Normative Deviation Modeling for EEG-based Depression Detection

Wan Jiang; Jingqi Cheng; Hui Zhang; Xu Zhao; Xiaojing Chen; Shengeng Tang; Jingjing Wu; Richang Hong

cs.LGOct 27, 2026

Normative Deviation Modeling for EEG-based Depression Detection

Authors: Wan Jiang, Jingqi Cheng, Hui Zhang, Xu Zhao, Xiaojing Chen, Shengeng Tang, Jingjing Wu, Richang Hong

Organizations: Hefei University of Technology · Anhui University

Abstract

Web-enabled digital health services are expanding access to physiological monitoring and mental health screening. However, electroencephalography (EEG)-based major depressive disorder (MDD) detection remains challenging across heterogeneous data sources, where individual variability and acquisition differences may affect disease-related representations. In this work, we investigate the use of healthy normative references for EEG-based MDD detection. Specifically, structural and distributional reference models are independently trained using healthy control (HC) recordings to characterize departures from healthy EEG patterns. We develop two diagnostic approaches that incorporate normative deviations at different representation levels. Scalar-level Normative Deviation-guided Diagnosis (S-NDD) summarizes structural and distributional deviations into scalar scores and integrates them with deep EEG representations. Feature-level Normative Deviation-guided Diagnosis (F-NDD) retains dimension-wise residuals and combines them with deep EEG representations through attention-based aggregation and auxiliary supervision. We also employ target HC-assisted score calibration for cross-source evaluation without updating the trained models. Experiments on two public EEG datasets examine both approaches under subject-independent within-source and bidirectional cross-source settings. The results demonstrate the effectiveness of normative deviation-guided diagnosis across different evaluation settings.

Figures & tables

Figure 1. Inter-electrode correlation patterns in HC and MDD EEG recordings. (a) and (b) show the mean Pearson correlation matrices for HC and MDD, respectively. (c) and (d) show their signed and absolute differences (MDD $-$ HC), respectively. All panels use the same electrode ordering derived from HC-based hierarchical clustering.

Figure 2. Overview of S-NDD and F-NDD. Both methods establish structural and distributional healthy references in distinct representation spaces. S-NDD quantifies scalar normative deviations for deep EEG feature fusion, whereas F-NDD preserves feature-level residuals for deviation-guided diagnosis. Optional target HC-assisted calibration adjusts cross-dataset predictions.

Discriminative Learning Methods
Methods	Window-level Metrics (%)					Subject-level Metrics (%)
Methods	Acc ( $\uparrow$ )	Pre ( $\uparrow$ )	Rec ( $\uparrow$ )	F1 ( $\uparrow$ )	AUC ( $\uparrow$ )	Acc ( $\uparrow$ )	Pre ( $\uparrow$ )	Rec ( $\uparrow$ )	F1 ( $\uparrow$ )	AUC ( $\uparrow$ )
XGBoost ( Hassan and Kaabouch, 2024 )	$83.25\pm 3.06$	$83.54\pm 7.49$	$82.41\pm 5.15$	$82.49\pm 5.85$	$88.73\pm 3.50$	$89.17\pm 4.66$	$93.71\pm 7.30$	$86.33\pm 5.53$	$89.03\pm 4.51$	$92.19\pm 3.92$
1D-CNN-LSTM ( Ay et al., 2019 )	$90.53\pm 3.25$	$90.17\pm 6.48$	$91.58\pm 7.27$	$90.52\pm 3.72$	$93.13\pm 3.38$	$93.26\pm 5.30$	$93.25\pm 9.92$	$95.83\pm 6.80$	$94.01\pm 5.18$	$95.13\pm 3.43$
1D-CNN-Trans ( Peng et al., 2025 )	$85.56\pm 4.61$	$84.90\pm 6.35$	$87.97\pm 5.76$	$84.98\pm 4.36$	$92.39\pm 5.03$	$86.52\pm 6.18$	$86.02\pm 6.69$	$89.63\pm 5.38$	$86.21\pm 5.24$	$93.96\pm 3.86$
CWT-1D-CNN ( Raghavendra et al., 2023 )	$91.54\pm 5.36$	$91.95\pm 9.66$	$92.30\pm 8.40$	$92.12\pm 5.40$	$92.57\pm 5.41$	$93.18\pm 6.62$	$92.80\pm 10.20$	$95.00\pm 8.74$	$93.31\pm 6.47$	$95.42\pm 5.15$

