Neuro-Geospatial Modelling of EEG Affective States Using Literature-Informed Environmental Context

Utsav Poudel; Jagannath Aryal; Subramaniyaswamy Vairavasundaram

cs.AIAug 21, 2026

Neuro-Geospatial Modelling of EEG Affective States Using Literature-Informed Environmental Context

Authors: Utsav Poudel, Jagannath Aryal, Subramaniyaswamy Vairavasundaram

Organizations: School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, India · Earth Observation and AI Research Group, Department of Infrastructure Engineering, The University of Melbourne, Melbourne, VIC 3010, Australia

Abstract

Environmental exposures such as air pollution and greenness have been associated with affective and cognitive outcomes, but EEG and environmental datasets are rarely jointly georeferenced. We investigate whether literature-informed environmental priors can serve as an auxiliary geospatial modality for EEG-based affective-state classification when individual-level exposure data are unavailable. We combine 30-channel EEG from the EAV benchmark (42 participants, aged 20-30 years) with environmental representations derived from OpenAQ, Sentinel-2, Sentinel-5P, and OpenStreetMap data for Astana. A dual-tower architecture combines EEG-Conformer representations with a graph-based environmental encoder. Because the datasets are not co-registered, environmental context is treated as a literature-informed prior rather than measured exposure. Subject-level repeated splits, permutation and label-shuffling controls, dose-response reversal, and domain-shift experiments distinguish architecture-level gains from prior-dependent gains. The multimodal model achieves 76.2% accuracy versus 67.4% for EEG alone. Controls disrupting environmental-label structure retain part of this gain, indicating that the improvement is not attributable solely to environmental information. Replacing the Astana environmental distribution with an independently modeled Singapore distribution reduces accuracy to 72.8%. These findings demonstrate technical feasibility but do not establish an observed or causal exposure-affect association. The study provides a framework for future jointly collected mobile EEG-environment studies. Implementation: https://github.com/r11up/geo-cog

Figures & tables

Figure 1: Data-integration pipeline. EEG recordings and environmental context are linked through PECM and aligned via a directly optimised squared- $\ell_{2}$ objective (Supplementary S2, S6).

Figure 2: Study area base maps, $10\times 10$ km grids for Astana (primary site; EAV EEG collected here) and Singapore (secondary site, environmental-domain shift evaluation only; no Singapore EEG data used).

Figure 3: Geographic context for the environmental-domain shift evaluation: Astana (EEG collection site) and Singapore (independently modelled environmental distribution only; no Singapore EEG data used), separated by 6,596 km, confirming the two environmental domains are structurally independent.

Dataset	Source	Modality	Key attributes	Spatial res.	Temporal res.
EAV [ 24 ]	Lab EEG (Astana)	Neural	30-ch, 42 participants, 5 affective states	—	500 Hz
OpenAQ [ 25 ]	API	Air quality	PM 2.5 , NO 2	Point	Hourly
Sentinel-5P [ 26 ]	Satellite	Atmospheric	NO 2 , O 3 , aerosol	3.5 km †	Daily
LUR model	OpenAQ+OSM+S5P	Atmospheric	PM 2.5 , NO 2	1 km 2	Annual
Sentinel-2 [ 27 ]	Satellite	Vegetation	NDVI	10 m	5-day
OpenStreetMap [ 29 ]	Open vector	Urban morphology	Buildings, roads, FAR	Parcel	Static

Table 1: Datasets used in the multimodal neuro-geospatial framework.

Figure 4: Dual-tower architecture: an EEG-Conformer encodes 30-channel EEG into a 256-D embedding; a graph neural network over a 100-node grid encodes LUR-estimated PM 2.5 /NO 2 , NDVI, FAR and GVI into a matched embedding. The two are fused via bidirectional cross-modal attention and classified into the five EAV affective states.

Figure 5: Experimental protocol: PECM’s label-conditioned sampling applies only to each split’s training partition; validation/test partitions always draw environmental input independently of the true label. The domain-shift evaluation re-fits only the environmental tower to Singapore data, freezing the EEG-Conformer.

