Are EEG Sequences Time Series? EEG Classification with Time Series Models and Joint Subject Training

Johannes Burchert; Thorben Werner; Vijaya Krishna Yalavarthi; Diego Coello de Portugal; Maximilian Stubbemann; Lars Schmidt-Thieme

cs.LGApr 10, 2024

Are EEG Sequences Time Series? EEG Classification with Time Series Models and Joint Subject Training

Authors: Johannes Burchert, Thorben Werner, Vijaya Krishna Yalavarthi, Diego Coello de Portugal, Maximilian Stubbemann, Lars Schmidt-Thieme

Organizations: Institute for Computer Science, University of Hildesheim, Hildesheim, German

Abstract

As with most other data domains, EEG data analysis relies on rich domain-specific preprocessing. Beyond such preprocessing, machine learners would hope to deal with such data as with any other time series data. For EEG classification many models have been developed with layer types and architectures we typically do not see in time series classification. Furthermore, typically separate models for each individual subject are learned, not one model for all of them. In this paper, we systematically study the differences between EEG classification models and generic time series classification models. We describe three different model setups to deal with EEG data from different subjects, subject-specific models (most EEG literature), subject-agnostic models and subject-conditional models. In experiments on three datasets, we demonstrate that off-the-shelf time series classification models trained per subject perform close to EEG classification models, but that do not quite reach the performance of domain-specific modeling. Additionally, we combine time-series models with subject embeddings to train one joint subject-conditional classifier on all subjects. The resulting models are competitive with dedicated EEG models in 2 out of 3 datasets, even outperforming all EEG methods on one of them.

Figures & tables

	Data Usage
Model	Single	All	Transfer	Meta	Specialised Layer	Architecture
MAtt ( Pan et al., 2022 )	✓	$-$	$-$	$-$	Riemann space transformation	Attention
MBEEGSE ( Altuwaijri et al., 2022 )	✓	$-$	$-$	$-$	EEGNetBlock	CNN
EEG-TCNet ( Ingolfsson et al., 2020 )	✓	$-$	$-$	$-$	TCN-Blocks	CNN
TCNet-Fusion ( Musallam et al., 2021 )	✓	$-$	$-$	$-$	notch filter	CNN
FBCNet ( Mane et al., 2021 )	✓	$-$	$-$	$-$	spectral filtering	CNN

Table 1: Comparison of the EEG Classification literature.

Figure 1: In the figure above, we demonstrate the three proposed methods of using the subject information, where (a) is Constant Indicator Channels, (b) the Constant Embedding Channels, and (c) Separate Embedding. Here

\oplus

denotes the concatenation of two tensors across the channel dimension, and for (b)

d

represents the embedding size. We have set the subject Id to 2 as an example.

Dataset	Subjects	Instances	Channels	Timesteps	Classes	Task
Mi	9	7.452	22	438	4	Motor Imagery
SSVEP	11	5.500	8	128	5	Visual Stimulus
ERN	16	5.440	56	160	2	Error recognition

Table 2: Summary for the three datasets with the number of subjects, instances, channels, timesteps and classes.

Models	MI	SSVEP	ERN
ShallowConvNet	61.84 $\pm$ 6.39	56.93 $\pm$ 6.97	71.86 $\pm$ 2.64
EEGNet	57.43 $\pm$ 6.25	53.72 $\pm$ 7.23	74.28 $\pm$ 2.47
SCCNet	71.95 $\pm$ 5.05	62.11 $\pm$ 7.70	70.93 $\pm$ 2.31
EEG-TCNet	67.09 $\pm$ 4.66	55.45 $\pm$ 7.66	77.05 $\pm$ 2.46
TCNet-Fusion	56.52 $\pm$ 3.07	45.00 $\pm$ 6.45	70.46 $\pm$ 2.94
FBCNet	71.45 $\pm$ 4.45	53.09 $\pm$ 5.67	60.47 $\pm$ 3.06

Table 3: Performance comparison for the datasets BCIC-IV-2a (MI), MAMEM EEG SSVEP Dataset II (SSVEP) and the BCI challenge error-related negativity (ERN). We report the average accuracy for MI and SSVEP and the AUC for ERN over 5 runs, respectively. The first block contains the common baselines for the EEG literature, where MAtt † are the results reported in the original paper and MAtt is our reproduction. The second block is composed of the proposed TSC models, ResNet and Inception, for the subject-specific approach. The last block consists of the subject-conditional approach where all subjects are trained jointly, and we utilize the subject information. The best result is highlighted in bold and the second best is underlined.

