As society develops, mental health is becoming increasingly important. Traditional mental health assessment methods, such as face-to-face interviews and observational techniques, are limited by several factors, including subjectivity, response delays, and high costs. To address these issues, recent research has utilized smart sensor data. Using smartphones to collect data provides valuable insights into student behavior while ensuring the privacy of the individuals involved. The transition from traditional questionnaires to sensor data has prompted the adoption of a multi-task learning (MTL) framework for a more comprehensive classification of mental health. This study investigates the relationship between behavioral attributes and mental health dimensions, using the Apriori algorithm for association rule mining and feature extraction. This method uncovers significant correlations, highlighting the close connection between mental health and behavioral data through detailed analysis. The study proposes the use of multi-task modeling to optimize information utilization and identify correlations across tasks. The introduction of the Deep Cross Network with Squeeze and Excitation Network and Progressive Layered Extraction (DC-SE-PLE), a deep learning neural network model, enhances hierarchical feature extraction and cross-task learning, significantly improving the model’s generalization and prediction accuracy.