Source-linked AI summary
An Automated System for Epilepsy Detection using EEG Brain Signals based on Deep Learning Approach
Ihsan Ullah, Muhammad Hussain, Emad-ul-Haq Qazi, Hatim Aboalsamh
TL;DR
Manual EEG inspection is laborious, and existing approaches face limited data and weaker performance on ternary epilepsy classification. The paper introduces an augmented ensemble of pyramidal 1D-CNN models with majority voting, achieving high accuracy across Bonn-dataset epilepsy-detection cases.
Problem
Limited EEG data and the difficulty of ternary classification motivate an automatic approach beyond hand-engineered traditional methods.
Method
The system uses augmented end-to-end P-1D-CNN models whose local decisions are fused by majority voting across fixed-size EEG windows.
Results
99.1±0.9% accuracy was reported in almost all epilepsy-detection cases on the University of Bonn dataset.
Takeaways & Limitations
The system is designed to assist neurologists while using fewer learnable parameters and limited data.
Abstract
from arXiv · showhide
Epilepsy is a neurological disorder and for its detection, encephalography (EEG) is a commonly used clinical approach. Manual inspection of EEG brain signals is a time-consuming and laborious process, which puts heavy burden on neurologists and affects their performance. Several automatic techniques have been proposed using traditional approaches to assist neurologists in detecting binary epilepsy scenarios e.g. seizure vs. non-seizure or normal vs. ictal. These methods do not perform well when classifying ternary case e.g. ictal vs. normal vs. inter-ictal; the maximum accuracy for this case by the state-of-the-art-methods is 97+-1%. To overcome this problem, we propose a system based on deep learning, which is an ensemble of pyramidal one-dimensional convolutional neural network (P-1D-CNN) models. In a CNN model, the bottleneck is the large number of learnable parameters. P-1D-CNN works on the concept of refinement approach and it results in 60% fewer parameters compared to traditional CNN models. Further to overcome the limitations of small amount of data, we proposed augmentation schemes for learning P-1D-CNN model. In almost all the cases concerning epilepsy detection, the proposed system gives an accuracy of 99.1+-0.9% on the University of Bonn dataset.
1. Introduction
Epilepsy EEG analysis is difficult because manual review is laborious, data are limited and noisy, and existing methods struggle with ternary classification. The paper proposes an augmented deep-learning ensemble to address these challenges.
- Manual EEG inspection requires substantial time and expert effort, burdening neurologists and reducing their efficiency.
- Limited training data, noise, artifacts, and patient-to-patient seizure-morphology variation complicate automatic recognition of normal, ictal, and inter-ictal signals.
- Traditional methods may classify binary seizure tasks accurately but perform poorly on normal-versus-ictal-versus-inter-ictal classification.
- Deep learning can automatically encode hierarchical EEG features and avoid selecting hand-engineered feature extractors and feature subsets.
- The proposed system combines P-1D-CNN models, augmentation schemes, and majority voting for binary and ternary EEG classification.
2. Literature Review
Prior epilepsy-detection systems largely rely on hand-crafted EEG features and traditional classifiers, with performance varying across binary and ternary tasks. The paper motivates deep learning as a data-adaptive alternative despite limited epilepsy data.
- Related methods classify epileptic and non-epileptic EEG signals using extracted discriminatory features and classifiers.
- Hand-crafted features such as Shannon entropy, standard deviation, and energy achieved high accuracy on selected Bonn-dataset tasks but lower accuracy on several other cases.
- The reviewed literature covers both binary epileptic-versus-non-epileptic and ternary ictal-versus-normal-versus-interictal scenarios.
- Most existing feature extractors are hand-crafted and not adapted to the data, while the limited dataset makes training deep models challenging.
3. The Proposed System
The proposed system segments EEG signals, classifies sub-signals with an ensemble of pyramidal one-dimensional CNNs, and fuses local decisions by majority vote. Its pyramid design reduces parameters while augmentation addresses limited training data.
- The Proposed System: The system splits EEG input into fixed-size overlapping windows, classifies each sub-signal with a P-1D-CNN model, and fuses local decisions by majority vote.
- The Proposed System: Overlapping windows are treated as independent training instances, and copies of trained P-1D-CNN models act as experts for different signal parts.
- P-1D-CNN Architecture: P-1D-CNN is an end-to-end architecture whose core layers include convolution, batch normalization, ReLU activation, fully connected, and dropout layers.
- P-1D-CNN Architecture: The model uses larger convolution strides instead of pooling to reduce redundant features while learning a hierarchy from low-level to high-level representations.
- P-1D-CNN Architecture: The final fully connected layer uses two or three nodes depending on whether the task is binary or ternary classification.
- P-1D-CNN Architecture: Convolutional kernels produce channels that extract different discriminative features, with the proposed model reducing kernels by 33% in subsequent layers.
- P-1D-CNN Architecture: Figure 3 displays the input signal and activations from Conv1, Conv2, and Conv3, using 24, 16, and 08 channels, respectively.
4. Model Selection and Parameter Tuning
The study uses the University of Bonn EEG dataset and evaluates data augmentation for training deep CNN models. Model selection and augmentation are examined through multiple 1D-CNN configurations.
- Experimental Design: The section introduces evaluation measures and identifies the best augmentation scheme and 1D-CNN model through comparative experiments.The paper organizes these procedures across data description, augmentation, evaluation, training, and model selection.
