Predicting student dropout in Saudi Universities using machine learning and explainable AI

Factors influencing student dropout

Understanding the factors influencing student dropout is essential for developing effective intervention strategies. Age, gender, employment status, and marital status are among the key demographic factors influencing student dropout. Several studies have shown that younger students and males are generally more likely to dropout, particularly during the early academic years, often due to financial or personal responsibilities (Mnyawami, Maziku & Mushi, 2022; Gonzalez-Nucamendi et al., 2023; Cocorad et al., 2021; Kabashi, Shabani & Caka, 2022; Ashour, 2020). Employment has also been linked to higher dropout rates for both genders, as balancing work and study responsibilities can increase academic pressure and disengagement (Sandoval-Palis et al., 2020). Moreover, family-related obligations, such as marriage and childcare, have been identified as additional factors contributing to dropout among female students (Ashour, 2020).

Academic performance is one of the critical factors influencing student dropout (Singh & Alhulail, 2022; Kim et al., 2023; Costa et al., 2020). Early difficulties, low GPA, academic dissatisfaction, and poor grades in core courses are consistently linked to higher dropout risk (Murshed, 2021). However, several studies emphasize that academic difficulties alone cannot fully explain student dropout, as dropout often results from the combined influence of academic, personal, and social factors (Yair, Rotem & Shustak, 2020).

Social and cultural factors may significantly influence student dropout. In some cultures, high parental pressure toward specific career paths and early marriage may conflict with students’ personal goals, leading to academic difficulties, psychological problems, and dropout (Ashour, 2020; Alaoui, Sigamoney & Dogra, 2021). In the UAE, where cultural norms are similar to those of Saudi Arabia, family pressure and early marriage are the key factors affecting the dropout rate for both genders (Ashour, 2020). Language barriers may also play a role, especially when the language of instruction differs from the student’s native language (Orong et al., 2020). These findings highlight the need for further investigation of these sociocultural factors within Saudi Arabia.

Machine learning and deep learning in student dropout

Recent studies have extensively applied machine learning and deep learning models to predict student dropout. For instance, Hoca & Dimililer (2025) presented a predictive framework using several algorithms such as Support Vector Machines (SVMs), K-Nearest Neighbors (KNNs), Logistic Regression (LR), Gaussian Naive Bayes (GNB), Artificial Neural Networks (ANNs), Random Forest (RF), Decision Tree (DT), CatBoost (CBs), and Classification And Regression Trees (CARTs). Among them, CatBoost achieved the highest accuracy of 83%. Similarly, Revathy, Kamalakkannan & Kavitha (2022) applied various machine learning models to predict student dropout. The achieved accuracies were 94% for SVM and linear discriminant analysis (LDA), 95% for LR, and 97.6% for KNN.

In another study, Hassan, Muse & Nadarajah (2024) utilized data from the 2022 National Education Accessibility Survey to evaluate multiple models, including DT, RF, SVM, GNB, KNN, and LR, where the Random Forest model achieved the highest accuracy of 95%. Furthermore, Aini et al. (2025) applied several machine learning algorithms, including DT (C4.5), Naive Bayes, SVM, and KNN, to predict student dropout using university academic data. The C4.5 and SVM models achieved the highest accuracy of 94.44%, whereas KNN and Naive Bayes achieved approximately 93%.

A systematic review by Andrade-Girn et al. (2023) showed that model performance primarily depends on the type and quality of input data. Ensemble and tree-based algorithms such as RF, Gradient Boosted Tree, LightGBM, and XGBoost achieved the highest predictive accuracy (94–99%). Deep learning models (e.g., Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN)) also demonstrated competitive accuracy (88–97%) but required larger datasets and longer training times. Traditional models (e.g., LR, SVM) performed well on smaller datasets but were outperformed by ensemble methods.

Several studies have consistently shown the effectiveness of boosting algorithms and ensemble models over traditional ML models. For instance, Villar & Andrade (2024) reported that LightGBM and CatBoost achieved AUC values exceeding 90%, outperforming other models. Similarly, Fernandez-Garcia et al. (2021) found that Gradient Boosting was the best model at enrollment, whereas SVM achieved a recall of 91.5% in the fourth semester. Wang & He (2025) also evaluated multiple machine learning algorithms to predict undergraduate dropout among domestic and international students. Among the tested models, LightGBM achieved the highest performance for domestic students with an accuracy of 90.14%, while CatBoost produced the best results for international students with an accuracy of 90.04%. Other models such as XGBoost, Random Forest, and Gradient Boosting also showed comparable performance, each exceeding 89% accuracy.