Model	Accuracy (%)	Macro F1	ROC-AUC
Random Forest (EEG-only)	$42.6\pm 4.1$	$0.389\pm 0.044$	$0.681\pm 0.033$
LSTM (EEG-only)	$51.3\pm 3.2$	$0.481\pm 0.036$	$0.724\pm 0.028$
Environment-only (matched-distribution)	$22.3\pm 1.4$	$0.214$	$0.507$
EEG-Conformer (EEG-only)	$67.4\pm 2.3$	$0.651\pm 0.026$	$0.824\pm 0.019$
EEG-Conformer + graph encoder (unidirectional)	$73.8\pm 2.0$	$0.714\pm 0.022$	$0.866\pm 0.016$
EEG-Conformer + graph encoder (bidirectional, proposed)	$\mathbf{76.2\pm 1.7}$	$\mathbf{0.741\pm 0.019}$	$\mathbf{0.884\pm 0.013}$

Table 2: Classification performance across five repeated subject-level splits, Astana (mean

\pm

SD); best per column in bold.

Test-time environmental conditioning	Accuracy (%)	Macro F1	ROC-AUC
EEG-only (no environmental input)	$67.4\pm 2.3$	$0.651\pm 0.026$	$0.824\pm 0.019$
Label-independent (throughout this paper)	$\mathbf{76.2\pm 1.7}$	$\mathbf{0.741\pm 0.019}$	$\mathbf{0.884\pm 0.013}$
Label-conditioned (diagnostic upper bound only)	$82.5\pm 1.5$	$0.808$	$0.921$

Table 3: Label-independent versus label-conditioned test-time sampling, proposed model, Astana. The label-conditioned condition upper-bounds test-time label information; label-independent is the protocol used throughout.

Model	Astana Acc. (%)	Singapore Acc. (%)	$\Delta_{\mathrm{shift}}$ (pp)
EEG-Conformer (EEG-only)	$67.4\pm 2.3$	$67.4\pm 2.3$	$0.0$
Environment-only	$22.3\pm 1.4$	$21.6\pm 1.5$	$-0.7$
EEG-Conformer + graph encoder (unidirectional)	$73.8\pm 2.0$	$70.1\pm 2.1$	$-3.7$
EEG-Conformer + graph encoder (bidirectional)	$\mathbf{76.2\pm 1.7}$	$\mathbf{72.8\pm 1.9}$	$-3.4$

Table 4: Environmental-domain shift results: Astana versus an independently modelled Singapore distribution (EEG-Conformer frozen; environmental tower fine-tuned on Singapore LUR data).

\Delta_{\mathrm{shift}}

= Singapore minus Astana accuracy.

Figure 6: Spatial error comparison, Astana (left) vs Singapore (right). Top: per-cell misclassification on the $1~\text{km}^{2}$ grid. Bottom: PECM regimes. Errors cluster in high-density, low-NDVI zones (financial/CBD districts, FAR $>5$ ), where the prior is least discriminative.

Figure 7: GraphLIME attribution scores by predicted affective class, across five repeated splits. Attributions describe reliance on the PECM prior, not independent estimates of environmental effects on affect. Error bars: 1 SD.

Condition	Accuracy (%)	Macro F1	ROC-AUC
No alignment (random env. pairing)	$71.3\pm 2.6$	$0.689\pm 0.028$	$0.851\pm 0.021$
Shuffled env. labels (PECM disabled)	$70.8\pm 2.5$	$0.683\pm 0.027$	$0.847\pm 0.020$
Reversed dose-response (negative control)	$70.9\pm 2.1$	$0.683$	$0.847$
CNN backbone (EEGNet, no Conformer)	$66.9\pm 3.0$	$0.641\pm 0.032$	$0.821\pm 0.024$
Unidirectional attention only	$73.8\pm 2.0$	$0.714\pm 0.022$	$0.866\pm 0.016$
Full proposed model	$\mathbf{76.2\pm 1.7}$	$\mathbf{0.741\pm 0.019}$	$\mathbf{0.884\pm 0.013}$

Table 5: Core ablation results, Astana. All conditions use the EEG-Conformer backbone and environmental tower (Section 2.7 ) unless stated; five further configurations in Supplementary Table S6.1.

