<p>Today, one of the primary causes of death worldwide is suicide, often linked to underlying mental health disorders such as depression. Detecting and addressing individuals with depression is crucial for providing appropriate counseling and mental health support services. Most individuals with depression tend to avoid seeking professional help, making social media platforms a valuable resource for identifying these individuals. Therefore, the best way to identify these people in the current era of social media is that a person has a high potential to express their opinions, thoughts, and opinions on various topics in the form of words in these environments. In recent years, advancements in artificial intelligence, particularly deep learning and natural language processing, have enabled us to extract precise and latent patterns from textual data. This study aims to offer a binary depression classification model that detects tweets indicative of depression signs on Twitter using a deep learning model. Due to the unavailability of a Persian depression dataset, Persian tweets were retrieved from Twitter using the Twitter API and annotated based on self-declaration by humans. The annotated texts underwent text preprocessing and were embedded using ParsBERT to gain contextual representation of words. These word vectors were then fed into a CNN-BiLSTM deep neural network to capture latent local and sequential patterns. The model trained on this dataset achieved a validation accuracy of 77% and an F1 score of 76% in depression detection, surpassing baseline machine learning and deep learning models, including LSTM and FNN. Moreover, the model’s applicability was validated by successfully labeling another dataset containing tweets about the Iran universities entrance exam, demonstrating its effectiveness and generalizability in practical tasks.</p>