Classifying student stress

The question

Stress affects students well past how they feel — it shows up in academic performance and in whether the ordinary business of a week gets done. Measuring it is the hard part, because almost every instrument relies on a student reporting on themselves, which is the one input nobody independently checks.

So: can a model predict a stress level from a short survey better than the student can?

• Whether a machine-learning model can predict stress level from psychological or non-psychological survey data.
• How combining psychological and non-psychological features changes prediction quality.
• How closely a student’s own assessment of their stress matches the USS-derived category.

The study

110 participants completed an online survey between 1 and 10 November 2025, covering demographics, lifestyle and all 21 items of the University Stress Scale. Responses were encoded, standardised and imputed, then built into three feature configurations — psychological only, non-psychological only, and both together — so the contribution of each kind of feature could be seen rather than assumed.

Four supervised models were compared across those configurations with Scikit-Learn: support vector machine, random forest, logistic regression and XGBoost. Results are reported on accuracy, Matthews correlation and weighted F1.

Two numbers that disagree

This is the part worth being careful about. The USS-only configuration reaches the higher accuracy at 0.739 — and it is also the worse-balanced model. The combined configuration scores lower on accuracy and substantially higher on Matthews correlation, which does not reward a model for guessing the most common class. The combined model is the paper’s own best result.

Quoting 0.739 on its own would be true and misleading, which is why both are here. With 110 responses across four models and three configurations, everything above should be read as a direction rather than a measurement.

What it actually showed

Psychological features carried most of the predictive signal, and XGBoost outperformed every other model tested. More striking was the mismatch between what students said about their stress and what the scale reported — 32 of 110 disagreed, which suggests self-assessment alone is a weak instrument.

The instrument was the subject all along — 32 of 110 students disagreed with what the scale reported about them.

Lead Author

• Framed the research question and designed the study as lead author.
• Ran survey collection across 110 participants between 1 and 10 November 2025.
• Handled data curation, encoding, standardisation and imputation.
• Built three feature configurations and compared four supervised models with Scikit-Learn.
• Wrote the paper and presented it at the AIS2C2 2025 international conference.

• Research design
• Survey instrumentation
• Data curation
• Feature engineering
• Supervised classification
• Scikit-Learn

Stress affects students in ways that reach beyond how they feel, into academic performance and the management of daily responsibilities. Measuring it is difficult, because most instruments rely on self-reporting, which often lacks reliability. This study asks whether machine learning can improve stress-level prediction from a short online survey covering demographics, lifestyle and all 21 items of the University Stress Scale.

The three questions that matter