# Depression Detection from Three-Channel Resting-State EEG Using a Hybrid Conv1D and Spectral–Statistical Fusion Model

## TL;DR

A new model using three-channel EEG data can detect depression with high accuracy, suggesting potential for portable and scalable screening tools.

## Contribution

A compact hybrid model combining Conv1D and spectral-statistical features for depression detection using only three EEG channels.

## Key findings

- The model achieves 93.43% window-level accuracy in distinguishing depression from healthy controls.
- It performs preliminary external validation on an independent portable EEG dataset without fine-tuning.
- The model is lightweight (40.19 MB) and compatible with hardware-efficient quantization for on-device use.

## Abstract

What are the main findings?
We propose a compact hybrid deep fusion model that combines a Conv1D representation of raw ≈15 s resting-state windows with per-channel spectral–statistical descriptors, trained with subject-independent (cross-subject) splits and class weighting to discriminate major depressive disorder from healthy controls using only three frontal EEG channels. The resulting model attains 93.43% window-level accuracy on the held-out MDD test subjects.The model is trained and selected on a single publicly available MDD dataset and then applied without additional adaptation to an independent three-channel UNICORN acquisition, providing a preliminary external feasibility check under a matched montage (Fp1–Fz–Fp2) and sampling rate (250 Hz).

We propose a compact hybrid deep fusion model that combines a Conv1D representation of raw ≈15 s resting-state windows with per-channel spectral–statistical descriptors, trained with subject-independent (cross-subject) splits and class weighting to discriminate major depressive disorder from healthy controls using only three frontal EEG channels. The resulting model attains 93.43% window-level accuracy on the held-out MDD test subjects.

The model is trained and selected on a single publicly available MDD dataset and then applied without additional adaptation to an independent three-channel UNICORN acquisition, providing a preliminary external feasibility check under a matched montage (Fp1–Fz–Fp2) and sampling rate (250 Hz).

What are the implications of the main findings?
The findings suggest that low-burden, three-channel EEG, when paired with a lightweight hybrid architecture, may inform the design of scalable screening workflows and portable implementations; however, evidence beyond the training dataset remains preliminary and requires larger multi-site validation.The work provides a reproducible processing and modeling pipeline, from preprocessing and feature extraction to training and subject-level aggregation, designed to facilitate cross-dataset comparability and reduce dataset-specific preprocessing sensitivity, supporting future external validation across sites and transparent benchmarking.

The findings suggest that low-burden, three-channel EEG, when paired with a lightweight hybrid architecture, may inform the design of scalable screening workflows and portable implementations; however, evidence beyond the training dataset remains preliminary and requires larger multi-site validation.

The work provides a reproducible processing and modeling pipeline, from preprocessing and feature extraction to training and subject-level aggregation, designed to facilitate cross-dataset comparability and reduce dataset-specific preprocessing sensitivity, supporting future external validation across sites and transparent benchmarking.

Major depressive disorder requires scalable, low-burden screening tools. We examined whether three-channel resting-state EEG can support reliable discrimination between major depressive disorder and healthy controls using a lightweight model compatible with portable implementations. This work makes three main contributions: (i) a compact hybrid fusion model combining raw-window Conv1D embeddings with per-channel spectral–statistical descriptors for three-channel resting-state EEG, (ii) a leakage-resistant subject-independent (cross-subject) evaluation protocol with subject-level inference via majority voting, and (iii) a preliminary external feasibility test on an independent portable three-channel cohort without fine-tuning. The proposed model fuses a Conv1D encoding of raw ≈15 s eyes-closed windows (3840 samples; 15.36 s at 250 Hz) with per-channel spectral and statistical descriptors. Training uses subject-independent splits to avoid leakage, class weighting, and data augmentation (including MixUp); hyperparameters are selected via randomized search with refinement. The model is trained on a publicly available MDD dataset and subsequently applied, without fine-tuning, on an independent acquisition of 20 subjects recorded with a portable three-channel device; we report both window-level and subject-level (majority-vote) performance. On the held-out test subjects from the public dataset, the hybrid model attains 93.43% window-level accuracy. The independent evaluation is reported as a preliminary external feasibility analysis; given the small cohort, we report subject-level performance with 95% confidence intervals to reflect uncertainty and avoid over-interpreting cross-device generalization. The model occupies approximately 40.19 MB on disk, and the architecture is compatible with post-training int8 (TFLite) quantization for resource-constrained hardware. These results, obtained on limited samples, support the feasibility of three-channel EEG for major depressive disorder detection using a lightweight hybrid architecture and motivate prospective clinical validation, on-device inference and quantization studies, and broader evaluation across centers and devices.