Suicide poses a serious mental health concern, especially as many individuals increasingly turn to social media platform to express their emotional distress. In this study, a deep learning method for identifying suicidal ideation in English language posts from the social media platform X is presented. The dataset consists of 35,917 posts. Feature ex traction uses TF-IDF, while semantic feature expansion uses multiple pre-trained GloVe embeddings, including glove.6B.50d, glove.6B.100d, glove.6B.200d (trained on Wikipedia 2014 and Gigaword 5 corpora), and twitter.27B.200d (trained on 2 billion tweets) Implementing a hybrid CNN-BiLSTM model, combining it with TF-IDF for feature extraction and GloVe-based word embeddings for semantic feature expansion. Four experimental scenarios are conducted in order to optimize model perfor mance by adjusting data split ratios, maximum feature, n gram configurations, and embedding types. The proposed model achieves its highest accuracy of 85.07% using GloVe Twit ter.27B.200d embeddings, representing a 12.74% improvement over the baseline. Results indicate that integrating local feature detection (via CNN), contextual sequence learning (via BiLSTM), and semantic representation (via GloVe) significantly enhances detection performance. This study demonstrates the potential of how hybrid deep learning models can be used to detect suicide ideation in short-form social media posts.