- Dataset: The University of Bonn dataset contains five sets, A–E, with 100 one-channel EEG instances per set.Sets A and B come from healthy volunteers, while sets C–E come from patients.
- Data Augmentation: The dataset is too small to train a deep model reliably, making augmentation necessary for improving generalization.The authors also note that acquiring and expert-labeling many additional EEG signals is impractical.
- Data Augmentation: Each record contains 4097 samples, and sliding-window segmentation is used to generate multiple EEG instances.A 512-sample window with stride 480 produces eight sub-signals per record in the described approach.
Scheme-1
Scheme-1 augments training data with heavily overlapping sliding windows and evaluates test signals through an ensemble of local predictions fused by majority vote.
- Training augmentation: Training signals of length 4097 are divided with a 512-sample window and stride 64, creating 57 sub-signals per signal.The windows overlap by 87.5%, producing 5130 training instances for each class.
- Testing procedure: Testing signals are divided into four 1024-sample sub-signals before further segmentation into 512-sample windows with 50% overlap.Each 1024-sample test segment yields three 512-sample instances.
- Ensemble decision: Three trained P-1D-CNN models independently analyze local test segments, and majority vote produces the global decision.Each ensemble model acts as an expert for a local part of the signal instance.
Scheme-2
Scheme-2 uses less-overlapping training windows and more-overlapping test windows, while model selection compares CNN architectures using cross-validation and multiple performance metrics.
- Scheme-2 augmentation: Scheme-2 uses a 512-sample training window with stride 128, corresponding to 25% overlap.For testing, 1024-sample segments are divided using 512-sample windows with 75% overlap.
- Scheme-2 ensemble: Five trained P-1D-CNN models analyze the Scheme-2 test segments, with majority vote used for fusion.The five local instances are each 512 samples long.
- Evaluation: Ten-fold cross-validation trains on 90% of signals and tests on the remaining 10% across accuracy, specificity, sensitivity, precision, f-measure, and g-mean.The average performance is calculated across the ten folds.
- Model selection: The eight candidate CNN models include traditional architectures with increasing filters and pyramid architectures that reduce filters by 33% with depth.Pyramid models M5–M8 use a coarse-to-fine refinement structure.
- Results: 96.45±0.13 average accuracy with augmentation scheme 1 exceeds 95.40±0.35 with scheme 2 across the evaluated models.Scheme 1 is therefore adopted for subsequent experiments.
- Results: Pyramid models perform better than or equal to traditional models, and M5 generally gives the best results with the fewest parameters.M5 uses dropout 0.5 and 20 fully connected neurons and is selected for later experiments.
5. Results and Discussion
Across ternary and binary epilepsy-detection experiments, the ensemble of P-1D-CNN models generally outperformed single models and reported state-of-the-art comparisons, with strongest results on the University of Bonn dataset.
- Experiment 1: 99.1% mean accuracy was achieved for normal vs interictal vs ictal classification with an ensemble of three P-1D-CNN models, outperforming Zhang et al. by 1.7%.A single P-1D-CNN model achieved 96.1% mean accuracy; the main confusion involved normal versus interictal or interictal versus ictal signals.
- Experiment 2: 99.8% mean accuracy was obtained for normal (AB) versus non-seizure epileptic (CD) classification using the ensemble.The single-model result was 98.2%, while mean sensitivity and specificity were 98% and 99%, respectively.
- Experiment 2: 99.95% mean accuracy was obtained for normal (AB) versus non-seizure and seizure epileptic (CDE) classification using the ensemble.The single-model mean accuracy was 98.1%, and both mean sensitivity and specificity were 98%.
- Experiment 3: 100% to 98.5% ensemble mean accuracies were reported across six normal-or-non-seizure versus seizure binary problems.For normal-versus-seizure problems, ensemble accuracy was almost 100%; C versus E reached 98.5%, and other non-seizure-versus-seizure problems reached 99.3% and 99.7%.
- Comparison with State-of-the-art Methods: 99.6% mean accuracy was reported across all sixteen binary and ternary cases in comparison with state-of-the-art methods.The authors describe these results as outperforming reported methods and indicating generalization of the ensemble system.
6. Conclusion
The paper concludes with an automatic EEG epilepsy-detection system based on memory-efficient pyramidal one-dimensional CNN ensembles and data augmentation. It reports 99.1±0.9% accuracy in almost all cases on the University of Bonn dataset, while identifying pre-seizure detection as future work.
- Conclusion: The proposed system addresses binary and ternary epilepsy-detection problems using an end-to-end pyramidal one-dimensional CNN model.The binary tasks include epileptic versus non-epileptic and seizure versus non-seizure; the ternary task includes ictal, normal, and interictal signals.
- Conclusion: Two data augmentation schemes were introduced to train the P-1D-CNN models despite the small dataset.The system uses an ensemble whose local decisions are fused by majority vote.
- Conclusion: 99.1±0.9% accuracy was reported in almost all epilepsy-detection cases on the University of Bonn dataset.The authors state that the system can be useful for similar EEG brain-signal classification problems.
- Conclusion: Future work will investigate detecting seizures prior to their occurrence rather than after occurrence.The paper characterizes pre-seizure detection as a challenging problem.