Component	Hyperparameter	Value
EEG-Conformer	Temporal conv. kernel; spatial channels; attention heads; feedforward dim; output dim; dropout	25 samples; 30 (depthwise); 8; 512; 256; 0.5
Environmental tower	GCN layers/dims; GRU hidden dim; GRU sequence length; output dim	2 ( $6\to 64\to 128$ ); 128; 1 (static snapshot); 256 (linear proj.)
Bi-attention	Heads; per-head dim; fused dim	8; 32; 512
Classifier	Hidden dims; dropout	$512\to 256\to 5$ ; 0.3
Training	Optimizer; LR; schedule; weight decay; batch size; early stopping	AdamW; $1{\times}10^{-4}$ ; cosine, 100 epochs; $1{\times}10^{-4}$ ; 64; patience 15 (macro F1)
Alignment	Weight $\lambda$ ; PECM bandwidth; PECM conditioning; LUR CV	0.1; 5-fold CV; training partition only; leave-1-station-out

Table S1: Model hyperparameters.

Affective state	Indicator	Quartile	Effect summary	Ref.
Anger, Sadness	PM 2.5 (LUR)	Upper	Reduced executive function; worse cognitive performance and accelerated decline	[ 2 , 31 , 47 ]
Anger, Sadness	NO 2 (LUR)	Upper	Cortical stress reactivity; neuroinflammatory pathway activation	[ 32 ]
Anger, Sadness	FAR	Upper	Attentional depletion; elevated stress biomarkers	[ 34 ]
Relaxed/Calm, Happiness	NDVI	Upper	Reduced physiological arousal; lower cortisol; improved mood	[ 33 , 11 ]
Relaxed/Calm, Happiness	GVI	Upper	Perceived greenness reduces stress biomarkers beyond NDVI	[ 35 , 11 ]
Neutral	All	Middle (IQR)	Baseline environmental conditions	—

Table S2: Dose-response evidence base grounding the PECM label-conditional sampling priors. Pollution concentrations are LUR-estimated.

Figure S1: PECM regime map for Astana based on PM 2.5 , NO 2 , NDVI and GVI. Cells are classified as high-pollution, high-vegetation or baseline; conflict areas align with the higher-error zones identified in the main-text spatial-error comparison figure.

Figure S2: Standardised environmental indicators for the 100 grid cells in Astana ( $10\times 10$ grid, $1~\text{km}^{2}$ each): PM 2.5 , NDVI, FAR and GVI (z-score standardised).

Figure S3: Per-cell Floor Area Ratio and LUR-estimated PM 2.5 for Astana and Singapore, each rendered as a $10\times 10$ spatial matrix with values colour-encoded. The panel illustrates that both indicators vary spatially within each city rather than reducing to a single city-level average.

Figure S4: Buffer-radius sensitivity for NDVI and LUR-estimated PM 2.5 across 100, 250 and 500 m buffers in Astana, shown relative to the native $1~\text{km}^{2}$ LUR baseline. The 250 m buffer minimises this sensitivity and best matches pedestrian-scale ranges reported in environmental psychology [ 33 , 11 ] .

Condition	Accuracy (%)	Macro F1	ROC-AUC
Graph encoder only, no GRU	$75.7\pm 1.8$	$0.735$	$0.881$
MLP-env, no graph structure	$73.9\pm 2.0$	$0.715$	$0.867$
Concatenation fusion (no cross-modal attention)	$74.6\pm 1.8$	$0.723$	$0.872$
Contrastive (InfoNCE) alignment	$75.4\pm 1.8$	$0.732$	$0.878$
RMA-projected alignment (end-to-end)	$76.0\pm 1.7$	$0.739$	$0.883$

Table S3: Additional ablation configurations, Astana primary study area (complements the six core configurations in the main-text ablation table). All conditions use the EEG-Conformer neural backbone and the environmental tower described in Supplementary S2 unless otherwise stated.

