🤖 machine learning

EEG-FuseFormer: A Transformer-Driven Feature Fusion Framework for Seizure Onset Prediction

The paper proposes EEG-FuseFormer, a transformer-driven feature fusion framework that combines CNN-LSTM and ResNet-18 features to achieve state-of-the-art seizure onset prediction accuracy (98.85% recall) on the CHB-MIT dataset, demonstrating superior performance in cross-patient scenarios and evaluating its computational efficiency across diverse hardware platforms.

Original authors: Vigneshwar Hariharan (National University of Singapore), Chithra Reghuvaran (University College Dublin), Arlene John (University of Twente), Nhat Pham (Cardiff University), Omer Rana (Cardiff Universi

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vigneshwar Hariharan (National University of Singapore), Chithra Reghuvaran (University College Dublin), Arlene John (University of Twente), Nhat Pham (Cardiff University), Omer Rana (Cardiff University), Deepu John (University College Dublin), Ganesh Neelakanta Iyer (National University of Singapore)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is like a busy orchestra playing a complex piece of music. Sometimes, this music suddenly goes off-key, leading to a "seizure." For people with epilepsy, these unexpected musical crashes can be dangerous. The goal of this paper is to build a super-smart "conductor" (an AI) that can listen to the orchestra and shout, "Warning! A crash is coming!" before it actually happens.

Here is how the researchers, Vigneshwar Hariharan and his team, built this conductor, explained in simple terms:

1. The Problem: Listening to a Noisy Room

Brain signals (EEG) are messy. They change over time, and every person's "music" sounds different. Trying to predict a seizure is like trying to guess when a storm is coming just by looking at a single cloud. It's hard because one type of cloud isn't enough to tell the whole story.

2. The Solution: A Two-Eyed Detective

The team created a new AI model called EEG-FuseFormer. Think of this model as a detective with two very different pairs of eyes, both looking at the same brain signal but seeing different things:

Eye 1 (The Time Traveler): This eye uses a system called 1D-CNN-LSTM. It looks at the raw brain waves as they happen second-by-second. It's great at spotting how the rhythm changes over time, like noticing if the music is speeding up or slowing down.
Eye 2 (The Frequency Analyst): This eye uses a system called ResNet-18. But before it looks, it turns the sound into a colorful picture called a "spectrogram" (using something called STFT). This picture shows the pitch and energy of the brain waves. It's like looking at a musical score to see which notes are being played loudly.

3. The Magic Glue: The Transformer

Usually, if you have two detectives, you might just ask them to vote on the answer. But this paper uses something smarter: a Transformer.

Imagine a wise mediator sitting between the two detectives. Instead of just averaging their opinions, the Transformer listens to both, figures out which detective is more confident about a specific clue, and blends their insights together perfectly. This "fusion" allows the AI to see the full picture: both the timing and the pitch of the brain's activity.

4. The Results: A Very Accurate Predictor

The team tested this new AI on a famous dataset of brain recordings from children (the CHB-MIT dataset).

The Score: The new "Two-Eyed" model was incredibly accurate, catching 98.85% of the impending seizures.
The Comparison: It did better than using just the "Time Traveler" eye or just the "Frequency Analyst" eye alone. It also beat many other high-tech methods found in previous research.

5. The "Stranger Danger" Test (Cross-Patient Testing)

Here is a tricky part: A model trained on Person A often fails when looking at Person B because everyone's brain music is unique.

The Old Way: If you train the AI on 12 people and test it on the 13th without any help, it gets confused and misses about half the seizures (only ~59% accuracy).
The New Trick (Target Adaptation): The researchers tried a "fine-tuning" trick. They took the smart AI, showed it just a tiny bit of the new person's brain data, and let it "re-learn" for a few minutes.
The Result: This small adjustment was a game-changer. The accuracy jumped from 59% to 93%. It's like giving the detective a quick briefing on the new suspect's habits before the case starts.

6. The Catch: Speed vs. Smarts

The paper also checked how fast this AI runs on different computers.

The Simple Model (CNN-LSTM): Fast and light, like a bicycle. It's good if you have a small, low-power device (like a wearable gadget).
The Complex Model (EEG-FuseFormer): Super smart but heavy, like a sports car with a big engine. It needs powerful computers (like high-end servers) to run quickly.
The Trade-off: You have to choose: Do you want the absolute best accuracy (use the heavy model on a powerful server), or do you need it to run on a small battery-powered device (use the simpler model)?

Summary

The paper introduces EEG-FuseFormer, a smart AI that combines two different ways of looking at brain waves to predict seizures. It works best when it gets a tiny bit of help to learn a new patient's specific brain patterns. While it is the most accurate method tested, it requires more computing power than simpler models, so doctors and engineers must choose the right tool based on where they plan to use it.

