Abstract: To achieve the early perception and modeling of abnormal student behavior, a method of student behavioral temporal modeling sensitive to abnormal behavior for mental health prediction (SBTM-SABMHP) was proposed. A heterogeneous information network was constructed to model students' current behavioral patterns using multiple behavioral sensing data collected by mobile devices such as accelerometers, sound sensors, and wireless fidelity (WI-FI). Furthermore, an attention mechanism-based abnormal behavior-sensitive gating module was built for historical behavioral temporal data to effectively integrate long- and short-term behaviors of students. In this way, students' temporal behaviors were modeled, and mental health status prediction was achieved. The comparative analysis experiments of the proposed model were conducted on the public dataset StudentLife. Results show that this method achieves the best performance on all evaluation metrics compared to all other baseline methods for student mental status prediction, demonstrating the effectiveness for student mental health prediction tasks.