PaperlayerView exact revision on arXiv

Source-linked AI summary

Normative Deviation Modeling for EEG-based Depression Detection

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

arXiv:2610.37338v1cs.LG

TL;DR

EEG-based MDD detection remains challenging because subject variability and acquisition differences can limit generalization across datasets. This paper trains healthy-control structural and distributional references and uses their deviations in two diagnostic approaches. Across two public datasets and multiple evaluation settings, the approaches show effectiveness, with calibration improving performance in selected cross-source settings.

  • Problem

    Subject variability and acquisition differences can entangle disease-related EEG patterns and limit generalization to unseen subjects and independently collected datasets.

    Source: §1 · p. 1

  • Method

    The paper learns structural and distributional references from healthy controls and integrates scalar deviations or feature-level residuals with deep EEG representations for diagnosis.

    Source: §1 · p. 2, §4.3 · p. 4

  • Results

    Across two public EEG datasets, S-NDD and F-NDD are effective in within-source and bidirectional cross-source evaluations, while target HC-assisted calibration improves performance in selected cross-source settings.

    Source: §6 · p. 8

  • Takeaways & Limitations

    Healthy normative references can complement supervised EEG representations for MDD diagnosis across the evaluated within-source and cross-source settings.

    Source: §6 · p. 8

    Generative EEG augmentation may introduce artifacts or depart from authentic recordings’ statistical properties, and its downstream classification benefits may vary across datasets.

    Source: §1 · p. 1

Abstract

from arXiv · show

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.

1 Introduction

EEG-based MDD detection is challenged by subject variability and acquisition differences that can undermine generalization. The paper proposes two ways to use healthy-control normative deviations alongside deep EEG representations and evaluates them across within-source and cross-source settings.

1 Introduction

  • Subject-dependent, noisy EEG and acquisition artifacts can entangle disease-related patterns, while limited psychiatric EEG cohorts make generalization to unseen subjects and datasets difficult.The introduction also notes the high cost of clinical data collection and annotation.Source: §1 · p. 1
  • Generative EEG augmentation may produce artifacts or depart from authentic recordings' statistical properties, and its downstream MDD-classification benefits may vary across datasets.Thus, increasing data quantity alone does not necessarily ensure more reliable detection.Source: §1 · p. 1
  • Healthy-control recordings may serve as normative references for individual departures, but deviations are not necessarily discriminative unless represented and integrated appropriately.The introduction reports preliminary inter-electrode correlation-pattern differences between HC and MDD recordings, illustrated in Figure 1.Source: §1 · pp. 1–2, Figure 1 · p. 1
  • S-NDD combines scalar structural and distributional deviation scores with deep EEG representations, while F-NDD retains dimension-wise residuals for attention-based aggregation and auxiliary supervision.Both approaches learn healthy references exclusively from HC recordings; target HC-assisted score calibration can accommodate cross-source score variation without updating model parameters.Source: §1 · p. 2
  • The study evaluates both approaches on two public EEG datasets using subject-independent within-source and bidirectional cross-source settings, and examines target HC-assisted calibration.Source: §1 · p. 2

2 Related Work

Prior work spans deep EEG-based MDD detection, normative modeling, and cross-domain diagnosis. These lines of research motivate testing whether healthy-reference deviations can support EEG diagnosis across subjects and data sources.

2.1 EEG-based MDD Detection

  • Deep EEG approaches model temporal, spectral, spatial, and graph dependencies, but discriminative representations may retain subject- and acquisition-specific variation that limits cross-dataset generalization.The reviewed methods include CNNs, graph models, hybrid architectures, and general EEG representation models.Source: §2.1 · p. 2

2.2 Normative Modeling and Healthy References

  • Normative modeling characterizes individual departures from a reference population, but neuroimaging findings do not establish its diagnostic utility for EEG-based MDD detection.The representation and use of deviations may affect effectiveness across subjects and data sources.Source: §2.2 · p. 2

2.3 Cross-domain EEG Diagnosis

  • Cross-domain methods reduce distribution discrepancies, yet cross-dataset EEG evaluation also involves acquisition, electrode, and population differences, and alignment alone may not separate disease-related from source-specific variation.Evaluation protocols also differ in whether target-domain data are available for adaptation or calibration.Source: §2.3 · p. 2

3 Preliminaries

The paper frames EEG-based MDD detection as identifying each subject’s MDD status from brain activity. It distinguishes testing on held-out subjects from applying source-trained models to independently collected target data.