Ensemble models that combine ML and DL also achieved high accuracy. Niyogisubizo et al. (2022) proposed a stacking model of RF, XGBoost, GB, and FNN, achieving an accuracy of 92.18%. Deep learning models, particularly the MLP, have also shown high performance. Delen, Davazdahemami & Dezfouli (2023) utilized an MLP to predict freshman student dropout with an accuracy of 88.4%. Similarly, Melo et al. (2022) reported a higher accuracy of 97% using MLP in predicting student dropout.

Explainable artificial intelligence in student dropout

The terms “interpretable” and “explainable” are used interchangeably in the field of XAI. Research by Burkart & Huber (2021) describes the difference between the two terms, as interpretability is used to describe the model’s internal workings, whereas explainability is used to explain the model results and identify which features affect the prediction. In this research, explainability techniques are used to explain the model’s results.

Among the widely used XAI techniques, SHAP has shown consistent performance in both global and local interpretability. Villar & Andrade (2024) used SHAP with CatBoost and LightGBM to visualize the top contributing features in dropout prediction. Similarly, Baranyi, Nagy & Molontay (2020) combined SHAP with permutation importance, identifying 12 key predictors based on both methods. Delen, Davazdahemami & Dezfouli (2023) also used SHAP to explain feature importance at both the global and individual levels.

Alwarthan, Aslam & Khan (2022) used LIME, a commonly used method, to identify the most influential features at the instance level. They also used a global surrogate model to explain the behavior of black-box models using interpretable algorithms. Melo et al. (2022) compared SHAP, LIME, and Shapley values using an XAI index to score each method’s overall performance. The results showed that SHAP achieved the highest score at 78%, whereas LIME scored 57% and Shapley values 35%.

Recent studies have emphasized the importance of evaluating explanation quality. For instance, Hoffman et al. (2023) proposed manual evaluation criteria, including checklists and user-rating scales for explainability. They introduced measures of explanation quality such as explanation goodness & satisfaction, user curiosity, user trust, and user and XAI performance. While such subjective approaches provide valuable insights, they depend heavily on user perception. To address this limitation, Amparore, Perotti & Bajardi (2021) proposed the Local Explanation evAluation Framework (LEAF),which offers a more systematic way to evaluate XAI methods such as SHAP and LIME. The framework evaluates explanations by measuring different metrics, including fidelity, local concordance, prescriptivity, and stability. Compared to earlier subjective evaluations, LEAF enables a consistent and objective comparison between explanation methods.

Student dropout in Saudi Arabia

Various studies have been conducted to predict student dropout from universities in Saudi Arabia. For example, Alwarthan, Aslam & Khan (2022) predicted at-risk students in the preparatory year at Imam Abdulrahman bin Faisal University using Synthetic Minority Over-sampling Technique (SMOTE) and feature selection methods such as Recursive Feature Elimination (RFE) and a Genetic Algorithm (GA). They used ANN, SVM, and RF models, which achieved accuracies of up to 99.6%. To enhance explainability, SHAP, LIME, and a global surrogate model were also applied. Similarly, Al-Sulami, Al-Masre & Al-Malki (2023) focused on predicting final grades and identifying at-risk students in an English foundation course at King Abdulaziz University. The study used vRNN, LSTM, and GRU models, with GRU achieving the best accuracy of 93.65% in the first course and 98.9% in the second. In addition, Uliyan et al. (2021) analyzed first- and second-year retention rates of preparatory students and graduation rates of Computer Science and Engineering students at the University of Ha’il. The dataset included grades, assessment results, participation, and attendance records from 35,000 students. They applied Bidirectional LSTM (BLSTM) and Conditional Random Fields (CRFs), achieving an accuracy of 90%. Similarly, Said et al. (2024) used data from a Saudi university that included 2,444 records of Information Systems and Computer Science students to predict both final grades and dropouts. They tested several models, including KNN, SVM, DT, LR, RF, GNB, MLP, and Extra trees (ET), achieving accuracies ranging from 59.7% to 80.8%. The ensemble Majority Voting (MV) model achieved the highest accuracy of 92.7%. Although these studies demonstrate the growing use of machine learning for student dropout prediction in Saudi universities, most of them are limited to single-institution datasets and focus on specific academic programs. The current research addresses this gap by creating a more diverse dataset covering multiple universities and academic programs.

