STUDENT MENTAL HEALTH AND BURNOUT PREDICTION SYSTEM

-Tanish Lodha

ABSTRACT

This project presents the design and implementation of a predictive analytics system capable of identifying students at risk of depression and burnout. Unlike traditional survey-based assessment methods that rely on subjective self-reporting, this prototype utilizes machine learning algorithms—integrating demographic data, academic performance metrics, and psychological indicators—to predict mental health conditions. The system analyzes student data including age, gender, CGPA, year of study, anxiety, and panic attack indicators to determine depression risk and enable early intervention.

INTRODUCTION

Using traditional mental health screening methods is complex, requiring significant counselor time and resources which most educational institutions cannot afford. Students often hesitate to self-report due to stigma, leading to delayed or incomplete data. Thus, to maintain cost effectiveness as well as accessibility, we have used machine learning models along with structured survey data to predict depression risk. This project aims to create a transparent, educational, and cost-effective prediction pipeline using a Python-based architecture.

PROBLEM STATEMENT

Educational institutions typically rely on highly resource-intensive systems such as periodic psychological assessments, one-on-one counseling sessions, or manual survey evaluations for student mental health screening. These methods are often time-consuming, cost-prohibitive, or suffer from low student participation due to stigma barriers.

We referred to student mental health statistics in higher education institutions as our case study. Universities worldwide report increasing rates of depression, anxiety, and burnout among students. However, replicating comprehensive mental health screening across large student populations is challenging; limited counseling staff cannot meet growing demand, and late identification of at-risk students often leads to severe academic and personal consequences.

This problem could be mitigated by implementing an automated, data-driven screening system that identifies at-risk students early, allowing counselors to prioritize follow-up and intervention.

Currently, most universities rely on students to self-refer for mental health support. Using a predictive system that can automatically analyze existing student data reduces manual workload, removes the chance of human oversight, and allows institutions to identify at-risk students much earlier.

We address the problem of designing and implementing a predictive system for student mental health screening that is capable of analyzing depression risk using accessible academic and demographic data without requiring expensive psychological assessments or specialized equipment. The system must accurately predict depression risk, identify key contributing factors, and provide interpretable results for counselors and administrators.

  OBJECTIVES

• Developing a machine learning algorithm that enables the prediction of depression risk in students based on accessible academic and demographic data without requiring clinical assessment.
• Integrating multiple predictive models including Logistic Regression, Decision Tree, and Random Forest to establish a reliable baseline for mental health prediction.
• Creating a responsive evaluation framework capable of detecting patterns, identifying at-risk students, and providing early warning indicators while maintaining prediction accuracy.

  EXISTING SYSTEMS

While predictive systems for mental health exist, most of them use complex deep learning models or require extensive clinical data. Both these solutions are quite resource intensive or require specialized expertise that may not be feasible for many educational institutions. In order to make our project accessible for everyone and easy to maintain, we have used traditional machine learning algorithms with structured survey data to predict depression risk.

High-end clinical mental health assessment tools that utilize comprehensive psychological evaluations are pre-existing and are being researched upon currently. These systems employ standardized questionnaires, clinical interviews, and professional assessment scales. When a student completes these assessments, mental health professionals analyze the results to determine risk levels. While this approach offers high reliability, it requires significant investment in trained personnel and time, making it impractical for large-scale, routine screening.

There are also sophisticated AI-based mental health prediction systems that utilize deep learning and natural language processing. These systems typically analyze social media posts, text responses, or speech patterns to detect mental health conditions. While these systems are highly autonomous and effective, they are computationally intensive, raise privacy concerns, and often rely on complex infrastructure that exceeds the scope of a cost-effective, educational prototype.

The traditional survey-based screening approach closely aligns with our project and has served as inspiration. These methods rely on self-reported questionnaires that measure various mental health indicators. However, our approach enhances this by applying machine learning to identify patterns that humans might miss, while using the same accessible survey data.

  SYSTEM OVERVIEW

• The system operates through an integrated pipeline that maps student data to mental health predictions, enabling institutions to autonomously identify at-risk students.
• The system begins by loading the student dataset. Each student record contains information about demographics, academic performance, and mental health indicators.
• Upon start-up, the system preprocesses the data by standardizing string values, converting categorical variables to numeric format, and handling missing values.
• The user (counselor or administrator) can input new student data or analyze existing records. The system processes the corresponding features and generates predictions about depression risk.
• Prediction is handled by multiple machine learning models including Logistic Regression, Decision Tree, and Random Forest, which convert input features into probability scores.
• To ensure the system maintains reliable predictions, we implemented cross-validation and hyperparameter tuning to optimize model performance.
• A critical design feature is class weighting. Because the dataset contains more non-depressed than depressed students, the system applies higher weights to the minority class to prevent bias toward predicting "No Depression."
• The system also generates a composite mental health score by summing depression, anxiety, and panic attack indicators, providing a quick measure of overall mental health burden.
• After prediction, the system outputs a classification report showing precision, recall, and F1-score for both depressed and non-depressed students.
• The system can save results and export predictions for counselor follow-up.