3 Preliminaries

  • The diagnostic objective is to identify MDD status by analyzing individual subjects’ EEG recordings.Source: §3 · p. 2
  • Within-source detection trains and evaluates models on disjoint subjects from the same data source.Source: §3 · p. 2
  • Cross-source detection applies source-trained models to an independently collected target dataset despite possible electrode-configuration and acquisition differences.Figure 2 depicts S-NDD and F-NDD as using healthy references, with optional target HC-assisted calibration for cross-dataset predictions.Source: §3 · p. 2, Figure 2 · p. 3

4 Normative Deviation-guided Diagnosis

The framework builds structural and distributional healthy references, then integrates deviations at scalar or feature level with learned EEG representations for MDD detection. It also offers target-HC-assisted score calibration for cross-source evaluation.

4.1 Dual Healthy Reference Modeling

  • VQ-VAE and diffusion models independently learn complementary structural and distributional references from healthy EEG recordings.EEG is spatially aligned to a common 19-electrode layout before modeling.Source: §4.1 · p. 3

4.2 Scalar-level Normative Deviation-guided Diagnosis (S-NDD)

  • S-NDD combines VQ-VAE reconstruction and diffusion denoising errors into a normalized scalar deviation score, then fuses it with deep EEG features.The pretrained EEG encoder and normative models remain frozen during diagnostic training.Source: §4.2 · p. 4

4.3 Feature-level Normative Deviation-guided Diagnosis (F-NDD)

  • F-NDD concatenates dimension-wise structural and distributional residuals, encodes them separately, and aggregates them with deep EEG features using shared attention.Fusion and auxiliary diagnostic supervision train the deviation representation; normative models remain frozen.Source: §4.3 · p. 4

4.4 Target HC-assisted Score Calibration.

  • Optional calibration estimates each model’s score center and scale from target HC recordings, without target MDD labels or model updates.Without target HC data, predictions remain uncalibrated.Source: §4.4 · p. 4

5 Experiments

Experiments evaluate S-NDD and F-NDD on two public EEG datasets in subject-independent within-source and bidirectional cross-source settings. The methods show complementary strengths across evaluation levels, with normative deviations also contributing in ablations and tested sensitivity sweeps.

5.1 Experimental Setup

  • The datasets differ in scale and montage: Mumtaz2016 has 34 MDD and 30 HC subjects with 19 electrodes, while MODMA has 24 MDD and 29 HC subjects with 128 electrodes.Evaluation includes subject-independent 10-fold cross-validation, with each subject's windows kept in one fold; reported metrics cover both window and subject levels.Source: §5.1 · pp. 5–6, §5.2 · p. 6

5.3 Cross-Dataset Generalization

  • Mumtaz2016-to-MODMA evaluation favors S-NDD at 78.50% accuracy and 69.96% F1; in reverse transfer, calibrated F-NDD leads with 89.15% accuracy and 91.56% F1.Reverse-transfer calibration raises F-NDD accuracy from 67.44% to 89.15% and F1 from 79.42% to 91.56%, while AUC changes from 93.41% to 94.21%.Source: Table 2 · p. 6, Table 3 · p. 6, §5.3 · pp. 7–8

5.2 MDD Detection Performance

  • On Mumtaz2016, S-NDD achieves 95.28% window-level accuracy, while F-NDD reaches 97.50% subject-level accuracy and 97.07% subject-level F1.S-NDD also leads in window-level recall, F1, and AUC; F-NDD leads in subject-level precision, while S-NDD has the highest subject-level AUC. Both variants outperform the augmentation baselines in accuracy and F1 at both levels.Source: §5.2 · p. 6

5.3 Cross-Dataset Generalization

  • Cross-source visualizations show distinct regional patterns across structural and distributional deviations in the two transfer directions.For the selected MODMA MDD case, S-NDD shows localized structural deviations alongside more widespread distributional deviations.Source: Figure 4 · p. 7, Figure 5 · p. 7, §5.3 · p. 7

