NeuroDyn-EEG: An Interpretable Pre-trained Model for EEG Based on Neural Dynamics

Yi Cui; Tong Zhao; Jiaxin Lei; Chuyi Yang; Yifan Cui; Ling Zhang; Yuxiang Yan; Bo Hong

cs.NESep 29, 2026

NeuroDyn-EEG: An Interpretable Pre-trained Model for EEG Based on Neural Dynamics

Authors: Yi Cui, Tong Zhao, Jiaxin Lei, Chuyi Yang, Yifan Cui, Ling Zhang, Yuxiang Yan, Bo Hong

Organizations: Gnosis Neurodynamics Co. Ltd, Room 202, West Area, Building North 1, No. 9 Yard, Shengmingyuan Road, Changping District, Beijing, China · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, China · University of California, Davis, department of psychology · Beijing Key Laboratory of Mental Disorders, National Clinical Research Center for Mental Disorders and National Center for Mental Disorders, Beijing Anding Hospital, Capital Medical University, Beijing, China · Advanced Innovation Center for Human Brain Protection, Capital Medical University, Beijing, China

Abstract

Clinical scalp electroencephalography (EEG) offers a noninvasive window into neural dynamics of neuropsychiatric disorders. However, discriminative deep models often lack anatomically indexed physiological interpretability. We propose NeuroDyn-EEG, a pretraining framework integrating generative priors from neural dynamics. It couples an extended Jansen-Rit neural mass model, leadfield-based source projection, and simulation-based parameter inversion. Trained on synthetic parameter-EEG pairs within physiological ranges, NeuroDyn-EEG estimates 11 regional parameter families across 90 AAL regions plus one global parameter from standard 19-channel EEG, using only ~2.43M trainable parameters. We evaluate the framework across three levels. First, controlled simulations demonstrate robust parameter recovery under diverse noise conditions, while real resting-state EEG evaluations confirm spectral and phase consistency in an inverse-forward closed loop. Second, on four clinical benchmarks (AD65, PD31, Figshare MDD, and TUAB), NeuroDyn-EEG achieves competitive classification performance, securing the highest BACC, AUROC, and AUCPR on PD31 and MDD, and highest BACC on AD65. Third, post hoc regional analyses reveal disease-specific alterations: local synaptic connectivity C_1 involves the most altered regions in AD65, whereas the firing threshold theta ranks first in MDD, offering testable mechanistic hypotheses. Overall, NeuroDyn-EEG maps scalp EEG to anatomically indexed dynamical parameters, bridging representation learning and mechanistic neurophysiology. Code: https://github.com/Gnosis-Neurodynamics/NeuroDyn-EEG.

Figures & tables

Figure 1: The three-stage modeling and validation framework of NeuroDyn-EEG. (1) Whole-brain dynamical simulation : Constructing a 90-node coupled dual-dynamic neural mass model based on the generalized Jansen model, generating a diverse synthetic “parameter–scalp EEG” dataset through physiologically constrained sampling and an intracranial source-to-scalp forward model. (2) Simulation-based multi-branch spatial-spectral parameter inversion model : Training a multi-branch spatial-spectral inverse network using synthetic scalp EEG-neural dynamic parameter pairs to extract interpretable, region-level parameters from scalp EEG. (3) Systematic evaluation : Quantifying model performance via parameter recovery, surrogate noise stress-testing, and closed-loop reconstruction; subsequently comparing downstream classification across real clinical cohorts; and finally examining whether inferred parameters exhibit anatomically localizable case–control differences.

Figure 2: NeuroDyn-EEG parameter inversion network architecture. The framework adopts a multi-branch spatial-spectral design: The scalp temporal branch extracts long-range temporal features via multi-scale convolutions and dilated residual blocks; the source-space branch projects scalp EEG onto 90 source nodes via the leadfield pseudoinverse; and the spectral branch encodes log-magnitude spectral features. Node self-attention and scalp–source cross-attention integrate these representations, and parameter-specific heads produce regional estimates across 90 AAL brain regions together with one global delay coordinate under the specified output constraints.

Figure 3: Simulation parameter recovery and real EEG closed-loop evaluation. ( a ) Sample-level predicted vs. true values across 12 targets on independent synthetic test data; dashed line indicates ideal identity ( $y=x$ ), with Pearson $r$ annotated. ( b ) Heatmap of region-wise Pearson $r$ across AAL cortical parcels. ( c ) Parameter-wise Pearson $r$ curves under four surrogate noise types (white noise, pink noise, EMG, EOG) across SNR 30 to $-5$ dB, referencing clean baseline performance. ( d ) Boxplots of multi-band Phase-Locking Value (PLV) in real resting-state EEG closed-loop reconstruction ( $\delta$ , $\theta$ , $\alpha$ bands). ( e ) Topographic maps of channel-wise reconstruction fidelity on real EEG ( $n=369$ ); Top: 1–40 Hz log PSD Pearson $r$ ; Middle: absolute $\alpha$ -peak frequency error (Hz); Bottom: $1/f$ aperiodic slope error. Black dots indicate standard 10–20 electrode positions.

