This study develops and evaluates a hybrid deep learning model for early detection of depression in Indonesian social media text. The approach combines convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) layers, enhanced by Word2Vec feature expansion to capture nuanced semantic context, and employs particle swarm optimization (PSO) to fine-tune hyperparameters such as learning rate, convolution filter sizes and LSTM unit counts. A similarity based corpus is created by integrating 58,115 Indonesian tweets annotated for depressive content with 111,458 Indonesian news articles, and all texts undergo tokenization, stopword removal and normalization. Both a baseline CNN BiLSTM model and the PSO optimized variant with Word2Vec feature expansion are trained with early stopping on validation loss and evaluated on a held-out test set using accuracy, precision, recall and F1 score. Experimental results demonstrate that the optimized CNN BiL STM model with PSO and Word2Vec feature expansion achieves an accuracy of 86.38% at Top 15 ranking. In comparison, the baseline CNN BiLSTMmodel without PSO andfeature expansion attains an accuracy of 78.96%. This represents a significant improvement of approximately 9.39% in overall accuracy and delivers substantial gains in precision and recall. These findings highlight the potential of optimized deep learning models with semantic feature enrichment for supporting early depression detection through text-based analysis in Indonesian social media contexts.