PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Modality-Agnostic Seizure Classification: A Unified Deep Learning Framework for Heterogeneous EEG and Intracranial EEG Signals

Ramya Sundaram, Muhammad Aslam Jarwar, M Uma, Sunder Ali Khowaja, Qurat-Ul-Ain Mastoi

arXiv:2610.16596v2cs.LGcs.AI

TL;DR

Existing seizure systems commonly analyze EEG or iEEG separately, despite their use across varied clinical monitoring conditions. This paper harmonizes and combines the modalities, balances the data, and classifies seizures with EEGNet-ViT; it reports higher accuracy than the listed comparison architectures.

  • Problem

    Many seizure systems analyze EEG or iEEG separately, while their differing recording conditions motivate a framework that can handle both modalities.

    Source: §I · p. 1

  • Method

    The framework harmonizes heterogeneous EEG and iEEG, uses cGAN-based augmentation to address class imbalance, and classifies signals with EEGNet-ViT.

    Source: §I · p. 2

  • Results

    98.63% accuracy was achieved by the proposed framework, highest among the listed architectures: CNN+Transformer 97.5%, EEGWaveNet 97%, CNN+CBAM 95%, and Transformer Net 83.81%.

    Source: §IV · p. 7

  • Takeaways & Limitations

    The experiments support the effectiveness of multimodal EEG–iEEG learning for robust seizure classification.

    Source: §I · p. 2, §V · p. 7

Abstract

from arXiv · show

Seizure prediction and detection, essential for clinical treatment of epilepsy, remain the crucial task of this neurological disease. EEG and iEEG are widely used to analyze epilepsy due to its capability of capturing discriminative brain activities at high time resolution. However, existing works generally analyze either EEG or iEEG seizure activity separately, where EEG analysis has been well-researched, while more and more works consider iEEG for its higher specificity at the spatial level and signal quality. Currently, most of the existing seizure prediction systems are single-modal in nature, which limits its application to a variety of clinical conditions. Therefore, a common framework to analyze seizures with both EEG and iEEG signals needs to be developed. This paper presents a new framework of multimodal seizure classification integrating the heterogeneous EEG and iEEG recordings from different subjects and sources. A specific framework, called Unified Multimodal Signal Harmonization (UMSH), is developed for harmonizing signals with varied sampling rates and recording conditions. Moreover, to handle the extremely class-imbalanced problem in the data sets, we use a Conditional Generative Adversarial Network (cGAN) based data augmentation technique to generate synthetic samples to balance the class distribution. The proposed framework uses the EEGNet-Vision Transformer (EEGNet-ViT) as the feature extraction and classification network to learn the local discriminative patterns as well as global contextual information in the multimodal seizure data. Experimental results demonstrate the effectiveness of the proposed framework for robust seizure classification, emphasizing the potential of integrated EEG-iEEG multimodal learning in advancing next-generation epilepsy monitoring and clinical decision-support systems.

I. INTRODUCTION

Seizure monitoring must address the complementary trade-offs of scalp EEG and iEEG, while existing systems often remain tied to one modality or recording environment. The paper proposes a unified EEG–iEEG classification framework with signal harmonization, class balancing, and a hybrid neural network.

I. INTRODUCTION

  • Scalp EEG is inexpensive and non-invasive but spatially imprecise, whereas iEEG offers greater signal accuracy and spatial resolution but is invasive, costly, and less available.The paper motivates a framework that can accommodate both modalities and diverse monitoring conditions.Source: §I · p. 1
  • The framework integrates EEG and iEEG recordings from independent datasets in a multimodal seizure-classification system.The stated contributions describe a common framework for the two modalities.Source: §I · p. 2
  • UMSH harmonizes heterogeneous EEG and iEEG signals by synchronizing sampling frequencies, normalizing, segmenting, and standardizing the data.The framework is intended to combine signals with differing recording properties into a unified dataset.Source: §I · pp. 1–2
  • A cGAN-based augmentation strategy generates synthetic samples to mitigate class imbalance and improve data diversity.The paper identifies the merged dataset’s imbalanced class distribution as a challenge for learning across seizure categories.Source: §I · p. 2
  • EEGNet-ViT combines local discriminative feature learning with global contextual representation for seizure classification.The introduction attributes local space-temporal features to EEGNet and long-range dependencies to the Vision Transformer.Source: §I · p. 2
  • The introduction reports that experiments demonstrate the effectiveness of multimodal EEG–iEEG learning for robust epilepsy analysis.Source: §I · p. 2

II. RELATED WORKS

Prior work has developed seizure detection and classification methods for scalp EEG and iEEG, but the reviewed systems generally treat the modalities separately. The paper frames a unified model as a response to their differing signal properties and acquisition conditions.

