A comparative analysis of machine learning, deep learning, and ensemble learning models for suicidal ideation detection on social media

Suicidal ideation is the consideration of self-inflicted death and has emerged as a serious and rising concern for worldwide public health systems, contributing to over 800,000 deaths annually worldwide. Individuals express intentions, distress, and thoughts of self-harm and death, thereby offering a crucial and unprecedented avenue for timely psychological intervention and surveillance. The core problem addressed by this research is the pronounced global insufficiency in detecting and mitigating suicidal ideation across linguistically diverse environments, specifically focusing on low-resource languages. Previous computational studies have primarily focused on English and other widely used languages, leaving a critical research gap for contexts such as Amharic. The main aim of this study was to investigate and to develop a robust and optimal deep learning model for identifying suicidal ideation in Amharic social media content. A total of 5,428 postings from social media such as Facebook and Telegram were carefully collected, manually annotated, and processed to provide the first publicly available dataset on Amharic suicidal ideation. After preprocessing, Word2Vec, GloVe, and FastText were among the word embedding techniques used for feature engineering. Five deep learning models and four machine learning models were then compared in this study. Based on the experimental results, ensemble learning models achieved slightly higher performance than the individual deep learning models in suicidal ideation detection. Specifically, hard voting ensemble learning models achieved the highest accuracy (83.3%) and precision (86.55%). CNN-BiLSTM achieved an accuracy of 81.45%, while the BiLSTM model achieved an accuracy of 81.4%, indicating that bidirectional sequence learning effectively captures contextual information in Amharic texts. Beyond its strong empirical performance, this study contributes a robust detection framework and introduces a public Amharic dataset, thereby establishing a foundational methodological resource for mental health research. Future studies should focus on developing scalable datasets and context-aware deep learning models that effectively handle multimodal data, multilingual transformer models, fusion-based features, and addressing bias in low-resource computational linguistics.