Subject	Inception	ResNet	MAtt
1	$78.96\pm 1.82$	$72.83\pm 3.22$	$\textbf{86.94}\pm 1.36$
2	$41.25\pm 2.63$	$38.12\pm 3.27$	$\textbf{56.94}\pm 2.42$
3	$82.09\pm 3.05$	$70.76\pm 3.10$	$\textbf{88.33}\pm 1.17$
4	$52.43\pm 2.40$	$49.72\pm 3.41$	$\textbf{67.85}\pm 3.42$
5	$38.75\pm 3.74$	$37.50\pm 4.51$	$\textbf{61.32}\pm 1.07$
6	$48.61\pm 1.78$	$44.58\pm 4.39$	$\textbf{52.50}\pm 2.52$

Table 4: Performance Comparison on the MI dataset for the subject-specific case. We report the average accuracy over 5 runs.

Subject	Inception	ResNet	MAtt
1	$80.40\pm 2.06$	$56.40\pm 3.98$	$\textbf{81.60}\pm 2.87$
2	$86.60\pm 1.62$	$75.20\pm 3.82$	$\textbf{89.40}\pm 1.36$
3	$\textbf{61.60}\pm 3.07$	$38.20\pm 2.64$	$58.20\pm 5.64$
4	$\textbf{25.00}\pm 4.00$	$20.40\pm 4.41$	$20.60\pm 3.88$
5	$25.00\pm 6.72$	$22.60\pm 1.74$	$\textbf{26.40}\pm 4.80$
6	$\textbf{79.20}\pm 1.72$	$38.20\pm 3.92$	$79.00\pm 2.68$

Table 5: Performance Comparison on the SSVEP dataset for the subject-specific case. We report the average accuracy over 5 runs.

Subject	Inception	ResNet	MAtt
2	$81.14\pm 4.60$	$74.03\pm 3.37$	$\textbf{81.86}\pm 2.56$
6	$\textbf{90.89}\pm 2.02$	$88.40\pm 2.67$	$68.34\pm 1.76$
7	$83.45\pm 6.80$	$\textbf{85.45}\pm 7.05$	$69.28\pm 10.90$
11	$59.50\pm 10.52$	$55.27\pm 9.19$	$\textbf{70.50}\pm 2.66$
12	$\textbf{74.49}\pm 3.67$	$68.62\pm 9.92$	$60.62\pm 4.70$
13	$56.34\pm 2.26$	$51.15\pm 4.16$	$\textbf{62.92}\pm 3.65$

Table 6: Performance Comparison on the ERN dataset for the subject-specific case. We report the average AUC over 5 runs.

Method	MI	SSVEP	ERN
ResNet
SA	55.42 $\pm$ 2.72	50.48 $\pm$ 0.48	69.76 $\pm$ 0.72
CIC	55.54 $\pm$ 1.72	50.39 $\pm$ 0.24	73.09 $\pm$ 0.66
CEC	55.21 $\pm$ 2.52	54.15 $\pm$ 1.19	73.06 $\pm$ 2.21
SE	49.08 $\pm$ 2.07	49.88 $\pm$ 0.75	68.77 $\pm$ 1.43
Inception

Table 7: Performance comparison with subject information for the joint training protocol for the Constant Indicator Channels (CIC), Constant Embedding Channels (CEC), and Separate Embedding (SE). (SA) is the subject-agnostic approach with no additional subject information. We report the average accuracy for MI and SSVEP and the AUC for ERN over 5 runs, respectively. The best result per dataset is marked in bold and the best method per model underlined.

Time Series Classification

Time series classification assigns labels to sequential data, using temporal patterns to support decisions in health, sensing, science, and industry.

See all 175 papers

LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification
Mar 11, 2026Ahmad Bdeir, Johannes Burchert, Tom Hanika +2Cross-Subject EEG Decoding Electroencephalography
Data Science Group, Universität Hildesheim · ISMLL, Universität Hildesheim
DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models
Oct 10, 2026Yuqiu Deng, Muyun Jiang, Zongpeng Zhang +8EEG Foundation Models Cross-Subject EEG Decoding
Uploading Inc · College of Computing and Data Science, Nanyang Technological University · Department of Biostatistics, School of Public Health, Peking University +2
MedTS-TTT: Test-Time Training for Medical Time Series Classification
Jun 19, 2026Mingzhi Chen, Yiyu Gui, Guibo LuoTime Series Classification Electrocardiogram Classification
Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University, Shenzhen, China

Mar 11, 2026J/K move · Enter open · S save

LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification

Ahmad Bdeir, Johannes Burchert, Tom Hanika, Lars Schmidt-Thieme+1

Data Science Group, Universität Hildesheim · ISMLL, Universität Hildesheim

Cross-Subject EEG Decoding Electroencephalography EEG Decoding Attention Mechanisms

Open paper AskPDF arXiv

Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain-computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with a Lorentz boost-based LoRA mechanism and hyperbolic projection layers to reduce overfitting in geometric representations. We evaluate LAtte with and without finetuning in three settings: subject-specific, subject-conditional, and leave-one-subject-out (LOSO) on five established EEG datasets, achieving a consistent improvement in performance over current state-of-the-art methods for smaller datasets and maintaining performance for larger datasets.

Abstract

Time Series Classification

Explore similar work

LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification

DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models

MedTS-TTT: Test-Time Training for Medical Time Series Classification