Tabular Prior-Data Fitted Networks for Objective Depression Screening with Portable EEG and ECG (Preprint)

BACKGROUND Major depressive disorder (MDD) is a a prevalent and debilitating psychiatric condition; however, early detection remains challenging due to reliance on subjective assessments. There is a critical need for scalable, objective tools to facilitate timely screening and intervention. OBJECTIVE To evaluate a portable multimodal system that combines single-channel Electroencephalogram (EEG) and Electrocardiogram (ECG) for rapid assessment of depression in real-world settings, including its utility for depression screening and severity prediction. METHODS We collected data recorded with a portable EEG–ECG system under four standardized conditions: eyes‑open rest, eyes‑closed rest, negative‑emotion image viewing, and a structured interview. To identify the MDD biomarkers, we derived standard heart rate variability (HRV) indices from ECG and computed power spectral density (PSD) features using EEG. Besides, brain–heart coupling was quantified by calculating the correlation between EEG and ECG. A state‑of‑the‑art tabular prior-data fitted network machine‑learning model (TabPFN) was trained with leave‑one‑out cross‑validation to (1) classify MDD versus HC using three input sets (EEG‑only, ECG‑only, and fused EEG and ECG) and (2) predict depression severity indexed by Hamilton Depression Rating Scale (HAMD‑24) scores. Furthermore, permutation feature importance was used to interpret the contribution of PSD and HRV features. RESULTS Fifty-six participants including 26 clinical diagnosed MDD and 30 healthy controls (HC) were included in this study. Relative to HC group, the MDD group showed consistent HRV alterations and disrupted brain–heart coupling, most pronounced during negative-emotion tasks, suggesting impaired central–autonomic regulation. ECG-only models demonstrated strong discrimination, while fused EEG and ECG achieved the highest cross-validated accuracy of 80.4% (area under the curve, AUC: 0.712). For severity estimation, multimodal physiological features predicted HAMD-24 (r = 0.659; P < .001). Permutation importance consistently prioritized signals from the Interview conditions and HRV reactivity between rest and task conditions. CONCLUSIONS These findings establish brain–heart biomarkers as a practical and objective basis for depression screening and severity prediction. The proposed portable system offers a promising step toward scalable, precise mental health assessment with potential applications in primary care and community settings.