Table 1. Performance comparison (%) on Mumtaz2016 under subject-independent 10-fold cross-validation. All baselines are evaluated using a unified protocol. Results are mean

\pm

std., with best and second-best in bold and underlined , respectively.

Figure 3. Normative deviation visualization for selected HC and MDD cases on Mumtaz2016 under within-dataset evaluation. (a) S-NDD and (b) F-NDD illustrate structural and distributional deviation patterns. Deviation maps are normalized for visualization, with non-comparable color scales across methods.

Discriminative Learning Methods
Methods	Acc ( $\uparrow$ )	Pre ( $\uparrow$ )	Rec ( $\uparrow$ )	F1 ( $\uparrow$ )	AUC ( $\uparrow$ )
XGBoost ( 2024 )	$58.45\pm 8.76$	$55.12\pm 9.68$	$60.30\pm 7.95$	$57.60\pm 8.33$	$59.18\pm 9.12$
1D-CNN-LSTM ( 2019 )	$57.20\pm 8.42$	$54.35\pm 9.15$	$58.12\pm 7.63$	$56.18\pm 8.84$	$58.14\pm 7.92$
1D-CNN-Trans ( 2025 )	$54.51\pm 2.69$	$55.90\pm 28.45$	$10.35\pm 8.23$	$15.77\pm 11.17$	$54.23\pm 6.87$
CWT-1D-CNN ( 2023 )	$59.80\pm 9.21$	$57.66\pm 8.54$	$57.34\pm 7.89$	$57.50\pm 9.46$	$60.47\pm 8.63$
CWT-2D-CNN ( 2023 )	$56.34\pm 5.40$	$56.52\pm 27.31$	$11.13\pm 12.88$	$16.62\pm 16.86$	$52.39\pm 5.77$

Table 2. Cross-dataset performance (%) on MODMA with Mumtaz2016 as the source. Results are mean

\pm

std., with best and second-best in bold and underlined , respectively.

Discriminative Learning Methods
Methods	Acc ( $\uparrow$ )	Pre ( $\uparrow$ )	Rec ( $\uparrow$ )	F1 ( $\uparrow$ )	AUC ( $\uparrow$ )
XGBoost ( 2024 )	$48.84\pm 8.39$	$76.73\pm 7.17$	$30.95\pm 17.62$	$41.91\pm 20.06$	$57.30\pm 3.24$
1D-CNN-LSTM ( 2019 )	$76.74\pm 8.39$	$90.55\pm 8.74$	$73.81\pm 20.62$	$79.38\pm 11.12$	$83.57\pm 4.92$
1D-CNN-Trans ( 2025 )	$56.59\pm 8.80$	$74.96\pm 9.49$	$55.95\pm 29.95$	$58.97\pm 21.37$	$61.11\pm 3.94$
CWT-1D-CNN ( 2023 )	$75.19\pm 1.34$	$82.48\pm 5.41$	$79.76\pm 11.48$	$80.52\pm 2.93$	$80.08\pm 5.99$
CWT-2D-CNN ( 2023 )	$48.84\pm 4.03$	$62.87\pm 6.21$	$54.76\pm 5.46$	$58.20\pm 1.53$	$46.98\pm 5.17$

Table 3. Cross-dataset performance (%) on Mumtaz2016 with MODMA as the source dataset. All methods use target HC-assisted score calibration. Results are mean

\pm

std., with best and second-best in bold and underlined , respectively.

Figure 4. Normative deviation visualization for selected HC and MDD cases on MODMA under cross-dataset evaluation with Mumtaz2016 as the source dataset. (a) S-NDD and (b) F-NDD illustrate structural and distributional deviation patterns.

