Depression and suicidal thoughts are significant global health concerns typically diagnosed through clinical assessments, which can be constrained by issues of accessibility and stigma. However, current methods often face challenges with this variability and struggle to integrate different models effectively and generalize across different settings, leading to reduced effectiveness when applied to new contexts, resulting in less accurate outcomes. This research presents a novel approach to suicide and depression detection from social media (SADDSM) by addressing the challenges of variability and model generalization. The process involves four key stages: first, preprocessing the input data through stop word removal, tokenization, and stemming to improve text clarity; then, extracting relevant features such as TF-IDF, style features, and enhanced word2vec features to capture semantic relationships and emotional cues. A modified mutual information score is used for feature fusion, selecting the most informative features. Subsequently, deep learning models like RNN, DBN, and improved LSTM are stacked to form an ensemble model that boosts accuracy while reducing overfitting. The performance is further optimized using the Dwarf Updated Pelican optimization algorithm (DU-POA) to fine-tune model weights, achieving an impressive 0.962 accuracy at 90% training data, outperforming existing techniques.