The COVID-19 pandemic has significantly increased depression cases globally, necessitating objective and stigma-free mental health assessments, particularly in educational sectors. Traditional manual screenings often induce discomfort, limiting individuals’ willingness to disclose their mental state. This study introduces a novel non-invasive, multimodal physiological signal-based automatic depression detection system. The framework utilizes electrocardiogram (ECG), photoplethysmography (PPG), electrodermal activity (EDA), respiration rate (Resp), and temperature (Temp) signals recorded from 44 university students, staff, and faculty members. Physiological signals were collected during Stroop and Emotional Stroop test tasks, administered before and after exposure to a video stimulus, with depression severity quantified using Patient Health Questionnaire-9 (PHQ-9) scores. The dataset was pre-processed and balanced using the SMOTE-Edited Nearest Neighbor (SMOTE-ENN) technique to mitigate overfitting and enhance classification robustness. A Bidirectional Long Short-Term Memory (Bi-LSTM) deep learning model was employed to capture sequential dependencies within physiological signals, enabling a four-class depression classification system. The proposed framework achieved an accuracy of 97.73%, outperforming existing physiological signal-based depression detection approaches. These findings highlight the potential of the proposed system as an effective early intervention tool for scalable and automated depression detection, particularly in remote healthcare applications and crisis scenarios such as the COVID-19 pandemic.