A Multi-Level CNN Framework for Accurate and Interpretable Assessment of College Students’ Mental Health

Accurate assessment of college students’ mental health is difficult, as traditional questionnaire-based methods are often subjective and limited in predictive capacity. This study proposes a multi-level convolutional neural network (CNN) that integrates multimodal data sources, including self-assessment scales, behavioral indicators, and academic performance records. Data preprocessing involved normalization, categorical encoding, and fusion into structured matrices suitable for CNN input. To enhance interpretability, visualization tools such as heatmaps and accuracy curves were employed. The proposed model achieved an accuracy of 95.6%, outperforming conventional classifiers including SVM, RF, and GBDT. Beyond overall accuracy, the CNN demonstrated superior recall and F1 scores, attributable to its ability to capture cross-domain correlations such as the combined effects of stress levels, sleep patterns, and social interaction on psychological outcomes. These results highlight both methodological advantages and practical relevance. The framework not only improves prediction accuracy and transparency but also provides a foundation for early intervention and future multi-institutional, privacy-preserving mental health monitoring.