Data preparation

This phase involves data collection and preprocessing to prepare the dataset for modeling. Data were collected through an online questionnaire distributed to students from various Saudi universities, including both dropouts and non-dropouts. The questionnaire was designed based on prior studies and refined through a pilot study to ensure its clarity and relevance. In addition, the questionnaire was distributed through multiple social media platforms, including Twitter, WhatsApp, and Telegram, and was also shared by various universities through their official student portals to enhance the diversity of responses. In total, responses were obtained from 30 public and several private universities across Saudi Arabia.

A non-probability convenience sampling approach was used since participation was voluntary and there was no centralized sampling frame for Saudi university students. This approach is commonly used when probability sampling is impractical in multi-institution contexts. Potential selection bias is acknowledged and mitigated by distributing the questionnaire broadly across universities (Ahmed, 2024).

Sample adequacy was evaluated using the events-per-variable (EPV) criterion, which measures the ratio between the number of dropout cases and the number of model factors. The dataset included 1,016 dropout cases and 25 factors, resulting in an EPV of 40.6. This value exceeds the minimum adequacy threshold ( $\ge$10 EPV) for reliable and stable predictive modeling. A higher EPV value also supports model stability and reduces the risk of overfitting by ensuring that each factor is informed by a sufficient number of outcome events (Pavlou et al., 2024).

Ethical approval for the study was granted by the Department of Information Systems at King Abdulaziz University. The research involved only academic and institutional data and did not include any private information of participants. In addition, informed consent was obtained from participants at the beginning of the survey. Participants were informed that completing the questionnaire indicated their voluntary agreement to participate. They were also notified that their anonymized responses would be used solely for research purposes.

Data preprocessing is an essential step in data analysis that transforms raw data into a structured format (Pandey, Patnaik & Gupta, 2020). In this step, duplicate data were removed, categorical variables were encoded, and class imbalance was addressed using the Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN). The SMOTE-ENN hybrid method improves data quality by first generating synthetic samples for the minority class and then applying ENN to remove noisy or misclassified instances (Batista, Prati & Monard, 2004). Compared to standalone oversampling methods (e.g., SMOTE, ADASYN, Borderline-SMOTE, SMOTE-Tomek), the hybrid SMOTE-ENN provides more reliable results by combining synthetic oversampling with noise reduction, effectively reducing class overlap and improving model performance (Husain et al., 2025; Mduma, 2023; Cuevas-Chvez et al., 2024).

Modeling and evaluation

This phase includes feature selection and training and evaluating ML models to predict student dropout.

Explainability models and evaluation

To enhance the explainability of the machine learning predictions, this research applied SHAP and LIME. These techniques were applied to explain individual predictions and identify the most influential features of student dropout.

SHAP is a method proposed by Lundberg & Lee (2017) to explain individual predictions and describe how a model’s predictions are influenced by each feature. It is based on the Shapley value, an approach from cooperative game theory that measures the contribution of each player to the result.

LIME was introduced by Ribeiro, Singh & Guestrin (2016) to explain individual predictions by perturbing the input data and training an interpretable surrogate model to approximate the behavior of the black-box model.

In addition, the LEAF framework was used to evaluate the quality of the SHAP and LIME explanations through four metrics: fidelity, local concordance, prescriptivity, and stability (Amparore, Perotti & Bajardi, 2021).

Data collection

The data collection began with a pilot study to ensure the clarity and relevance of the questionnaire. Based on the feedback, the final version was distributed to students from various universities to collect data for the research. Data were collected between July 2024 and January 2025.

Data preprocessing

The data preprocessing steps resulted in the removal of 39 duplicate records and the mapping of categorical features (e.g., gender, age, employment status) to their corresponding numerical values. In addition, as shown in (Fig. 2), the non-dropout class contains 77.5% of the dataset, while the dropout class contains only 22.5%, which represents a highly imbalanced class distribution that can negatively affect the performance of the ML model. To address this imbalance, the SMOTE-ENN was implemented with a sampling strategy of 0.7, which balances the minority (dropout) class to 70% of the majority (non-dropout) class. Default neighborhood parameters were used ( $k_{neighbors}$ = 5 for SMOTE and $n_{neighbors}$ = 3 for ENN) following standard configurations in the imbalanced-learn library. Using SMOTE-ENN resulted in a more balanced dataset, as shown in (Fig. 3), with similar sample sizes for both classes (Class 0: 55.7%, Class 1: 44.3%).

