Suicidal ideation posted on social media platform has become a critical concern in mental health research, calling for the implementation of effective detection systems. However, the detection of suicidal ideation content is challenged by informal vocabulary and jargon used in tweets. This paper proposes a hybrid deep model that combines Convolutional Neural Networks (CNN) to extract local, spatial text patterns and Gated Recurrent Units (GRU) to understand sequential context in order to counter these problems. The model employs a multi-source feature approach, with TF-IDF for statistical feature weighting, pre-trained GloVe vectors trained on a Wikipedia and Gigaword 5 token corpus of 6 billion tokens (Wikipedia 2014 + Gigaword 5) for general semantic enhancement, and FastText embeddings trained on the domain-specific Kaggle SuicideWatch corpus, with 232,074 posts, for advanced feature enhancement. We collected and labeled 35,917 tweets into suicidal and non-suicidal classes. The data was thoroughly preprocessed by normalization, stopword elimination, and lemmatization to get it ready for analysis. The hybrid CNN-GRU model was further trained and optimized by hyperparameter tuning. In comparison with the baseline CNN-GRU model (91.28% accuracy), our highest-performing setting posted 93.87% accuracy. This is a remarkable boost of 2.59 percentage points, confirming the high efficacy of the proposed research. The result confirm that integrating statistical features with GloVe rich feature enhancement and FastText semantic embedding provides a highly effective method especially on noisy social media inputs.