Suicidal ideation is a serious global concern linked to mental health issues such as depression, hopelessness, and sadness. Social media platforms like X offer valuable data, as individuals often express their emotional states online, presenting an opportunity for early detection. However, previous studies have typically relied on only one or two feature extraction methods, limiting model performance. This study proposes a hybrid deep learning approach to detect suicidal ideation using an English dataset of 27,637 posts, with 12,020 labeled as suicidal and 15,617 as nonsuicidal. Feature extraction combines TF-IDF for word importance, BERT for contextual understanding, NRC Emotion Lexicon for emotional cues, and FastText for subword feature expansion. The research is carried out in four scenarios: (1) determining TF-IDF feature size, with 2,500 features achieving 67.89% accuracy using CNN; (2) finding best data split ratios at 90:10, with BERT + CNN-BiLSTM achieving 95.44% accuracy; (3) combining features, where BERT + NRC + CNN-BiLSTM reached 95.68%; and (4) applying FastText expansion, with BERT + NRC + FastText + CNN-BiLSTM achieving the highest accuracy of 95.91%. This research provides a comprehensive framework for suicidal ideation detection and supports the development of early warning systems for suicide prevention on social media.