Source-linked AI summary
Learning Robust Features using Deep Learning for Automatic Seizure Detection
Pierre Thodoroff, Joelle Pineau, Andrew Lim
TL;DR
Automatic seizure detection is difficult because EEG seizure manifestations vary within and between patients, while manual analysis is laborious and specialist-dependent. The paper combines image-based EEG representations with a recurrent convolutional neural network that captures spectral, temporal, and spatial information. It reports improved cross-patient detection, robustness to missing channels and electrode montages, and state-of-the-art patient-specific performance.
Problem
Manual EEG analysis is laborious and seizure manifestations vary substantially within and between patients, limiting reliable automatic detection across patients.
Method
A recurrent convolutional neural network processes image-based EEG representations to learn robust features from spectral, temporal, and spatial information, supported by sample-efficient training.
Results
The model exceeds previous cross-patient results, achieving 85% average sensitivity versus 67% for REVEAL and reducing false positive rate from 1.7/hours to 0.8/hours.
Takeaways & Limitations
The model reaches state-of-the-art patient-specific performance and remains robust to missing channels and different electrode montages.
Takeaways & Limitations
Sensitivity was low for four patients, and cross-patient false positive rates remained higher than for patient-specific detection.
Abstract
from arXiv · showhide
We present and evaluate the capacity of a deep neural network to learn robust features from EEG to automatically detect seizures. This is a challenging problem because seizure manifestations on EEG are extremely variable both inter- and intra-patient. By simultaneously capturing spectral, temporal and spatial information our recurrent convolutional neural network learns a general spatially invariant representation of a seizure. The proposed approach exceeds significantly previous results obtained on cross-patient classifiers both in terms of sensitivity and false positive rate. Furthermore, our model proves to be robust to missing channel and variable electrode montage.
1. Introduction
Automatic seizure detection is motivated by the laborious, specialist-dependent analysis of long EEG recordings. Because seizure EEG patterns vary within and between patients, the paper proposes deep learning to learn robust, spatially invariant representations and reports stronger cross-patient performance.
- Neurologists must analyze hours-long EEG recordings manually, creating a laborious process that depends on scarce specialists.
- Seizure manifestations vary substantially both within individual patients over time and between different patients.
- The proposed recurrent convolutional neural network jointly captures spectral, temporal, and spatial information while learning a spatially invariant seizure representation.
- On cross-patient detection, the approach exceeds previous results, while matching state-of-the-art sensitivity and false positive rate on patient-specific detection.
2. Problem Definition
The study addresses automatic offline seizure detection in EEG, where recordings must support diagnosis, monitoring, and treatment planning. It evaluates patient-specific and cross-patient detection on a diverse CHB-MIT dataset using labeled 30-second segments.
- EEG is the primary diagnostic tool, but detecting and classifying seizures in hours-to-days recordings requires laborious visual analysis by trained neurologists.
- Automated detectors can support patient monitoring, diagnosis, and treatment planning, especially where expert neurologists are unavailable.
- Patient-specific detectors use annotated data from the same patient, whereas cross-patient detectors operate without patient-specific training data.
- The CHB-MIT dataset contains 969 hours of scalp EEG from 23 patients, 24 cases, and 173 seizures spanning several seizure types.
- The task is to determine whether each annotated 30-second EEG segment contains a seizure.
3. Previous Work
Prior work distinguishes offline seizure detection from seizure prediction and commonly uses handcrafted EEG features. Patient-specific detection performs well, whereas cross-patient generalization is more difficult because seizure manifestations differ across patients.
- Offline detection labels completed recordings, whereas online prediction aims to anticipate seizures before they occur.
- Sensitivity measures correctly identified seizures, while false detection rate counts false alarms per hour; effective systems must balance both metrics.
- Cross-patient detection is harder because seizure locations, shapes, and durations can vary dramatically between patients.
4. Methods
The method converts multichannel EEG into an image-based representation and processes it with a jointly trained recurrent convolutional network. Convolution captures spatially robust features, while recurrent components model temporal sequences for patient-specific and cross-patient detection.
- 4. Methods: The pipeline projects multichannel EEG into images, then predicts whether each image contains a seizure using a recurrent convolutional neural network.
- 4.1 Image based representation: Electrode coordinates are projected from 3D onto a 2D surface, and three image channels encode magnitudes from the 0-7, 7-14, and 14-49 Hertz bands.
- 4.2 Recurrent convolutional neural network: Convolutional layers extract representations robust to spatial translation, replacing complex handcrafted feature engineering.
- 4.2 Recurrent convolutional neural network: LSTM units preserve information across time, with bidirectional processing using both chronological and reverse chronological sequences.
- 4.2 Recurrent convolutional neural network: Each 1-second image produces a 64-dimensional output vector, and the recurrent network processes 30 such blocks corresponding to 30 seconds of EEG.
- 4.3 Training neural network in a sample-efficient manner: Training addresses scarce and imbalanced data through negative-sample subsampling, convolutional pre-training, transfer learning, and prediction averaging across three initializations.
5. Results on the CHB-MIT dataset
On CHB-MIT, the neural detector was evaluated against established patient-specific and cross-patient baselines. It matched patient-specific performance while substantially improving cross-patient sensitivity and false positive rate.
- Evaluation setup: The evaluation compares the neural patient-specific detector with Shoeb’s detector and the cross-patient model with REVEAL.The study also notes that benchmarking is complicated by differing settings and disagreement among experts about seizure definitions.
- Patient-specific detection: Patient-specific results were broadly similar for sensitivity and false positive rate between the neural model and Shoeb’s SVM detector.Both methods achieved accuracy considered suitable for clinical use in patient-specific classification.
- Patient-specific detection: The neural model was significantly more robust to missing channels than the SVM with expert features.This robustness was observed in the missing-channel comparison.
- Cross-patient detection: 85% average sensitivity exceeded REVEAL’s 67% in cross-patient detection.The comparison used the REVEAL algorithm applied to the CHB-MIT dataset.
- Cross-patient detection: False positive rate decreased from 1.7/hours with REVEAL to 0.8/hours with the proposed model.This result was reported for the cross-patient comparison.
6. Discussion
The model learns a seizure representation from spatial, temporal, and frequency information, with clinical advantages including electrode-montage flexibility and seizure localization. Cross-patient performance remains limited by low sensitivity for some patients and high prediction variance on small datasets.
- The neural model learns robust seizure features from spatial, temporal, and frequency information.
- Sliding-window occlusion identifies brain regions critical to classification, indicating seizure activity around left parietal, frontal, or temporal lobes.The image-based representation also supports classification across different electrode montages.
- Cross-patient sensitivity was low for four patients, likely because their seizure patterns differed from the training data.The authors highlight the need for more data.
- Small parameter changes can substantially affect false positive rate when training uses few negative samples and testing uses many more.The authors describe this as an inherent problem when deep-learning architectures are applied to small datasets.