Figure 5. Normative deviation visualization for selected HC and MDD cases on Mumtaz2016 under cross-dataset evaluation with MODMA as the source dataset. (a) S-NDD and (b) F-NDD illustrate structural and distributional deviation patterns.

Figure 6. F-NDD scores before and after target HC calibration for (a) HC and (b) MDD. Green lines indicate corrected predictions, crosses denote errors, and the dashed line marks the decision threshold.

S-NDD Ablation Study
Methods	Acc ( $\uparrow$ )	Pre ( $\uparrow$ )	Rec ( $\uparrow$ )	F1 ( $\uparrow$ )	AUC ( $\uparrow$ )
w/o Both	$85.56\pm 4.61$	$84.90\pm 6.35$	$87.97\pm 5.76$	$84.98\pm 4.36$	$92.39\pm 5.03$
w/o Dist.	$90.40\pm 5.94$	$90.69\pm 5.99$	$92.72\pm 5.72$	$90.91\pm 5.05$	$98.24\pm 1.67$
w/o Struct.	$93.11\pm 4.85$	$92.38\pm 7.72$	$94.34\pm 8.09$	$92.95\pm 4.84$	$97.92\pm 2.96$
Dec. Fusion	$93.12\pm 5.05$	$94.10\pm 7.49$	$91.62\pm 7.60$	$92.00\pm 8.41$	$98.53\pm 1.44$
Mahalanobis	$92.62\pm 5.01$	$90.91\pm 7.59$	$95.31\pm 5.25$	$93.06\pm 6.72$	$97.84\pm 2.97$

Table 4. Ablation results (%) of S-NDD and F-NDD on Mumtaz2016. Results are reported as mean

\pm

std., with the best and second-best results in bold and underlined, respectively.

Figure 7. Sensitivity of S-NDD to (a) codebook size $K$ and (b) fusion weight $\alpha$ on Mumtaz2016. Orange dots indicating the highest means.

Figure 8. Sensitivity of F-NDD to (a) codebook size

K

and (b) auxiliary loss weight

\lambda_{\mathrm{dev}}

on Mumtaz2016. Orange dots indicating the highest means.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Configuration
Condition classifier	1D-CNN–Transformer; Adam, lr $=10^{-3}$ , wd $=10^{-4}$ ; dropout $=0.5$ , patience $=8$ .
Diffusion training	1D-CNN–Transformer with FiLM; $T=1000$ , $\beta\in[10^{-4},0.02]$ ; MSE, lr $=10^{-3}$ , 50 epochs, patience $=8$ .
Data synthesis	570 anchors, $k=3$ per anchor; 100-step DDIM, guidance $=1.0$ .
Downstream training	1D-CNN–Transformer; lr $=10^{-4}$ , 100 epochs; $\delta=0.1$ , patience $=15$ .

Appendix

Table 5. Stage-wise training and sampling configurations of Diff-Aug. lr and wd denote the learning rate and weight decay, respectively.

Stage	Configuration
Anchor selection	$s_{i}=0.7\widetilde{D}_{\mathrm{str}}+0.3\widetilde{D}_{\mathrm{dist}}$ ; normalization: 5th–95th percentiles; hard range: 70th–90th percentiles; exclude the top 10%.
Data synthesis	Diff-Aug predictor; 100-step DDIM; classifier guidance $=2.5$ ; five pseudo-segments per anchor.
Downstream training	$H_{i}\in\mathbb{R}^{L\times 128}$ , $W_{\mathrm{att}}\in\mathbb{R}^{128\times 1}$ ; attention: $\operatorname{Softmax}_{L}(s_{i}H_{i}W_{\mathrm{att}})$ ; optimization as in Diff-Aug.

Appendix

Table 6. Key configurations of Hard Neg. Settings not listed here follow those of Diff-Aug.