  MATERIALS USED

Our predictive system is constructed using Python as the primary programming language, designed with a modular structure to give adequate functionality and flexibility while keeping computational requirements low. It is equipped with multiple machine learning algorithms including Logistic Regression, Decision Tree, and Random Forest for accurate prediction. Data processing is aided by Pandas and NumPy libraries, which ensure efficient data manipulation and numerical computation. The entire system is managed through Jupyter Notebook or Python scripts, which coordinate data loading, preprocessing, model training, and evaluation efficiently. The system runs on standard computing hardware requiring approximately 4GB RAM and any modern processor.

Key Technologies

SOFTWARE AND PROGRAMMING

The entire platform is programmed using Python within the Jupyter Notebook environment. The code below walks through the full implementation: from data loading and preprocessing, through feature engineering, to model training and evaluation.

1. Load Dataset

2. String Standardization & Gender/Marital Normalization

3. Year of Study Cleaning

4. CGPA Parsing

5. Binary Encoding

6. Missing Value Handling

7. Label Encoding

8. Course Grouping & One-Hot Encoding

9. Mental Health Score Computation

10. Final Dataset Export

11. Feature Selection & Train/Test Split

12. Logistic Regression Training

13. Decision Tree Classifier

14. Model Definitions (Pipeline)

15. Cross-Validation & Evaluation Loop

16. Grid Search – Logistic Regression

17. Grid Search – Decision Tree

18. Random Forest with Class Weighting

  RESULT AND PERFORMANCE EVALUATION

The results and performance evaluation of the system demonstrated that the models operate reliably and effectively to predict depression risk. The Random Forest classifier was able to predict depression with consistent accuracy, achieving 95% cross-validation accuracy. The models responded well to training data, resulting in stable and reliable predictions across different subsets of students. The ensemble approach using Random Forest provided sufficient predictive power and generalization, maintaining high performance on unseen test data. Additionally, the class weighting technique ensured consistent performance on the minority class without significant drops in precision. Overall the system showed efficient predictive capability, good generalization to new data, and stable performance across evaluation metrics, indicating that the design and integration of both data preprocessing and modeling were successful.

Performance Metrics Summary

  ADVANTAGES AND LIMITATIONS

Advantages

• By utilizing ensemble machine learning methods, the system maintains high predictive accuracy. The cross-validation technique ensures that even if the model sees slightly different training data, it maintains reliable performance.
• The use of a structured preprocessing pipeline ensures the system's behavior is predictable and reproducible. Unlike complex deep learning models that can be difficult to debug, this rule-based preprocessing allows for easy troubleshooting and performance verification.
• The system uses standard, open-source Python libraries and runs on basic computing hardware. This reduces the total project cost significantly compared to AI-based solutions while providing sufficient prediction capability for routine mental health screening.
• The non-invasive data collection method—using only survey responses—ensures the system can operate in educational settings without requiring expensive equipment or specialized training.

Limitations

• The prediction logic assumes the input data is honest and accurately reported. If students provide inaccurate responses, the predictions will be unreliable.
• The system does not account for temporal changes in mental health. It provides a snapshot based on current data rather than tracking trends over time.
• Because the system relies on pre-defined features, it cannot dynamically incorporate new types of information without manual updates to the feature set.
• The current model is designed for individual institutional use. It does not support cross-institutional analysis or federated learning across multiple universities.

  POTENTIAL IMPROVEMENTS

• Adding time-series analysis would allow the system to detect and track mental health trends over multiple survey periods. This provides an independent check on cross-sectional predictions, creating a more comprehensive assessment framework.
• While complex, upgrading to include deep learning models could allow the system to identify more subtle patterns in the data and potentially achieve higher accuracy.
• The current system uses static feature selection. Integrating automated feature selection algorithms would allow the system to identify the most predictive variables dynamically for different student populations.
• While currently operating on structured survey data, integrating natural language processing would allow for analysis of open-ended student responses, providing richer insights into mental health status.
• Adding a real-time dashboard would allow counselors to visualize at-risk students, track prediction trends, and prioritize follow-up interventions more effectively.
• To make the system truly autonomous for continuous screening, a scheduled retraining protocol should be added. The system would periodically retrain models on new data to maintain accuracy as student populations change over time.
• The system has strong potential to support sustainable development in the future. By contributing to more efficient mental health screening and early intervention, it aligns with the goals set by the United Nations under Sustainable Development Goal 3, which focuses on good health and well-being. Such technologies can play an important role in building healthier, more supportive educational environments for the future.