Feature selection

Four feature selection methods, including F-statistic, Chi-square, RFE, and MI, were applied to generate five feature subsets consisting of 22, 19, 16, 13, and 10 features. The feature subsets were generated systematically by ranking features according to their importance scores within each selection technique and iteratively removing the three lowest-ranked features at each step. This process continued until only 10 features remained, as performance began to decline with further reduction.

Figure 4 presents the comparison of the top 22 selected features across the four feature selection methods. This 22-feature subset represents the first and most comprehensive configuration before the reduction to smaller subsets. It was used to examine the relationships between the features and to compare how each selection method ranked them according to its statistical criterion.

The results show that Age, Employment Status, and Marital Status consistently obtained the highest importance across all methods. Year and University Name also maintained high and stable rankings across the techniques. However, GPA showed a different pattern compared with the other factors. It had moderate ranks under the F-statistic, Chi-square, and RFE methods, but became the most influential under MI. This result indicates that the relationship between GPA and dropout is non-linear, where students with lower GPAs are more likely to dropout. Similarly, Family/Friends Support showed a clear non-linear pattern. It ranked among the lowest under the F-statistic, Chi-square, and RFE methods, but became the second most influential factor under MI. In contrast, Gender showed an opposite pattern. It was ranked first under the linear-based methods but dropped to the lowest position under MI, which suggests that its predictive influence is primarily linear. It is also worth noting that three features were excluded from each feature selection method. The F-statistic and Chi-square methods did not identify City Moving, Living Expenses, and English Language as important factors. The RFE method did not identify Living Expenses, English Language, and Peer Influence. In contrast, the MI method did not identify City Moving, Health Issues, and Studying Difficulty among the influential factors. These variations reflect the methodological differences between linear and non-linear approaches, as each method captures distinct aspects of the data distribution and feature relationships.

ML model training and evaluation

Each algorithm was trained using empirically optimized parameters to ensure stable performance. Default settings from the scikit-learn library were applied unless otherwise specified. For instance, the KNN model used $k=5$ and the RF used $200$ estimators with a maximum depth of $10$. For the Gradient Boosting and XGBoost algorithms, a learning rate of $0.05$ was selected to achieve a balance between training stability and predictive performance. Using a smaller learning rate helps the model learn more gradually, reducing the risk of overfitting while maintaining high accuracy across iterations. All models were trained with a fixed random state ( $random\mathrm{\_}state=42$) to ensure reproducibility.

In addition, models were evaluated using an 80/20 train–test split to ensure a consistent comparison across the twelve algorithms and feature subsets. As shown in (Tables 2–9), each model was evaluated across the four feature selection techniques: F-statistic, chi-square, RFE, and MI. For each technique, the models were trained using the full set of features as well as subsets of 22, 19, 16, 13, and 10 features.

Tables 2 and 3 present the accuracy and F1-score results of various machine learning models using the F-statistic feature selection. Overall, KNN achieved the highest performance in both metrics, with an accuracy and F1-score of 97.5% when using all features. The voting classifier and CB also achieved strong performance. In contrast, models such as LDA, GNB, and LR had lower accuracies and F1-scores, especially when the number of features was reduced.

Tables 4 and 5 summarize the results using the chi-square feature selection technique. Similar to the F statistic results, the KNN and ensemble models maintained strong performance across different feature sets. However, model performance decreases as the number of selected features decreases.

Tables 6 and 7 present the results of RFE. This technique also showed strong performance in the KNN and ensemble models. The accuracy and F1-score of most models remained stable when using 16–22 features but started to decline as the number of features was reduced.

Tables 8 and 9 summarize the results obtained using MI. While there were some differences in model performance, the overall trend remained consistent with that of the other feature selection techniques.