Method	Model Size	AD65			PD31
Method	Model Size	BACC	AUCROC	AUCPR	BACC	AUCROC	AUCPR
LaBraM	5.8M	0.775 $\pm$ 0.083	0.890 $\pm$ 0.085	0.915 $\pm$ 0.065	0.567 $\pm$ 0.091	0.722 $\pm$ 0.111	0.721 $\pm$ 0.162
CBraMod	4.9M	0.526 $\pm$ 0.031	0.681 $\pm$ 0.104	0.769 $\pm$ 0.046	0.533 $\pm$ 0.075	0.678 $\pm$ 0.127	0.709 $\pm$ 0.143
BrainOmni tiny	8.4M	0.767 $\pm$ 0.140	0.853 $\pm$ 0.121	0.896 $\pm$ 0.089	0.533 $\pm$ 0.075	0.756 $\pm$ 0.214	0.778 $\pm$ 0.227
BrainOmni base	33M	0.713 $\pm$ 0.116	0.871 $\pm$ 0.103	0.917 $\pm$ 0.065	0.600 $\pm$ 0.149	0.683 $\pm$ 0.239	0.722 $\pm$ 0.213
VAEEG	0.46M	0.798 $\pm$ 0.136	0.938 $\pm$ 0.071	0.954 $\pm$ 0.050	0.717 $\pm$ 0.139	0.789 $\pm$ 0.159	0.751 $\pm$ 0.191

Table 1: Quantitative performance and model efficiency comparison. (a) Neurodegenerative disease datasets; (b) General clinical and psychiatric datasets. Size denotes the number of trainable parameters. Bold and underline denote best and second-best results, respectively.

Figure 4: Region-wise case–control structure of inferred JR parameters in AD65 and Figshare MDD. ( a ) AD65 : Bar plot shows the number of AAL regions satisfying Benjamini–Hochberg corrected $q_{\mathrm{BH}}<0.001$ for each JR parameter family; right surface map displays the significance distribution for the top-ranked parameter C1 across AAL-90 space. ( b ) Figshare MDD : Following the same protocol, the top-ranked parameter is $\theta$ . Colorbars represent binary significance masks (1 = region significant at $q_{\mathrm{BH}}<0.001$ ; 0 = non-significant). The global scalar $\mathrm{delay\_scale}$ is excluded from regional counts.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Description	Range	Unit	Distribution
$\tau_{e1}$	Excitatory time constant, slow branch	$[10,35]$	$\mathrm{ms}$	HTN
$\tau_{i1}$	Inhibitory time constant, slow branch	$[10,35]$	$\mathrm{ms}$	Conditional uniform
$\tau_{e2}$	Excitatory time constant, fast branch	$[3.9,8.4]$	$\mathrm{ms}$	HTN
$\tau_{i2}$	Inhibitory time constant, fast branch	$[7.3,16.8]$	$\mathrm{ms}$	Conditional uniform
$C_{1}$	Local coupling factor: pyramidal to excitatory interneurons	$[0.5,1.5]$	$1$	HTN
$C_{2}$	Local coupling factor: excitatory interneurons to pyramidal cells	$[0.4,1.2]$	$1$	HTN

Appendix

Supplementary Table S1: Sampled dynamical parameters for forward simulation. Bounds specify the simulation prior.

Parameter	Description	Value or rule	Unit
$H_{e1}/H_{i1}$	Excitatory/inhibitory synaptic gains, slow branch	$3.25/22$	$\mathrm{mV}$
$H_{e2}$	Excitatory synaptic gain, fast branch	$32.5/1000\tau_{e2}$	$\mathrm{mV}$
$H_{i2}$	Inhibitory synaptic gain, fast branch	$440/1000\tau_{i2}$	$\mathrm{mV}$
$C_{\mathrm{avg}}$	Reference local connectivity scale	135	$1$
$\mu/\sigma_{p}$	Mean/standard deviation of the Gaussian background input	$220/22$	$\mathrm{s}^{-1}$
$\sigma_{\mathrm{ref}}$	Reference amplitude for interregional coupling normalization	$22$	$\mathrm{s}^{-1}$

Appendix

Supplementary Table S2: Fixed simulator constants, gain constraints, and structural settings.

