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

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.

💡 Research Summary

This paper introduces QuanvNeXt, an end‑to‑end quantum‑enhanced neural network for diagnosing major depressive disorder (MDD) from raw electroencephalogram (EEG) recordings. The core innovation lies in replacing conventional 1‑D convolutional filters with Quanv1D layers, which embed each temporal patch into a quantum state using amplitude encoding, process it through parameterized unitary gates, and decode the result via Pauli‑Z measurements. To address the overly smooth attention distribution caused by the original softmax normalization, the authors incorporate temperature scaling (α/t), granting fine‑grained control over the temporal focus of each quantum filter.

Building on Quanv1D, the architecture features a novel Cross Residual block that fuses three design principles: (i) residual skip connections (ResNet) for stable gradient flow and preservation of raw signal information, (ii) dense feature aggregation (DenseNet) to promote multi‑scale feature reuse without inflating parameter count, and (iii) channel shuffling (ShuffleNet) in two stages (four‑group then eight‑group) to mitigate channel‑specific bias and encourage cross‑channel interactions. Layer normalization replaces batch normalization to ensure stability under small batch sizes and variable sequence lengths.

The authors evaluate QuanvNeXt on two publicly available EEG datasets. Dataset 1 comprises 19‑channel, 256 Hz recordings from 30 healthy controls and 34 MDD patients; Dataset 2 contains 128‑channel, 250 Hz data from 25 MDD patients and 29 controls. Both datasets undergo rigorous preprocessing: band‑pass filtering, dual notch filtering (to suppress 50 Hz power‑line interference), Artifact Subspace Reconstruction, ICA‑based ocular and muscular artifact removal, and manual inspection. After cleaning, only 10 healthy and 19 depressed recordings remain for Dataset 1, and 10 healthy and 10 depressed recordings for Dataset 2. Segments are extracted using 8‑second windows with 90 % overlap, yielding tensors of shape (64, 19, ≈2000) and (64, 128, ≈2000) respectively. Subject‑wise 70/30 train‑test splits are performed before windowing, followed by random undersampling to balance classes and subject‑independent Z‑normalization based on training statistics.

Training employs the Adam optimizer with cosine‑annealing learning‑rate scheduling and L2 regularization. Across five‑fold cross‑validation, QuanvNeXt achieves an average classification accuracy of 93.1 % and an average AUC‑ROC of 97.2 %, surpassing strong baselines such as InceptionTime (91.7 % accuracy, 95.9 % AUC‑ROC). To assess robustness, Gaussian noise with amplitudes ε = 0.01, 0.05, and 0.1 is added to test inputs. Expected Calibration Error (ECE) remains low—0.0436 for Dataset 1 and 0.1159 for Dataset 2 at the highest perturbation—indicating well‑calibrated predictions even under substantial noise.

Explainable AI analyses are conducted using two complementary post‑hoc techniques. Hierarchical feature mapping visualizes the contribution of each Quanv1D and Cross Residual layer to the input, revealing that early quantum filters attend strongly to alpha and beta band power differences, while deeper layers capture longer‑range temporal dependencies. Latent manifold projection (t‑SNE/UMAP) of the final feature vectors shows clear separation between healthy and depressed subjects, confirming that the model learns discriminative spectro‑temporal patterns rather than spurious artifacts.

The paper’s contributions are threefold: (1) it pioneers the application of quantum machine‑learning to mental‑health EEG analysis, (2) it proposes a Cross Residual block tailored for time‑series data that enhances feature diversity without increasing parameter count, and (3) it delivers state‑of‑the‑art performance on two depression datasets while providing thorough uncertainty quantification and interpretability.

Limitations include reliance on quantum circuit simulators rather than real quantum hardware, modest dataset sizes that may limit generalizability, and a lack of systematic sensitivity analysis for the temperature scaling and circuit depth hyper‑parameters. Future work should benchmark QuanvNeXt on actual quantum processors, expand evaluation to larger multi‑site cohorts, explore automated hyper‑parameter optimization, and investigate multimodal extensions (e.g., EEG + fMRI) to further improve diagnostic accuracy.