DISCERN: Disentangled Multilingual Mental Health Risk Detection With Calibrated, High-Recall Predictions

Abstract

Identifying mental health risk on social media is high-stakes, demanding low false negative rates (FNRs) and calibrated, trustworthy predictions. Prior work typically fuses linguistic and semantic signals without psycholinguistic grounding or disentanglement, lacks integrated calibration, prioritizes standard metrics over FNR, and remains largely English-only. We propose a disentangled multilingual mental health risk detection with calibrated, high-recall predictions (DISCERN), a safety-oriented framework that disentangles psycholinguistically grounded linguistic signals from semantic content via dual encoders, dynamically fused through hierarchical gated attention (HGA). Evaluated on seven social media datasets spanning suicidal ideation, stress, and depression in English, Hindi, and Hinglish, DISCERN incorporates uncertainty quantification and calibration during training, reports expected calibration error (ECE), and applies posthoc calibration. It achieves lower FNR than all supervised baselines across all datasets, including a ${\boldsymbol{\sim}}$9% relative reduction over the strongest baseline on the Hinglish suicide-risk set, while producing statistically well-calibrated predictions, thereby supporting ethical, human–artificial intelligence (AI) cooperative triage for multilingual mental health monitoring.