However, KNN was the best performing model, achieving an accuracy and F1-score of 97.6% with the 22 features selected through RFE, as shown in (Tables 6 and 7). The corresponding Wilson Confidence Interval (95% CI) was [96.3–98.4%], indicating stable and reliable performance. To better understand why the RFE configuration achieved the best results, the relative importance of the selected factors was examined. The RFE results showed that factors such as gender, age, employment status, marital status, academic year, GPA, failed courses, and academic warnings were among the top contributing features. Factors such as studying difficulty, university satisfaction, language barrier, major preknowledge, and family/employer pressure showed moderate contributions. In contrast, factors such as cultural expectation, city moving, commitments, family/friends support, psychological issues, and health issues had weaker contributions. Additionally, peer influence and living expenses were excluded entirely by the RFE algorithm. Accordingly, KNN was selected as the primary model for the explanation phase.

Explainability model results and evaluation

In this phase, the KNN model was selected due to its superior performance when using the 22 features identified through RFE. The explanation process focused on analyzing individual students to explore how the model reached its prediction in different situations. Several student cases were selected and analyzed using SHAP and LIME. In each case, the model predicted that the student would/would not drop out with high confidence, assigning a prediction probability of 1.00 in both the SHAP and LIME explanations. The average dropout and non-dropout were subsequently calculated using SHAP and LIME to identify the most influential factors for each class. A SHAP summary plot was also presented to provide an overview of feature importance across all the students. In addition, the LEAF framework was used to evaluate the quality of these explanations through four metrics: fidelity, local concordance, prescriptivity, and stability.

Dropout Case 1 The SHAP waterfall plot (Fig. 5) shows that the most influential factor in the prediction was the faculty name “Preparatory Year”, which alone contributed $+0.39$ toward the prediction. This finding indicates that the preparatory year was strongly associated with higher dropout rates in the dataset. This finding is consistent with those of previous studies, which reported that most students tend to drop out during their first year. Other key factors include the university name “PNU” ( $+0.11$) and the academic year “First Year” ( $+0.09$), both of which contribute positively to the prediction. The students lack of family/friend support, the presence of two failed courses, and family/employer pressure also played roles in student dropout.

The LIME explanation (Fig. 6) supports this result by highlighting similar factors. It shows that the faculty and university name (preparatory year in PNU), a lower academic year, and a higher number of failed courses (more than 1) strongly lead to a dropout prediction. Despite the students having some positive indicators, such as prior knowledge of their major and no language barrier, these indicators were outweighed by stronger dropout factors.

Dropout Case 2 The SHAP waterfall plot (Fig. 7) shows that the greatest contributing factor was the GPA “Good”, which added ( $+0.18$) to the output. Other major contributors were the university “KKU” ( $+0.17$), the academic year “Second Year” ( $+0.13$), lack of support from family/friends ( $+0.13$), and lack of prior major knowledge ( $+0.10$). Together, these features led to a dropout prediction. Additional factors included family/employer pressure, the faculty name “College of Art & Design,” and the presence of failed courses, each adding a small but consistent contribution ( $+0.03$).

The LIME explanation (Fig. 8) supports this result by highlighting the same top factors. This finding shows that the dropout decision was driven by factors such as low GPA, lower academic year, no prior major knowledge, and lack of support from family/friends. However, the student had a few factors that reduced dropout risk, such as being unemployed, having no language barriers, and being single, but these positive factors were outweighed by more influential dropout factors.

Non-Dropout Case 1 The SHAP plot (Fig. 9) shows that the most influential factor reducing the likelihood of dropout was the faculty name “College of CS and IT”, which contributed ( $-0.15$) to the prediction. This suggests that students in this faculty are less likely to drop out based on patterns in the dataset. Another strong factor was the student’s GPA, “Excellent” ( $-0.10$). This finding is consistent with previous research showing that higher academic performance is strongly associated with lower dropout rates. Other features, such as the academic year “Above Third Year” ( $-0.06$), employment status “Unemployed” ( $-0.02$), and the university name “KSU” ( $-0.02$), also contributed to a lower dropout risk. A few features contributed slightly to dropout, such as the student’s age ( $+0.07$), language barrier ( $+0.05$), and marital status ( $+0.04$), but their influence was outweighed by stronger academic and social factors.

The LIME explanation (Fig. 10) confirms this result, highlighting that a high GPA between 2.0 and 3.0, an advanced academic year, and strong support from family/friends support are among the top factors reducing dropout risk.

