Fine-Grained Symptom-Aware Arabic Mental Health Disorder Classification Using MentalMARBERT

Abstract Mental health NLP research in Arabic remains underexplored and has largely focused on binary detection, overlooking symptom-level distinctions among clinically overlapping disorders. This study proposes a symptom-aware two-stage framework based on MentalMARBERT, a domain-adapted Arabic language model, for classifying depression, anxiety, bipolar disorder, post-traumatic stress disorder, and obsessive-compulsive disorder. Stage 1 distinguishes None from Disorder posts, whereas Stage 2 classifies disorder-related posts and jointly predicts multi-label symptoms. The resulting symptom representation is integrated with the textual representation through gated fusion, while a masking strategy, informed by the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR), emphasizes disorder-relevant symptoms and suppresses less informative signals. The framework was evaluated using stratified five-fold cross-validation on 50,670 Arabic posts. The proposed two-stage framework achieved an Accuracy of 0.9723 and a Macro-F1 of 0.9680 in Stage 1, while its DSM-masked symptom-aware Stage 2 achieved a Macro-F1 of 0.8646, outperforming the text-only baseline (0.8498) and the unmasked symptom-aware model (0.8447). DSM-5-TR informed masking also improved symptom extraction, increasing the Micro-F1 from 0.6258 to 0.8127 and the Jaccard score from 0.5112 to 0.7365. Limiting the symptom space to the 20 most frequent labels further increased the symptom-level Macro-F1 to 0.7699, indicating that reducing rare labels produces a cleaner predictive signal. These findings demonstrate the value of symptom-level modeling for fine-grained Arabic mental health classification and symptom extraction.