at a wide range of details including their entrance exam scores, their grades from the first semester, how often they attended lectures, their scores on continuous assessments, and even whether they had failed any courses in the past. The researchers then asked four different computer models to analyze this information and predict which students would be at risk of poor academic performance. They defined "at-risk" as students whose performance fell below a certain standard, effectively creating two groups: those doing well and those who were struggling. The team wanted to see if the computers could identify the struggling group early on, using only the data available at the start of the semester, and to determine which specific factors were the strongest indicators of trouble.

The results revealed a clear picture of what separates a thriving student from one who is in danger. Out of the 400 students studied, nearly one in four was identified as being at risk. When the researchers looked closely at the data, they found that the students who were struggling shared a very specific profile. They had entered the university with lower entrance exam scores, but more importantly, their performance during their first semester was a much stronger signal of future trouble. The most powerful predictor was the grade point average from the first semester; students who were at risk had an average score of 2.43, while those doing well averaged 3.24. The number of courses a student had failed previously, known as carryover courses, was the second most important factor. Students who were at risk had failed an average of 2.5 courses, compared to just one for their peers. Other factors like continuous assessment scores and lecture attendance also played a role, with struggling students attending fewer classes and scoring lower on ongoing tests. Interestingly, the data showed that a student's gender or the specific way they entered the university did not help predict who would struggle, suggesting that academic behavior matters far more than background demographics in this setting.

To find the best way to make these predictions, the researchers compared four different machine learning techniques, each of which approaches the problem in a slightly different way. One method, called a decision forest, builds many simple decision trees to make a collective guess. Another, known as a support vector machine, tries to draw a complex line that separates the struggling students from the successful ones. When the models were tested on new data they had never seen before, the results were surprising. The model that performed best during the initial training phase did not hold up as well when faced with new students. Instead, the support vector machine and a method called logistic regression proved to be the most reliable. The support vector machine correctly identified 16 out of the 20 struggling students in the test group, missing only four, while also keeping the number of false alarms—students flagged as struggling when they were actually doing fine—relatively low. The computer models were able to distinguish between the two groups with a high degree of accuracy, proving that the patterns in the data were real and consistent.

The study also looked deeply into why the models made the decisions they did, ensuring the results were not just a statistical trick but reflected real educational dynamics. The analysis confirmed that what a student does during their first semester is far more telling than where they started. While entrance exam scores did matter, they were overshadowed by how the student performed once they were actually in the classroom. The computer learned that a low grade point average, a high number of failed courses, and poor attendance were the strongest signals that a student was in trouble. This aligns with the idea that learning is an active process; students who are not engaging with their studies, missing classes, or failing to grasp the material early on are the ones who need help the most. The researchers noted that while the models were highly effective at spotting these patterns, they should be viewed as a tool to support human judgment, not to replace it. A computer can flag a student for attention, but it is the university's counselors and teachers who must provide the understanding and support that follows.

Ultimately, this research demonstrates that universities can use the data they already collect to build a safety net for their students. By focusing on early indicators like first-semester grades and course attendance, institutions can identify students who are likely to struggle long before they fail a final exam. The study suggests that Nnamdi Azikiwe University, and others like it, could implement these early-warning systems to guide their academic advising. Instead of waiting for a crisis to unfold, staff could reach out to students the moment the data shows they are drifting off course. The findings offer a practical path forward: use the patterns in the records to trigger timely intervention, ensuring that the students who need help the most receive it before their academic difficulties become impossible to overcome.

Technical Summary: Machine Learning Prediction of Undergraduate Academic Performance for Early Identification of At-Risk Students

Problem Statement
The study addresses the challenge of identifying undergraduate students at risk of poor academic performance early enough to allow for effective intervention. While universities possess vast amounts of student data, there is often a lack of validated, institution-specific predictive frameworks to utilize this data before poor outcomes become entrenched. The research specifically targets Nnamdi Azikiwe University (NAU) in Nigeria, aiming to move beyond post-hoc analysis to proactive identification. The core problem is determining whether machine learning (ML) models can reliably classify students as "at-risk" or "not at-risk" using routine academic records available early in a student's journey, and which specific variables drive these predictions.

Methodology
The research employs a quantitative predictive design using secondary data from 400 undergraduate students at NAU.