Stage	Configuration
Data synthesis	Diff-Aug predictor; 100-step DDIM; classifier guidance $=1.0$ .
Artifact cleaning	Fourth-order Butterworth low-pass filter, cutoff $=50$ Hz; zero-phase forward–backward filtering via scipy.signal.filtfilt .
Downstream training	Original and filtered pseudo-segments; classifier and optimization as in Diff-Aug.

Appendix

Table 7. Key configurations of Clean. Settings not listed here follow those of Diff-Aug.

Method	Synthesis	Params (M)	FLOPs (G)	Latency (ms)
1D-CNN-Trans	$\times$	0.59	0.14	$\sim$ 0.6
DiffMDD ∗	$\checkmark$	1.32	101.69	$\sim$ 94.5
S-NDD (Ours) †	$\times$	1.70	1.42	$\sim$ 2.9
F-NDD (Ours) ‡	$\times$	0.457	0.00405	$\sim$ 16.5

Appendix

Table 8. Computational cost and inference latency measured on an NVIDIA RTX 6000 GPU.

Depression Detection

Depression detection uses language, behavior, speech, and biological signals to identify depressive symptoms and estimate their severity.

See all 77 papers

Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection
May 29, 2026Xiaojing Chen, Jingqi Cheng, Xu Zhao +2Electroencephalography Depression Detection
School of Computer Science and Technology, Hefei University of Technology Hefei, China · School of Computer Science and Information Engineering, Hefei University of Technology Hefei, China
Separating Diagnosis from Disease Representation: Dual-View EEG Learning with Neural-Dynamics-Guided Deformation
Sep 26, 2026Jiaying Wang, Shouqian Shi, Yutong Chen +5Electroencephalography EEG Decoding
Nanjing University · Tianjin University
EEGDM: Label-Efficient EEG Representation Learning with Generative Diffusion Model
Aug 13, 2025Jia Hong Puah, Sim Kuan Goh, Ziwei Zhang +6Electroencephalography EEG Seizure Detection
School of Artificial Intelligence and Robotics, Xiamen University Malaysia · School of Energy and Chemical Engineering, Xiamen University Malaysia · Department of Biomedical Engineering, Universiti Malaya +3

May 29, 2026J/K move · Enter open · S save

Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection

Xiaojing Chen, Jingqi Cheng, Xu Zhao, Wan Jiang+1

School of Computer Science and Technology, Hefei University of Technology Hefei, China · School of Computer Science and Information Engineering, Hefei University of Technology Hefei, China

Electroencephalography Depression Detection

Open paper AskPDF arXiv

Deep learning-based Major Depressive Disorder (MDD) detection using Electroencephalography (EEG) is fundamentally constrained by the "small-sample dilemma." Prevailing generative data augmentation methods not only incur heavy computational overhead but also risk introducing synthetic noise, thereby blurring classification boundaries. To challenge the traditional "data quantity first" convention, we propose a novel framework "Beyond Augmentation": Score-Guided Classification (SGC). SGC does not synthesize pseudo-samples; instead, it utilizes an unsupervised generative network architecture to model the structural and statistical anomaly degrees of samples, serving as the core "Pathological Prior". This prior, after robust normalization, is explicitly fused with deep feature representations, thereby precisely guiding the classifier's decision boundary. Furthermore, to dynamically adapt to varying channel configurations, we propose a Cross-Channel Spatial Adaptation module, utilizing a spatial mapping mechanism to effectively resolve the hardware heterogeneity of mismatched channels in multi-center datasets. Extensive experiments on the Mumtaz2016 and high-density MODMA datasets demonstrate the effectiveness and exceptional generalizability of our method under the challenging "zero data augmentation" setting and at "zero sample synthesis cost". Keywords: Electroencephalography (EEG), Depression Detection, Anomaly Score, Diffusion Models, Few-Shot Learning

Abstract

Depression Detection

Explore similar work

Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection

Separating Diagnosis from Disease Representation: Dual-View EEG Learning with Neural-Dynamics-Guided Deformation

EEGDM: Label-Efficient EEG Representation Learning with Generative Diffusion Model