AI Flags Stressed and Struggling Students Before They Drop Out

Every year, universities lose students not because they lack ability, but because life outside the lecture hall becomes unmanageable. A new study from researchers at Amrita Vishwa Vidyapeetham in India suggests that an algorithm can spot those students early, and, crucially, explain exactly why it flagged them. Writing in the journal Discover Artificial Intelligence, a team led by Vinayak Hegde and Pallavi M. Shanthappa describes an explainable machine learning framework that predicts dropout risk and sorts students into high, medium, and low risk categories, giving administrators a transparent decision-support tool rather than an inscrutable black box.

The problem the researchers set out to tackle is far from niche. Student dropout carries heavy costs at every level: for individuals, an unfinished degree often means unfulfilled ambitions, narrower career prospects, and wasted tuition; for institutions, attrition erodes revenue and performance indicators; for society, it shrinks the skilled workforce and is associated with higher unemployment and lower living standards. Decades of research have shown that attrition is rarely caused by a single factor. Academic performance, attendance, financial pressure, psychological well-being, health, motivation, and family circumstances all intertwine, which is precisely why early identification has proven so difficult.

Machine learning has been applied to this challenge before, often with impressive headline numbers. Previous studies have reported accuracies ranging from around 70 percent to above 98 percent using algorithms such as LightGBM, CatBoost, XGBoost, support vector machines, and deep learning ensembles built on CNN and LSTM networks. But the Amrita team identified a persistent gap: most of these systems focus narrowly on academic indicators, treat dropout as a simple yes-or-no question, and offer little insight into why a particular student was flagged. Advanced models frequently operate as black boxes, which makes administrators hesitant to act on their outputs.

To build their dataset, the researchers surveyed 253 undergraduate students at their institution between October and November 2025 using a structured Google Form distributed through class WhatsApp groups. Participation was voluntary, informed consent was obtained, and no names or registration numbers were collected. The survey captured eight features spanning academic and non-academic territory: cumulative grade point average, attendance, academic stress level, financial difficulty, health issues, course interest, thoughts of discontinuing studies, and personal or family problems. Responses were normalized and transformed through ordinal encoding, which mapped graded answers such as Never, Sometimes, and Often, or Low, Medium, and High, onto numerical values of 0, 1, and 2.

That encoding choice turned out to matter. By preserving the gradational nature of survey responses rather than flattening them into binary indicators, the researchers gave the models richer patterns to learn from. The target variable itself was engineered: rather than asking students directly whether they had dropped out, the authors constructed a dropout-risk label from predefined conditions combining academic performance, attendance, stress, finances, health, course interest, discontinuation intention, and family problems. The team candidly acknowledges a limitation here, since self-reported discontinuation thoughts served both as an input feature and as one condition defining the target, a dependency that may inflate measured performance.

Three classifiers were trained on a stratified 70:30 split of the data, keeping class proportions consistent across the 177 training and 76 test records: Logistic Regression, a Decision Tree, and a Random Forest, all implemented in Python with the Scikit-learn library and configured with balanced class weights to counteract imbalance. Logistic Regression actually posted the highest raw accuracy at 0.92, but the Random Forest won out on the metrics that matter most for this task. It achieved a precision of 0.96, a recall of 0.88, an F1-score of 0.92, and an area under the ROC curve of 0.911, indicating excellent discrimination between the two engineered classes. Because missing an at-risk student carries a high cost in institutional settings, the team prioritized the balance of precision and recall over raw accuracy and selected the ensemble model as the final predictor.

The most striking result came from the feature importance analysis. Stress emerged as the single most influential factor associated with dropout risk, followed by financial difficulties, health issues, and thoughts of discontinuation. Academic staples such as CGPA and attendance ranked comparatively low. In other words, the model learned that psychological and socioeconomic pressures outweigh grades when it comes to predicting who will walk away. This aligns with a growing body of literature emphasizing well-being and financial strain in student retention, but it challenges the grades-and-attendance focus of many traditional early warning systems, suggesting institutions should widen the lens through which they identify struggling students.

Where the framework departs most clearly from prior work is in its risk stratification and explainability layers. Predicted probabilities from the Random Forest are bucketed into three tiers: High Risk at 0.90 or above, Medium Risk between 0.70 and 0.90, and Low Risk below 0.70. The authors are careful to note these thresholds are heuristic demonstration values, not statistically optimized or calibrated boundaries, and that the raw probabilities should be read as relative confidence scores rather than true likelihoods. For individual explanations, the system multiplies each feature’s global importance score by the student’s encoded value, ranks the results, and surfaces the top three contributing factors. The team frames this honestly as a feature-importance-based heuristic rather than a full local explanation method like SHAP or LIME, and stresses that the identified factors are influential features, not causal claims.

The whole pipeline is wrapped in a working prototype built on the Flask web framework. The trained model is serialized with Joblib and loaded at runtime; uploaded CSV files are validated, preprocessed with the same encoding rules used in training, and scored in batch. A dashboard built with HTML, CSS, JavaScript, and Chart.js presents an overview page with risk distribution pie charts and trend visualizations, plus a per-student view showing dropout probability, color-coded risk level, and the top three reasons behind each prediction. High-risk students appear in red, medium in yellow, and low in green, allowing counselors to see at a glance who may need academic counseling, financial assistance, or health support, and who simply warrants periodic monitoring.

The researchers are appropriately measured about what their system can and cannot do. The dataset is small and drawn from a single institution, responses are self-reported and may carry bias, and the model was evaluated on a single train-test split without external validation. No formal usability study with administrators or counselors has yet been conducted, so the dashboard’s real-world effectiveness in improving retention remains unproven. Still, the direction of travel is clear: future work will pursue probability calibration, larger multi-institutional datasets, longitudinal records, real-time integration with student management systems, and fairness audits. If those steps pan out, the quiet signal buried in a student’s stress level or bank balance could reach a counselor months before the withdrawal form is ever filed.

Subject of Research: Explainable machine learning for early prediction and stratification of student dropout risk in higher education

Article Title: An explainable machine learning framework for early student dropout risk prediction and stratification

Article References: Hegde, V., Palaniswamy, A., Bhar, D., Revanna, M., Prakash, N., Prabhu, A., & Shanthappa, P. M. (2026). An explainable machine learning framework for early student dropout risk prediction and stratification. Discover Artificial Intelligence, 6(1), Article 1224. https://doi.org/10.1007/s44163-026-02288-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02288-8

Keywords: student dropout, machine learning, random forest, explainable AI, educational data mining, risk stratification, feature importance, higher education, Flask dashboard, predictive modeling, student retention, explainable