Technical Summary: EEG-FuseFormer

Problem Statement
Epilepsy is a prevalent neurological disorder characterized by unpredictable seizure events, posing significant risks to patient quality of life. While automated seizure onset prediction using Electroencephalography (EEG) offers a potential mitigation strategy, the inherent complexity of EEG signals—exhibiting irregular temporal variations and significant inter-patient variability—makes accurate detection challenging. Existing approaches often rely on single feature sets or patient-specific models that fail to generalize to unseen patients. Furthermore, simply fusing multiple features can introduce runtime overhead and irrelevant data, necessitating a more sophisticated fusion mechanism that balances performance with computational efficiency.

Methodology
The paper proposes EEG-FuseFormer, a transformer-driven feature fusion framework designed to predict seizure onset by leveraging both time-domain and frequency-domain characteristics of EEG signals. The methodology is structured as follows:

Dataset and Pre-processing: The study utilizes the CHB-MIT scalp EEG dataset, selecting 13 pediatric patients with intractable seizures. Raw signals are filtered (1 Hz high-pass and 50–60 Hz notch) and segmented into 5-second windows. To address class imbalance between interictal (normal) and preictal (pre-seizure) states, a sliding window approach with overlapping segments is employed, followed by random under-sampling. Post-ictal segments are excluded to avoid noise.
Dual-Channel Feature Extraction: The architecture processes data through two parallel channels:
1. Channel 1 (Time-Domain): A 1D-CNN-LSTM architecture processes raw filtered signals. It utilizes two convolutional blocks (with batch normalization, ReLU, and max-pooling) to extract spatial and temporal features, followed by a two-layer LSTM to capture temporal dependencies.
2. Channel 2 (Frequency-Domain): Raw signals are converted into Short-Time Fourier Transform (STFT) spectrograms. These are fed into a ResNet-18 network (modified to accept 128×90 concatenated spectrogram inputs) to extract spatial features from the time-frequency representation.
Transformer Fusion: Features from both channels are projected into a common 64-dimensional vector space. These vectors are then fused using a Transformer Encoder consisting of two layers, each with an 8-head multi-head self-attention mechanism and a feed-forward network. This allows the model to dynamically weigh the importance of features from both architectures.
Classification: The fused features pass through dense layers and a softmax function to classify the state as either interictal or preictal.
Evaluation Frameworks: The model is evaluated under two scenarios:
1. Patient-Specific: Training and testing on the same patient.
2. Cross-Patient: Training on $N-1$ patients and testing on the remaining patient. This includes a Target Adaptation phase where the pre-trained model is fine-tuned on a small sample of the target patient's data (3–5 epochs) before final testing.

Key Contributions

Novel Architecture: Introduction of EEG-FuseFormer, which integrates the temporal modeling capabilities of CNN-LSTM with the spatial feature extraction of ResNet-18 via a Transformer encoder, moving beyond simple concatenation or attention mechanisms.
Cross-Patient Generalization: A rigorous evaluation of the model's ability to generalize across patients, demonstrating that Target Adaptation significantly improves performance metrics compared to conventional cross-patient validation without fine-tuning.
Computational Trade-off Analysis: A comprehensive assessment of the model's runtime and complexity across diverse hardware platforms (NVIDIA L4, A100 SXM4, and Jetson Nano), highlighting the performance-complexity trade-off between the proposed fusion model and its constituent parts.

Results

Patient-Specific Performance: On the CHB-MIT dataset, EEG-FuseFormer achieved a mean Recall of 98.85% and a mean F1-Score of 97.92%, outperforming standalone ResNet-18 (97.53% Recall) and 1D-CNN-LSTM (93.4% Recall). It also maintained a low False Alarm Rate (FAR) of 0.12/hour.
Cross-Patient Performance: In the cross-patient setting without adaptation, recall dropped significantly (e.g., ~59% for EEG-FuseFormer). However, with Target Adaptation, the recall improved to 93%, precision to 82%, and F1-score to 88%. This represents a substantial improvement over the non-adapted baseline and outperforms other state-of-the-art methods in similar settings.
Comparison with State-of-the-Art: The proposed model surpasses existing methods (e.g., Truong et al., Zhang et al., Jemal et al.) in recall while requiring significantly less manual feature engineering than approaches relying on statistical moments or entropy characteristics.
Complexity: While EEG-FuseFormer offers the highest predictive accuracy, it incurs higher computational costs (12.40M trainable parameters) compared to ResNet-18 (11.17M) and 1D-CNN-LSTM (0.12M). Runtime analysis indicates that while feasible on high-end GPUs, the model requires careful consideration for edge deployment compared to lighter architectures.

Significance
The paper claims that EEG-FuseFormer provides a robust solution for seizure onset prediction by effectively fusing multi-scale features through a transformer mechanism. Its primary significance lies in demonstrating that feature fusion combined with target adaptation can bridge the gap between patient-specific accuracy and cross-patient generalizability. The study underscores that while deep fusion models offer superior performance, the selection of a specific architecture (CNN-LSTM, ResNet-18, or EEG-FuseFormer) must be guided by the specific constraints of the target application, balancing the need for high accuracy against computational resources and deployment environments (cloud vs. edge).