Data Sources and Variables: The dataset includes admission scores (UTME, Post-UTME), O'level points, first-semester GPA, continuous assessment scores, attendance rates, number of carryover courses, age, gender, and mode of entry.
Target Variable: Academic performance was dichotomized into two classes: "at-risk" (performance below a defined threshold) and "not at-risk." In the sample, 99 students (24.8%) were classified as at-risk.
Preprocessing: Data was cleaned for missing values and duplicates. To address class imbalance (24.8% at-risk), the Synthetic Minority Over-sampling Technique (SMOTE) was applied strictly within the training folds to prevent data leakage. The dataset was split into training (320 students) and independent test (80 students) sets.
Models Evaluated: Four supervised machine learning algorithms were trained and compared:
1. Logistic Regression
2. Random Forest
3. Support Vector Machine (SVM) with Radial Basis Function (RBF) kernel
4. XGBoost
Training and Validation: Models underwent 10-fold cross-validation on the training set for hyperparameter tuning (via grid search). Final performance was evaluated on the held-out test set using Accuracy, Precision, Recall, F1-score, and ROC-AUC.
Interpretability: Feature importance was analyzed using Random Forest's built-in impurity-based importance and SHAP (SHapley Additive exPlanations) values to determine the direction and magnitude of variable influence.

Key Results

Descriptive Statistics: Significant differences ( $p < .001$ ) were found between at-risk and non-at-risk groups across almost all academic variables. At-risk students had lower admission scores, lower continuous assessment scores, lower attendance, and significantly more carryover courses. First-semester GPA showed the widest gap (2.43 for at-risk vs. 3.24 for non-at-risk). Gender and mode of entry showed no significant difference between groups.
Model Performance (Cross-Validation): During 10-fold cross-validation, Random Forest achieved the highest accuracy (0.859) and F1-score (0.733), followed closely by Logistic Regression and SVM. All models demonstrated strong discriminatory power, with ROC-AUC values ranging from 0.887 to 0.911.
Model Performance (Independent Test Set): On the unseen test data, the performance ranking shifted. SVM achieved the highest accuracy (0.825), F1-score (0.696), and ROC-AUC (0.911). Logistic Regression performed similarly well (Accuracy: 0.812, Recall: 0.800). Conversely, Random Forest, which led in cross-validation, yielded the lowest test performance (Accuracy: 0.762, F1: 0.578), suggesting potential overfitting or instability with the smaller test sample.
Feature Importance: Both impurity-based importance and SHAP analysis identified First-semester GPA (importance ~~0.317/SHAP ~0.143) and Carryover courses (~~0.226/SHAP ~0.122) as the dominant predictors, collectively accounting for over 50% of the model's predictive power. Continuous assessment was the third most important factor. Admission scores (UTME, Post-UTME) and O'level points contributed moderately, while gender and mode of entry had negligible impact.
Error Analysis: In the test set, SVM and Logistic Regression correctly identified 16 of the 20 at-risk students (Recall = 0.80), whereas Random Forest identified only 13.

Key Contributions

Institution-Specific Evidence: The study provides empirical evidence from a Nigerian public university context, addressing a gap where most existing literature focuses on Western institutions, online learning environments, or specific disciplines.
Comparative Model Analysis: It offers a direct comparison of Logistic Regression, Random Forest, SVM, and XGBoost within this specific demographic, highlighting that cross-validation performance does not always guarantee superior test-set performance.
Variable Hierarchy: The research establishes that early university performance indicators (First-semester GPA, carryovers) are more predictive of future risk than pre-admission characteristics (UTME scores) or demographic factors (gender, entry mode) in this context.
Interpretability: By utilizing SHAP values, the study moves beyond "black box" predictions, confirming that the models rely on logical, educator-observable factors (e.g., low GPA and high carryovers push the prediction toward "at-risk").

Significance and Claims
The paper claims that machine learning can serve as a viable tool for early warning systems in higher education, specifically for Nnamdi Azikiwe University. The findings suggest that:

Early Identification is Feasible: Reliable predictions can be made using data available early in the academic cycle (first semester).
Data-Driven Intervention: Universities can use these models to trigger timely academic advising and support programs rather than waiting for final examination failures.
Complementary Role: The authors emphasize that these systems should complement, not replace, professional judgment. Predictions are intended as "early warning signals" to trigger support, not as permanent labels of student ability.
Caveats: The study modestly notes limitations, including the relatively small sample size (400 students) and the small test set (80 students), which may affect the stability of performance metrics. It recommends that such systems be validated with larger datasets and across different cohorts before widespread implementation.

The study concludes that while no single algorithm is universally superior, SVM and Logistic Regression showed robust performance on unseen data in this specific context, and the integration of such predictive tools into student support programs is a recommended step for improving academic outcomes.

Technical Summary: Machine Learning Prediction of Undergraduate Academic Performance for Early Identification of At-Risk Students

More like this