5.4 Ablation Study

  • Ablations show that removing both deviation components reduces accuracy from 95.28% to 85.56% for S-NDD and from 95.20% to 88.75% for F-NDD.Removing either component also reduces accuracy; S-NDD is more affected by excluding distributional deviations, and F-NDD performs best when spectral and inter-electrode correlation features are combined.Source: Table 4 · p. 8, §5.4 · p. 8

5.5 Sensitivity Analysis

  • On Mumtaz2016, S-NDD has highest mean accuracy, F1, and recall at K=512 and α=0.7; F-NDD has highest mean accuracy and recall at K=64 and F1 at K=256.For F-NDD, λdev=0.5 yields the highest means across all three metrics, and performance varies little across tested configurations.Source: Figure 7 · p. 8, Figure 8 · p. 8, §5.5 · p. 8

6 Conclusion and Future Work

S-NDD and F-NDD use healthy-control references to model normative deviations, with distinct strengths across diagnostic levels and transfer directions. The conclusion also identifies residual acquisition differences as a target for future work.

6 Conclusion and Future Work

  • Across two public EEG datasets, S-NDD and F-NDD show effectiveness in within-source and bidirectional cross-source evaluation, with distinct strengths across diagnostic levels and transfer directions.Optional target HC-assisted score calibration further improves performance in selected cross-source settings.Source: §6 · p. 8
  • Future work will investigate unsupervised domain adaptation to mitigate acquisition-related differences remaining after spatial interpolation.The conclusion notes that spatial interpolation mainly aligns electrode geometry.Source: §6 · p. 8

A.1 Cross-electrode Spatial Adaptation

  • Spherical spline interpolation maps heterogeneous EEG montages to the 19-electrode International 10–20 system.For MODMA, source and target electrode coordinates come from MNE GSN-HydroCel-128 and standard 10–20 templates, expressed in the MNE head coordinate system and normalized to the unit sphere.Source: §A.1 · p. 10
  • The interpolation system uses a smoothness order of m=4, truncation order N_g=50, and regularization parameter λ=10^-5.Its final row enforces 1⊤c=0.Source: §A.1 · p. 10

A.2 S-NDD Implementation Details

S-NDD computes structural and distributional deviations from fixed normative references and combines their normalized joint score with deep EEG features for diagnosis.

A.2 S-NDD Implementation Details

  • S-NDD performs spatial adaptation, scores deviations with two frozen references, and predicts without iterative sample generation.The implementation samples one timestep and one noise realization per window.Source: §A.2 · p. 10
  • At inference, S-NDD fuses the normalized deviation score with deep EEG representations before diagnostic classification.The branch uses one sampled timestep and noise realization per window.Source: §A.2 · p. 10
  • Structural scoring excludes codebook and commitment losses, while distributional scoring uses reconstruction error rather than the noise-prediction training loss.The sampled timestep sets the noise level but is not an explicit predictor input.Source: §A.2 · p. 10
  • The joint deviation score is normalized using training-partition P5 and P95 bounds, which remain fixed during inference.If P95−P5<10^-8, the upper bound is set to P5+1.Source: §A.2 · p. 10

A.3 F-NDD Implementation Details

F-NDD derives handcrafted EEG features, standardizes them using source-training healthy-control statistics, and encodes feature-level deviations alongside EEG representations for diagnosis.

A.3 F-NDD Implementation Details

  • The feature-space VQ-VAE and DDPM references are trained independently on source-training HC features and remain frozen during diagnostic inference.The inference pipeline prepares diagnostic inputs using fixed source-HC scaling and asinh transforms.Source: §A.3 · pp. 10–11
  • F-NDD encodes each window’s EEG features and structural and distributional residuals, then uses masked attention to aggregate deviation representations across windows.The pooled EEG and deviation representations are concatenated for classification, with optional HC logit calibration and ensemble prediction.Source: §A.3 · p. 11
  • Spectral features use five frequency bands from 0.5 to 45 Hz, and Pearson channel correlations are clipped, arctanh-transformed, and retained as a 19×19 matrix.The five bands are [0.5, 4), [4, 8), [8, 13), [13, 30), and [30, 45) Hz.Source: §A.3 · p. 11
  • F-NDD extracts 456 handcrafted features per window: 95 spectral entries and 361 transformed channel-correlation entries.The 19×19 correlation entries retain symmetric duplicates and zero diagonal values, so the 456 features are not independent.Source: §A.3 · p. 11
  • Each feature coordinate is normalized using its median and IQR estimated exclusively from healthy-control windows in the source-training fold.Normalization and residual-scaling statistics remain fixed for validation and testing, and target HC used for score calibration does not contribute to them.Source: §A.3 · p. 11