Target	$\tau_{e1}$	$\tau_{i1}$	$\tau_{e2}$	$\tau_{i2}$	$\theta$	$\beta$	$r_{\max}$	$C_{1}$	$C_{2}$	$C_{3}$	$C_{4}$
$w_{k}$	1.2	1.0	1.5	1.5	1.0	1.5	1.0	5.0	5.0	20.0	20.0

Appendix

Supplementary Table S3: Loss weights

w_{k}

for multi-task MAE training (

k

indexes physical targets).

Parameter family	$\alpha^{\mathrm{subj}}$	$\alpha^{\mathrm{node}}$	Notes
$\tau_{e1},\;\tau_{e2}$	$1/4$	$1/12$	Smaller within-sample dispersion
$\theta,\;\beta,\;r_{\max}$	$1/4$	$1/8$	Intermediate within-sample dispersion
$C_{1},\;C_{2},\;C_{3},\;C_{4}$	$1/4$	$1/6$	Larger within-sample dispersion
$\tau_{i1},\;\tau_{i2}$	—	—	Conditional on the matching excitatory time constant; Equation (S1.4)
$\delta_{s}$ ( delay_scale )	—	—	One uniform draw per sample

Appendix

Supplementary Table S4: Hierarchical sampling hyperparameters entering Equations (S1.2)–(S1.3). For a bound width

\Delta_{k}=h_{k}-\ell_{k}

, the parent-Gaussian standard deviations are

\sigma_{\mathrm{subj}}^{(k)}=\alpha^{\mathrm{subj}}\Delta_{k}

and

\sigma_{\mathrm{node}}^{(k)}=\alpha^{\mathrm{node}}\Delta_{k}

. The choice

\alpha^{\mathrm{subj}}=1/4

makes the width of the parent Gaussian’s

\pm 2\sigma

interval equal to

\Delta_{k}

. Truncation determines the resulting distribution moments.

Module	Layer / operation	Parameters	Output size
Input and preprocessing	Scalp EEG segment	$C=19$ , $f_{s}=256$ Hz, $T=1280$ (5 s)	$B\times 19\times 1280$
	Temporal and spatial demeaning	Channel-wise temporal mean removal; sample-wise cross-channel mean removal	$B\times 19\times 1280$
	Peak-to-peak amplitude screening	Retained range: $[3,100]~\mu$ V	$B\times 19\times 1280$
Scalp temporal branch	Four parallel Conv1D paths	$C_{\mathrm{in}}=19$ , $C_{\mathrm{out}}=32$ per path; $k\in\{3,5,7,9\}$ ; $s=2$ ; same padding; no bias	$4\times(B\times 32\times 640)$
	Channel concatenation	$4\times 32=128$ channels	$B\times 128\times 640$
	Dilated residual block 1	Two Conv1D layers; $C=128$ , $k=5$ , $d=2$ ; pre-BN; LeakyReLU $(0.2)$ ; dropout $=0.1$	$B\times 128\times 640$

Appendix

Supplementary Table S5: Layer-wise specification of the NeuroDyn-EEG multi-branch spatial–spectral inversion network.

B

denotes the batch size. Tensor dimensions follow the channel-first convention.

Figure S1: Parameter-wise $R^{2}$ robustness results under four surrogate noise types. The 12 subpanels correspond to the 12 model output targets; each subpanel presents $R^{2}$ values for white noise, pink noise, EMG-like noise, and EOG-like noise across SNR 30 to $-5$ dB, referencing the clean baseline result for that parameter. $R^{2}=1$ indicates perfect agreement; $R^{2}=0$ indicates that prediction mean squared error equals the ground-truth mean baseline error; $R^{2}<0$ indicates residual sum of squares exceeds that baseline error. All curves are computed independently per parameter.

Figure S2: Parameter-wise MAE robustness results under four surrogate noise types. Subpanels, noise types, SNR levels, and clean baseline references are identical to Supplementary Figure S1 . Lower MAE indicates smaller absolute prediction error; the vertical axis preserves the original physical scale of each parameter. Therefore, curves should be used to compare different noise types and SNR conditions within each parameter panel, rather than comparing recovery difficulty across distinct parameters based on absolute MAE magnitude.

