QuanvNeXt: An end-to-end quanvolutional neural network for EEG-based detection of major depressive disorder

TL;DR

QuanvNeXt tackles EEG-based major depressive disorder diagnosis using an end-to-end quanvolutional architecture with a novel Cross Residual Block. It achieves state-of-the-art accuracy and MCC on two open EEG datasets while maintaining a lightweight parameter footprint, and it provides well-calibrated predictions alongside post hoc explainability analyses. The work demonstrates the practical potential of quantum-inspired approaches for neuropsychiatric disorder detection and highlights avenues for hardware realization and reliability improvements.

Abstract

This study presents QuanvNeXt, an end-to-end fully quanvolutional model for EEG-based depression diagnosis. QuanvNeXt incorporates a novel Cross Residual block, which reduces feature homogeneity and strengthens cross-feature relationships while retaining parameter efficiency. We evaluated QuanvNeXt on two open-source datasets, where it achieved an average accuracy of 93.1% and an average AUC-ROC of 97.2%, outperforming state-of-the-art baselines such as InceptionTime (91.7% accuracy, 95.9% AUC-ROC). An uncertainty analysis across Gaussian noise levels demonstrated well-calibrated predictions, with ECE scores remaining low (0.0436, Dataset 1) to moderate (0.1159, Dataset 2) even at the highest perturbation (ε = 0.1). Additionally, a post-hoc explainable AI analysis confirmed that QuanvNeXt effectively identifies and learns spectrotemporal patterns that distinguish between healthy controls and major depressive disorder. Overall, QuanvNeXt establishes an efficient and reliable approach for EEG-based depression diagnosis.

Figures (5)

• Figure 1: Workflow of the Quanv1D layer, which can handle arbitrary input and output sizes. After extracting patches, multiple trainable quantum circuits generate feature maps.
• Figure 2: Cross Residual block. The block, proposed for time series, combines residual skip connections, dense feature aggregation, and channel shuffling for efficient representation learning. Note: BN: batch normalization; LN: layer normalization.
• Figure 3: Hierarchical feature maps of HC and MDD subjects in Dataset 1, showing time-domain representations and corresponding spectrograms of the embedded activations. (a) Features after the Embedding layer; (b) Features in the middle of the network, i.e., after the 2nd Cross Residual block; (c) Features towards the end of the network, i.e., after the 4th and final Cross Residual block. Note: time and frequency axes correspond to latent representations derived from the embedding process, not the raw EEG domain.
• Figure 4: Hierarchical feature maps of HC and MDD subjects in Dataset 2, showing time-domain representations and corresponding spectrograms of the embedded activations. (a) Features after the Embedding layer; (b) Features in the middle of the network, i.e., after the 2nd Cross Residual block; (c) Features towards the end of the network, i.e., after the 4th and final Cross Residual block. Note: time and frequency axes correspond to latent representations derived from the embedding process, not the raw EEG domain.
• Figure 5: Comparison of raw EEG features and model-learned latent representations using UMAP. Points represent individual subjects, colored by group (HC vs MDD). The left column shows raw input data, and the right column shows the features after model projection. (a) Dataset 1; (b) Dataset 2.