B Cross-Dataset Evaluation Protocols

The cross-dataset protocols keep normative references and diagnostic models fixed during transfer, while distinguishing source-only evaluation, HC-assisted calibration, and target-data adaptation.

B Cross-Dataset Evaluation Protocols

  • Mumtaz2016-to-MODMA S-NDD evaluation is source-only and zero-shot, with model weights and normalization statistics fixed.MODMA recordings are mapped to the 19-electrode montage; neither target data nor target HC updates the trained model or statistics.Source: B Cross-Dataset Evaluation Protocols · p. 11
  • MODMA-to-Mumtaz2016 F-NDD calibration uses seven target HC record groups excluded from diagnostic testing, while references and the diagnostic network remain fixed.Test record groups overlap across calibration runs, so runs are not independent subject cohorts.Source: B Cross-Dataset Evaluation Protocols · p. 11
  • Domain-adaptation comparisons include DANN, DAN, Deep CORAL, and TFST-SDKA, but differ in target-data access and prediction aggregation.Target-HC-restricted adaptation is not equivalent to unrestricted unlabeled-target UDA, and the forward S-NDD and reverse F-NDD settings are not a controlled comparison of deviation granularity.Source: B Cross-Dataset Evaluation Protocols · p. 11

C Generative Baseline Implementation

Four generative augmentation strategies are implemented: Diff-Aug, Hard Neg, Clean, and Hybrid. They vary in how EEG segments are generated, selected, filtered, or spatially mixed.

Diff-Aug

  • Diff-Aug uses classifier-conditioned 1D-CNN–Transformer diffusion with FiLM and guided DDIM to synthesize segments from training anchors, then retrains the classifier with originals.Unless otherwise specified, methods use PyTorch and Adam with batch size 32 on an NVIDIA RTX 6000 GPU.Source: C Generative Baseline Implementation · p. 12

Hard Neg

  • Hard Neg selects anchors from the 70th–90th percentiles of normalized deviation scores, excludes the highest 10%, and generates five pseudo-segments per anchor.Generation uses 100-step DDIM sampling with classifier guidance of 2.5; the deviation score also modulates sequence-feature attention during classification.Source: C Generative Baseline Implementation · p. 12

Clean

  • Clean applies a fourth-order Butterworth low-pass filter with a 50 Hz cutoff to each generated segment before classifier training.The filter uses zero-phase forward–backward filtering via scipy.signal.filtfilt.Source: C Generative Baseline Implementation · p. 12

Hybrid

D Computational Efficiency Analysis

Normative-deviation methods avoid iterative sample synthesis, reducing computational cost relative to DiffMDD. S-NDD has lower reported latency than F-NDD, while F-NDD has lower profiled neural-operation FLOPs.

Computational cost and inference latency

  • About 72× fewer FLOPs and more than 30× lower latency: S-NDD uses 1.42 GFLOPs and about 2.9 ms, versus DiffMDD’s 101.69 GFLOPs and about 94.5 ms.Their parameter counts are comparable: 1.70M for S-NDD versus 1.32M for DiffMDD.Source: D Computational Efficiency Analysis · p. 12
  • F-NDD has 0.457M parameters and 0.00405 GFLOPs of profiled neural operations, with approximately 16.5 ms inference latency.Its latency includes feature extraction and data transfers, while its FLOPs exclude CPU feature processing; S-NDD’s fixed cost includes one VQ-VAE reconstruction pass, one denoising score evaluation, and the diagnostic classifier.Source: D Computational Efficiency Analysis · p. 12

Related papers

This paper cites

Loading 2610.37338v1…