Non-Dropout Case 2 According to the SHAP plot (Fig. 11), the strongest factor influencing the prediction was the student’s GPA, which contributed ( $-0.14$) to the non-dropout prediction. Other features that reduce the likelihood of dropout include the faculty name “College of Architecture & Planning” ( $-0.06$), the university name “YIC” ( $-0.05$), and the academic year “Above Third Year” ( $-0.04$). Minor contributions also came from the presence of support from family/friends and the student’s employment status, i.e., “unemployed.”

The LIME plot (Fig. 12) supports these findings. It highlights a GPA above 3.0, advanced year level, unemployment, and the student’s university and faculty name as top contributors toward the non-dropout prediction. Additional features, such as marital status, the presence of family/friends support, and prior major knowledge, support the prediction.

SHAP summary

The SHAP summary plot (Fig. 13) presents the average impact of each feature on the model’s dropout predictions. GPA was the most influential factor, where lower GPA values (in blue) were strongly associated with higher dropout likelihood. This finding aligns with previous research indicating that academic performance plays a key role in predicting student dropout.

The academic year also had a strong influence as students in earlier years were more likely to drop out. Thus, students are most vulnerable at the beginning of their academic journey to drop out. As shown by the red points, being employed (part-time, full-time, or owning a business) was linked to a greater likelihood of dropping out. Similarly, lower-level program types (e.g., diplomas or bachelor’s degree programs) were associated with a higher dropout probability. Other contributing factors included lack of family/friends support, dissatisfaction with the university, no prior major knowledge, and being married. Interestingly, older students also showed an increased likelihood of dropping out, suggesting that age-related responsibilities may influence student dropout. Overall, the distribution of SHAP values supports the conclusion that personal, academic, and sociocultural factors jointly affect student dropout.

LEAF evaluation

LEAF was applied to both SHAP and LIME to assess the quality of the model explanations. Table 14 summarizes the definitions of each LEAF metric: Fidelity, Local Concordance, Prescriptivity, and Stability.

Table 14:

LEAF metrics description.

Metric	Description
Fidelity	The extent to which the explanation matches the model’s actual prediction.
Local concordance	How well the explanation matches the model’s confidence in its prediction.
Prescriptivity	The extent to which the explanation offers actionable insight.
Stability	The consistency of the explanation when small changes are made to the input data.

DOI: 10.7717/peerj-cs.3490/table-14

As shown in (Table 15), LIME outperforms SHAP across all four metrics. In terms of fidelity, the LIME explanations matched the actual predictions, with a score of 0.80, compared with 0.72 for SHAP. Similarly, local concordance was higher for LIME (0.74) than for SHAP (0.68). For prescriptivity, LIME achieved a score of 0.96, whereas SHAP scored 0.88. Finally, in terms of stability, LIME again scored higher (0.94) than SHAP did (0.85). In addition, the differences across all metrics were statistically significant ( $p<0.05$). These results demonstrate that LIME produces more reliable and actionable explanations than SHAP for the student dropout prediction model.

Table 15:

LEAF evaluation results with statistical significance test.

Metric	LIME	SHAP	p-value
Fidelity	0.80	0.72	0.037
Local concordance	0.74	0.68	0.041
Prescriptivity	0.96	0.88	0.028
Stability	0.94	0.85	0.018

DOI: 10.7717/peerj-cs.3490/table-15

Figure 14 provides a visual comparison between LIME and SHAP based on the four LEAF evaluation metrics. This illustrates that LIME achieved higher scores than SHAP did for all the metrics. This finding supports the results shown in (Table 15) and confirms that LIME is consistently more effective.

Future work

Future research directions for this research include expanding the dataset by collecting a larger sample from different universities. This would enable the application of more complex deep learning and transformer-based architectures to capture more nonlinear dropout patterns. The design could also be extended by incorporating repeated k-fold cross-validation and statistical testing (e.g., Friedman or Wilcoxon) to further enhance performance comparisons across multiple runs. In addition, researchers in other regions with similar academic structures, sociocultural characteristics, or dropout risk factors are encouraged to replicate this study using the same set of factors. Such cross-context validation would help determine whether these factors remain strong factors of student dropout in different educational contexts. Finally, collaborating with academic institutions can help translate these findings into practical interventions, such as early warning systems or targeted support programs. These efforts may enhance student retention and contribute to achieving educational goals.