ROI	AAL Brain Region	$t$	dof	$p$	$q_{\mathrm{BH}}$	$\|d\|$
2	R Precentral Gyrus	5.23	62.2	$2.1\times 10^{-6}$	$4.3\times 10^{-5}$	1.26
3	L Superior Frontal Gyrus	4.69	63.0	$1.5\times 10^{-5}$	$1.7\times 10^{-4}$	1.14
5	L Superior Orbital Gyrus	5.74	61.4	$3.1\times 10^{-7}$	$1.2\times 10^{-5}$	1.42
6	R Superior Orbital Gyrus	5.11	62.6	$3.2\times 10^{-6}$	$5.3\times 10^{-5}$	1.24
9	L Middle Orbital Gyrus	5.56	62.4	$5.9\times 10^{-7}$	$2.0\times 10^{-5}$	1.34
11	L IFG (p. Opercularis)	4.36	46.8	$7.0\times 10^{-5}$	$5.1\times 10^{-4}$	1.14

Appendix

Supplementary Table S6: Region-wise Welch’s

t

-tests for the C1 parameter family in the AD65 cohort (significant regions with

q_{\mathrm{BH}}<0.001

;

n_{\mathrm{AD}}=36

n_{\mathrm{control}}=29

ROI	AAL Brain Region	$t$	dof	$p$	$q_{\mathrm{BH}}$	$\|d\|$
3	L Superior Frontal Gyrus	-3.18	47.6	$4.8\times 10^{-6}$	$1.0\times 10^{-4}$	0.87
5	L Superior Orbital Gyrus	-6.27	44.7	$1.3\times 10^{-7}$	$5.2\times 10^{-6}$	1.76
13	L IFG (p. Triangularis)	-4.78	50.0	$1.6\times 10^{-5}$	$2.6\times 10^{-4}$	1.32
15	L IFG (p. Orbitalis)	-5.90	50.0	$3.1\times 10^{-7}$	$9.7\times 10^{-6}$	1.63
19	L Posterior-Medial Frontal	-5.96	45.6	$3.4\times 10^{-7}$	$1.0\times 10^{-5}$	1.67
23	L Superior Medial Gyrus	-5.78	42.5	$7.8\times 10^{-7}$	$2.1\times 10^{-5}$	1.63

Appendix

Supplementary Table S7: Region-wise Welch’s

t

-tests for the

\theta

parameter family in the Figshare MDD cohort (significant regions with

q_{\mathrm{BH}}<0.001

;

n_{\mathrm{MDD}}=35

n_{\mathrm{control}}=30

Explore similar work

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces
May 14, 2026Konstantinos Kontras, Trui Osselaer, Stylianos G. Mouslech +12Electroencephalography Foundation Models Brain-Computer Interface
KU Leuven · MIT
Identifying Neural Source Dynamics from Unknown Local Interventions
Sep 28, 2026Ayana Mussabayeva, Jiaqi Sun, Anuar Aimoldin +2Electroencephalography Neural Dynamics
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE · Carnegie Mellon University (CMU), Pittsburgh, PA, USA
NeurDuo-EEG: A Long-Sequence EEG Foundation Model with Persistent State and Explicit Memory
Sep 29, 2026Yifan Wang, Haiping Liu, Yang Cui +13Electroencephalography Foundation Models Electroencephalography
University of Manchester · ETH Zürich · Shanghai Jiao Tong University +5

May 14, 2026, cs.LGJ/K move · Enter open · S save

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

Konstantinos Kontras, Trui Osselaer, Stylianos G. Mouslech, Angeliki-Ilektra Karaiskou+11

KU Leuven · MIT

Electroencephalography Foundation Models Brain-Computer Interface Electrocardiogram Foundation Models Brain Age Wireless Foundation Models

Open paper AskPDF arXiv

Foundation models (FMs) promise to extract unified representations that generalize across downstream tasks. They have emerged across fields, including electroencephalography (EEG), but it is less clear how effective they are in this particular field. Published evaluations differ in datasets, in the EEG-specific preprocessing that might influence reported results, and in the reported metrics, frequently obscuring the clinical relevance in EEG. We introduce NeuroAtlas, the largest EEG benchmark to date: 42 datasets and 260k hours covering clinical EEG (epilepsy, sleep medicine, brain age estimation) and brain-computer interfaces, and include multiple datasets per task along with bespoke clinical evaluation metrics. Besides evaluating EEG-FMs with respect to supervised baselines, we present results from generic time-series FMs. We report three findings. First, EEG-specific FMs do not consistently outperform time-series FMs, which have neither EEG-focused architectures nor been pretrained on EEG. Second, standard machine learning metrics are insufficient to assess clinical utility: thus, we thoroughly evaluate more appropriate measures such as the quality of event-level decision-making, hypnogram-derived features, and the brain-age gap in the domains of epilepsy, sleep, and brain age, respectively. Third, model rankings and performance can vary substantially within domains. We conclude that pretrained models perform largely on par, with only narrow advantages for a few, and that current models do not yet deliver on the promise of an out-of-the-box unified EEG model. NeuroAtlas exposes this gap and provides the datasets and metrics for the next generation of unified EEG FMs.

Abstract

Explore similar work

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

Identifying Neural Source Dynamics from Unknown Local Interventions

NeurDuo-EEG: A Long-Sequence EEG Foundation Model with Persistent State and Explicit Memory