Depression detection with machine learning in a low-resource language: A case study in Hebrew digital health communities

Depression is a debilitating global health crisis; while digital social platforms offer an avenue for early detection through user-generated text, leveraging these digital traces is essential for timely clinical intervention. However, current computational linguistics for mental health are overwhelmingly English-centric. Standard natural language processing (NLP) methodologies may fail when applied to morphologically-rich low-resource languages like Hebrew, where non-concatenative structures and dense root-patterns obscure semantic markers of psychological distress. To bridge this gap, we propose a tailored machine learning framework explicitly designed for morphologically rich languages. We rigorously evaluated 16 classical and ensemble models across more than 130,000 configurations, isolating the optimal interactions between character-level n-grams and specialized Hebrew emotion lexicons to maximize feature extraction. This methodology is empirically tested on a curated dataset from "Camoni," Israel’s largest dedicated digital health community, which provides a high-value, clinically adjacent environment free from the noise of generalist social media platforms. Our optimized Stacking ensemble framework achieved an 84.1% accuracy rate, significantly outperforming standard baselines as well as fine-tuned, state-of-the-art Transformer models like AlephBERT. The integration of 3-to-5 character n-grams proved vital in navigating Hebrew’s morphological sparsity, proving that computationally efficient, expert-engineered features can surpass deep learning in low-resource clinical domains. This study provides a generalizable validated blueprint for deploying mental health screening tools. We urge the clinical and computational integration of these NLP frameworks into digital health ecosystems to facilitate early, life-saving interventions across diverse and underrepresented linguistic landscapes.