A. EEG-Based Seizure Detection and Classification

  • EEG seizure studies span handcrafted features, channel selection, hybrid classifiers, and deep models evaluated on datasets including CHB-MIT, Bonn, and TUSZ.The review also covers multiclass classification and methods combining feature selection, fusion, and classification.Source: A. EEG-Based Seizure Detection and… · p. 2
  • iEEG research includes hyperdimensional computing, transformer and multitask models, lightweight attention CNNs, and convolutional-transformer approaches.The review describes applications in seizure detection, prediction, and epilepsy-state classification.Source: B. iEEG-Based Seizure Detection and… · p. 2
  • The reviewed literature generally studies scalp EEG or iEEG independently, with signal-property, electrode-topology, spatial-resolution, noise, and sampling-rate differences complicating a joint model.The paper identifies this as the research gap addressed by its unified classification framework.Source: B. iEEG-Based Seizure Detection and… · pp. 2–3

III. PROPOSED METHODOLOGY

The methodology combines scalp EEG and iEEG into a unified dataset, harmonizes their representations, and uses a balancing method before classification. Its stated workflow uses EEGNet and ViT to extract and classify seizure-related features.

III. PROPOSED METHODOLOGY

  • The proposed pipeline harmonizes EEG and iEEG, balances the merged data, then extracts features and classifies them with hybrid EEGNet and ViT models.The method is intended to accommodate both recording modalities despite differences in signal characteristics, channel setup, and sampling frequencies.Source: §III · p. 3

A. Dataset Creation

  • The UMSH algorithm resamples iEEG from 5000 Hz to 256 Hz, concatenates four consecutive samples, and creates 12-second, 3072-point EEG windows.The algorithm also extracts ictal and non-seizure iEEG samples, ictal EEG segments, and preictal EEG segments from 10 minutes before a seizure; it standardizes and combines the samples.Source: §III · p. 3
  • The dataset combines CHB-MIT scalp EEG with Mayo Clinic iEEG recordings from Kaggle.The paper describes forming one dataset for generalized seizure classification.Source: A. Dataset Creation · p. 3

1) CHB-MIT dataset:

The paper draws on CHB-MIT scalp EEG and a multicentre intracranial EEG dataset. The two sources differ in recording setup and sampling rate, providing distinct data for the multimodal work.

1) CHB-MIT dataset:

  • CHB-MIT provides long-term scalp EEG from 23 pediatric patients, recorded across 23–28 channels at 256 Hz, with 173 annotated seizures.The dataset includes around 500 hours of continuous recordings.Source: 1) CHB-MIT dataset · p. 3

2) Multicentre intracranial EEG dataset:

3) Multimodal EEG–iEEG Dataset Creation:

The multimodal dataset combines selected EEG and iEEG seizure-status segments, while UMSH standardizes their representations. Its iEEG harmonization resamples signals and joins contiguous segments into longer windows.

3) Multimodal EEG–iEEG Dataset Creation:

  • The combined dataset uses iEEG non-seizure and ictal segments and CHB-MIT EEG preictal and ictal segments, yielding three clinical classes.The classes are Non-Seizure, Preictal, and Ictal.Source: 3) Multimodal EEG–iEEG Dataset Creation · p. 4

B. Unified Multimodel Signal Harmonization

1) iEEG Signal Harmonization:

  • iEEG signals are resampled from 5000 Hz to 256 Hz, then four contiguous 3-second records are concatenated into 12-second segments.The resulting 12-second iEEG segment has 3072 data points; the four-segment concatenation is applied to seizure and non-seizure classes.Source: 1) iEEG Signal Harmonization · p. 4

2) EEG Signal Harmonization:

EEG and iEEG are processed into matching 12-second, 256 Hz windows with 3072 points, then standardized into a shared representation. The combined dataset remains imbalanced across seizure-status classes.

2) EEG Signal Harmonization:

  • CHB-MIT EEG is reduced to a representative signal by averaging three clinically relevant channels, then divided into 12-second windows of 3072 samples.Ictal segments and preictal segments from 10 minutes before seizure onset are selected.Source: 2) EEG Signal Harmonization · p. 4

3) Unified Multimodal Signal Harmonization:

  • After preprocessing, EEG and iEEG share a 256 Hz sampling frequency, 12-second duration, and 3072-point segment length for a unified feature space.The combined dataset includes EEG Preictal, EEG Ictal, iEEG non-Seizure, and iEEG Seizure classes; standardization reduces modality differences.Source: 3) Unified Multimodal Signal… · pp. 4–5
  • The combined dataset has 27,709 Non-Seizure samples versus 6,925 Ictal and 8,613 Preictal samples, motivating class balancing.Source: 3) Unified Multimodal Signal… · p. 5

C. Conditional Generative Adversarial Network (CGAN)-Based Data Balancing

The method uses a class-conditioned GAN to oversample minority seizure classes. Its generator creates synthetic examples from noise and labels, while a discriminator distinguishes generated from real samples.

