Random Forest Based Prediction of Student Stress Levels from Digital Activity Data

Abstract

Early-semester students experience academic adaptation while relying intensively on digital devices, creating a need for accessible, non clinical stress screening. This study aimed to determine whether six self-reported digital-activity variables could distinguish low, moderate, and high DASS-21 derived stress categories and support a student facing screening prototype. A quantitative cross sectional survey and software prototyping design involved 66 Informatics Engineering students. The analytical procedure combined median imputation, standardization, SMOTE within each training fold, Random Forest classification, repeated stratified five fold cross validation with ten repeats, comparison algorithms, class-specific metrics, permutation importance, learning-curve analysis, and functional testing. Metric auditing produced a single out of fold accuracy of 62.12% and macro F1-score of 53.88%. Under repeated validation, the SMOTE Random Forest pipeline achieved 61.97 ± 10.92% accuracy and 55.82 ± 9.87% macro F1-score. The majority DummyClassifier obtained higher accuracy at 65.16 ± 3.47% but only 26.29 ± 0.84% macro F1-score and zero recall for moderate and high stress. SMOTE Random Forest recall was 20.67 ± 23.94% for moderate stress and 80.00 ± 30.00% for high stress. Logistic Regression produced the highest comparison-model accuracy of 70.89%, although its macro F1-score of 55.23% was slightly below that of SMOTE Random Forest. Total screen time was the only predictor with clearly positive permutation importance. Overlapping feature distributions, unstable minority class estimates, and a persistent training validation gap limited performance. StresCheck therefore constitutes an exploratory proof of concept and requires larger, balanced, externally validated data before screening deployment.