  FUTURE SCOPE

• While the current system uses standard machine learning models, it lacks deep learning capabilities. Future work will involve implementing neural network architectures which could allow the system to capture more complex, non-linear relationships in the data, potentially improving prediction accuracy.
• The existing architecture is designed for single-institution operation. Future iterations will focus on a federated learning protocol, where multiple institutions could train models collaboratively without sharing sensitive student data. This would enable more robust models trained on diverse populations while maintaining privacy.
• To overcome the current dependency on structured survey data, future development will integrate natural language processing on student writing samples or social media data. By using transformer-based models, the system could identify linguistic patterns associated with depression risk, allowing for more passive and continuous screening.
• For true proactive intervention, the system requires integration with counseling center workflows. Future work will include the design of an automated alert system. By setting risk thresholds, the system can automatically notify counselors when students are identified as high-risk, enabling immediate follow-up.
• To expand the variety of predictive features, the system could integrate academic system data including attendance records, grade changes, and course withdrawal patterns. By detecting sudden drops in academic performance, the system could identify potential mental health crises even before students report symptoms.
• A future application of this project may include integration with campus wellness programs wherein the system would automatically recommend personalized interventions based on identified risk factors, such as stress management workshops, counseling sessions, or peer support groups.

  CONCLUSION

The development of the Student Mental Health and Burnout Prediction System demonstrates that effective mental health screening can be achieved through machine learning rather than relying solely on resource-intensive clinical assessments. By transitioning from a reactive, self-referral model to a proactive, data-driven prediction system, this project has successfully established a scalable and cost-effective framework for student mental health screening.

The system's core achievement lies in its structured preprocessing pipeline and ensemble modeling approach. By decomposing complex prediction tasks into defined preprocessing steps and model training phases, the system maintains reliable accuracy even with imbalanced data. The integration of class weighting techniques ensures that the model pays appropriate attention to the minority class, prioritizing the identification of at-risk students.

Furthermore, this project validates that a traditional machine learning approach, when combined with thorough preprocessing and cross-validation, is capable of reliable operation for mental health prediction in educational settings. While the prototype currently functions as a batch prediction system, the modular nature of its code provides a robust foundation for future expansions, such as real-time prediction and counselor alert systems.

Ultimately, this project highlights the essential synergy between data preprocessing, thoughtful feature engineering, and appropriate model selection. It provides a transparent, modifiable, and educational platform that serves as a blueprint for institutional mental health screening, proving that effective prediction is not just a function of model complexity, but a result of elegant, well-integrated system design.

  FINAL OUTPUT

The Student Mental Health and Burnout Prediction System is designed to predict depression risk using six key features: age, gender, year of study, CGPA, anxiety status, and panic attack status. The system continuously preprocesses input data to standardize string values, convert categorical variables to numeric format, and handle missing values. When a student's data is processed, the Random Forest model analyzes the features and outputs a depression risk prediction. The system achieved 95% cross-validation accuracy and 88% test accuracy. After applying class weights, the model achieved a recall of 78% for depression detection, meaning it successfully identified approximately 78% of depressed students. The model maintained a precision of 88% for depression detection, meaning when it predicts depression, it is correct nearly nine times out of ten. This resulted in an F1-score of approximately 83% for the depression class, with an AUC score of 97%, demonstrating a strong ability to distinguish between depressed and non-depressed students. Real-time performance metrics are printed for monitoring and evaluation, making the system a reliable and effective platform for early identification of students at risk of depression.

  REFERENCES

[1] World Health Organization. (2022). World mental health report: Transforming mental health for all. Geneva: WHO Press. ISBN 978-92-4-004933-8. Retrieved from https://www.who.int/publications/i/item/9789240049338

[2] Eisenberg, D., Hunt, J., & Speer, N. (2013). Mental health in American colleges and universities: Variation across student subgroups and across campuses. Journal of Nervous and Mental Disease, 201(1), 60–67. https://doi.org/10.1097/NMD.0b013e31827ab077

[3] Kessler, R. C., Barker, P. R., Colpe, L. J., Epstein, J. F., Gfroerer, J. C., Hiripi, E., ... & Zaslavsky, A. M. (2003). Screening for serious mental illness in the general population. Archives of General Psychiatry, 60(2), 184–189. https://doi.org/10.1001/archpsyc.60.2.184

[4] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. Retrieved from https://jmlr.org/papers/v12/pedregosa11a.html

[5] Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321–357. https://doi.org/10.1613/jair.953