¹The First Hospital of Hebei Medical University, Mental Health Center, Shijiazhuang, China

Background: Atypical depression (AD), a prevalent but often misdiagnosed subtype of major depressive disorder (MDD), is urgently needed the objective biomarkers.

Aims & Objectives: This study aimed to investigate the potential of speech analysis using the Context-Aware Masking Plus (CAM++) deep-learning model: first, to distinguish MDD patients from healthy controls (HCs), and second, within the depressed group, to objectively differentiate atypical depression (AD) from non-AD depression.

Method: A cross-sectional study enrolled 80 MDD patients (19 AD, 61 non-AD) meeting Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) criteria and 69 HCs. Standardized speech samples (4-min per participant), including interactive, reading, and descriptive tasks, were recorded in a controlled environment. We propose a novel end-to-end deep learning architecture (CAM++) that integrates multi-scale temporal context modeling to identify depression-relevant speech features directly from raw audio. The symptom severity was assessed using HAMD-17, HAMA, and MADRS. Performance metrics (sensitivity, specificity, PPV, NPV) were calculated.

Results: Clinically, the AD group had significantly higher BMI, HAMA, and MADRS scores compared to the non-AD group (all p<0.05). The model achieved high diagnostic performance, with over 90% sensitivity and 100% specificity for MDD classification, and similarly robust accuracy for identifying AD subtypes.

Discussion & Conclusions: Speech analysis using the CAM++ AI model effectively distinguishes MDD patients from healthy individuals and shows significant promise in differentiating the AD subtype from other forms of depression. These findings highlight the potential utility of speech-based deep learning models in identifying objective markers for atypical depression, warranting further validation in larger cohorts.