  • The joint EEG–iEEG dataset is highly imbalanced, so a conditional GAN generates class-specific synthetic samples for minority classes before classification.Conditioning on class labels is used instead of standard oversampling.Source: C. Conditional Generative Adversarial… · p. 5
  • The generator takes a noise vector and class label to make a sample, while the discriminator evaluates real and generated samples with their labels.Adversarial learning trains the generator to produce better fake minority-class samples.Source: C. Conditional Generative Adversarial… · p. 5
  • CGAN oversamples the Ictal and Preictal classes until the distribution is balanced, with 10,000 synthetic samples extracted for every case.The resulting balanced EEG–iEEG data are used to train the unified classifier.Source: C. Conditional Generative Adversarial… · p. 5

D. Improved EEGNet-ViT architecture for multimodal classification

The architecture uses EEGNet to extract efficient local spatiotemporal features, then projects them into tokens for a Vision Transformer to capture global context. Spatial filtering, pooling, and dropout compress features while retaining discriminative information.

D. Improved EEGNet-ViT architecture for multimodal classification

  • EEGNet is selected for its lightweight design and efficient learning of spatial-temporal electrophysiological features using depthwise and separable convolutions.These convolution types reduce trainable parameters compared with standard CNNs.Source: D. Improved EEGNet-ViT architecture… · pp. 5–6
  • EEGNet uses temporal convolution to extract frequency-specific patterns, followed by depthwise spatial filtering of the temporal features.The temporal layer uses 16 filters with a (64,1) kernel; the spatial layer uses a (1,16) kernel and depth multiplier 2.Source: D. Improved EEGNet-ViT architecture… · p. 6
  • Average pooling and dropout compress the representation before separable convolution updates temporal-spatial information efficiently.Pooling reduces the temporal representation from 3072 to 768 samples and then to 96 tokens; dropout is set to 0.4.Source: D. Improved EEGNet-ViT architecture… · p. 6
  • A time-distributed dense layer maps EEGNet feature vectors into 64-dimensional embeddings, which the Vision Transformer uses to model global context and long-range dependencies.The feature maps are flattened and reshaped into a sequential token representation before entering the Transformer.Source: D. Improved EEGNet-ViT architecture… · p. 6

IV. RESULTS AND ANALYSIS

The unified framework achieved the highest reported accuracy among the compared architectures and maintained strong metrics across training, testing, and validation. Its evaluation used balanced training data with a separate validation set and showed few errors, mainly between ictal and preictal classes.

IV. RESULTS AND ANALYSIS

  • Training used 49,800 samples, including CGAN-generated ictal and preictal samples; the distinct 20,955-sample validation set used original ictal and preictal samples.The validation set was formed from records excluded from training.Source: §IV · p. 6, Table V · p. 6, Table VI · p. 7
  • Training, testing, and validation accuracy were 98.87%, 97.71%, and 98.63%, respectively, with F1-scores of 98.86%, 97.68%, and 98.63%.The authors report similarly high precision and recall across the three datasets.Source: §IV · p. 7
  • The confusion matrices show few misclassifications across training, testing, and validation, with most errors between the clinically comparable ictal and preictal states.The validation matrix is described as having very low error rates.Source: §IV · p. 7
  • 98.63% accuracy was highest; stated gains were 1.63% over Transformer Net, 1.13% over CNN+CBAM, 3.63% over EEGWaveNet, and 14.82% over CNN+Transformer.Figure 6 presents the comparison with existing deep-learning architectures.Source: §IV · p. 7

A. Ablation Analysis of EEG and iEEG Modalities

Joint training and testing on EEG and iEEG produced the strongest reported accuracy, while cross-modal training and testing in either direction also generalized but performed less well.

A. Ablation Analysis of EEG and iEEG Modalities

  • 98.63% accuracy was achieved with joint EEG-iEEG training and testing, versus 95.45% for EEG Training-iEEG Testing and 94.67% for iEEG Training-EEG Testing.The cross-modal settings showed generalization but fell below the joint-modality setting.Source: A. Ablation Analysis of EEG and iEEG… · p. 7

B. Ablation Analysis of proposed framework

The conclusion reports the strongest accuracy for joint EEG-iEEG training and testing, while the ablation compares training without CGAN against a setting using UMSH and CGAN. The authors identify larger multicenter validation and additional physiological modalities as future work.

B. Ablation Analysis of proposed framework

V. CONCLUSION

  • 98.63% accuracy with joint EEG-iEEG training and testing exceeded EEG Training-iEEG Testing (95.45%) and iEEG Training-EEG Testing (94.67%).The conclusion attributes the result to the combined modalities representing more seizure-related information.Source: §V · p. 7
  • Future work will add physiological modalities, introduce explainable AI, and validate the framework on larger multicenter datasets.The conclusion frames these as directions for extending the seizure prediction framework.Source: §V · p. 7
Loading 2610.16596v2…