Figure S5: UMAP projections of the Astana latent space before (left) and after (right) training with the cross-modal alignment objective. Circles denote neural embeddings, triangles denote environmental embeddings; increased overlap after alignment reflects the learned representation geometry rather than an independently verified environmental mechanism. UMAP is used here only to visualise representations produced by the model’s own alignment objective (defined in the main text), not as an alignment method in itself.

PM 2.5 /NO 2 resolution	Accuracy (%)	Macro F1	ROC-AUC	$\Delta$ Acc.
$1~\text{km}^{2}$ (native LUR, proposed)	$76.2\pm 1.7$	$0.741\pm 0.019$	$0.884\pm 0.013$	—
$5~\text{km}^{2}$	$74.0\pm 2.1$	$0.718\pm 0.023$	$0.869\pm 0.016$	$-2.2$
$10~\text{km}^{2}$	$72.1\pm 2.3$	$0.697\pm 0.025$	$0.856\pm 0.018$	$-4.1$

Table S4: Spatial scale sensitivity (MAUP), primary Astana study area. NDVI, FAR and GVI held at native resolution.

Explore similar work

SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction
Oct 4, 2026Minsu Kim, Ye-Sung Kim, Hyeseong Jeon +3EEG Foundation Models Predictive Representation Learning
Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea · Department of Artificial Intelligence, Hanyang University, Seoul, Republic of Korea · Department of Biomedical Engineering, Hanyang University, Seoul, Republic of Korea
What the LLM Should Not Say: Boundary-Aware Context Grounding for A Seven-Channel EEG Agent
Jun 25, 2026Zhiyuan Xu, Yueqing Dai, Junling Li +1Electroencephalography LLM Grounding
Shanghai Pulse Element Intelligent Technology Co., Ltd. (NeuraDock)
Hierarchic-EEG2Text: Assessing EEG-To-Text Decoding across Hierarchical Abstraction Levels
Feb 24, 2026Anupam Sharma, Harish Katti, Prajwal Singh +2EEG Decoding EEG-Based Visual Decoding
IIT Gandhinagar Gandhinagar, Gujarat, India · NIMH, NIH USA

Oct 4, 2026J/K move · Enter open · S save

SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction

Minsu Kim, Ye-Sung Kim, Hyeseong Jeon, Wooseok Hyung+2

Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea · Department of Artificial Intelligence, Hanyang University, Seoul, Republic of Korea · Department of Biomedical Engineering, Hanyang University, Seoul, Republic of Korea

EEG Foundation Models Predictive Representation Learning Self-Supervised Time Series Representation Learning

Open paper AskPDF arXiv

Electroencephalography (EEG) provides a non-invasive measure of ongoing neural activity, but building general-purpose EEG models remains challenging due to the heterogeneity of subjects, devices, and electrode montages. Existing EEG foundation models predominantly rely on reconstruction-based objectives defined on the observed signal, which contains both neural and non-neural components. We introduce SPERA (Spherical Prior EEG Representation Architecture), an EEG foundation model that adopts the joint-embedding predictive architecture (JEPA) to predict in latent space. SPERA introduces a Legendre-polynomial spatial prior, incorporated into attention to encode varying scalp electrode geometries. Two further components adapt the model to EEG: factorized temporal and spatial attention interleaved with periodic full-attention blocks, and a relational spectral regularizer aligning latent similarity structure with spectral views. Pretrained on approximately 80,000 hours of EEG from 29,048 subjects across 106 datasets, SPERA achieves the highest average balanced accuracy across nine downstream tasks spanning clinical, cognitive, and BCI applications. SPERA further exhibits strong parameter efficiency under linear probing and robustness across varying recording conditions, suggesting its potential as a general-purpose backbone for diverse EEG analyses.

Abstract

Explore similar work

SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction

What the LLM Should Not Say: Boundary-Aware Context Grounding for A Seven-Channel EEG Agent

Hierarchic-EEG2Text: Assessing EEG-To-Text Decoding across Hierarchical Abstraction Levels