Robust learning from corrupted EEG with dynamic spatial filtering

NeuroImage, Journal Year: 2022, Volume and Issue: 251, P. 118994 - 118994

Published: Feb. 16, 2022

Building machine learning models using EEG recorded outside of the laboratory setting requires methods robust to noisy data and randomly missing channels. This need is particularly great when working with sparse montages (1-6 channels), often encountered in consumer-grade or mobile devices. Neither classical nor deep neural networks trained end-to-end on are typically designed tested for robustness corruption, especially While some studies have proposed strategies channels, these approaches not practical used computing power limited (e.g., wearables, cell phones). To tackle this problem, we propose dynamic spatial filtering (DSF), a multi-head attention module that can be plugged before first layer network handle channels by focus good ignore bad ones. We DSF public encompassing ∼4000 recordings simulated channel corruption private dataset ∼100 at-home natural corruption. Our approach achieves same performance as baseline no noise applied, but outperforms baselines much 29.4% accuracy significant present. Moreover, outputs interpretable, making it possible monitor effective importance real-time. has potential enable analysis challenging settings where hampers reading brain signals.

Language: Английский

---

MP-SeizNet: A multi-path CNN Bi-LSTM Network for seizure-type classification using EEG

Biomedical Signal Processing and Control, Journal Year: 2023, Volume and Issue: 84, P. 104780 - 104780

Published: March 7, 2023

Language: Английский

---

Automated EEG pathology detection based on different convolutional neural network models: Deep learning approach

Computers in Biology and Medicine, Journal Year: 2021, Volume and Issue: 133, P. 104434 - 104434

Published: April 25, 2021

Language: Английский

---

Automatic Detection of Abnormal EEG Signals Using WaveNet and LSTM

Sensors, Journal Year: 2023, Volume and Issue: 23(13), P. 5960 - 5960

Published: June 27, 2023

Neurological disorders have an extreme impact on global health, affecting estimated one billion individuals worldwide. According to the World Health Organization (WHO), these neurological contribute approximately six million deaths annually, representing a significant burden. Early and accurate identification of brain pathological features in electroencephalogram (EEG) recordings is crucial for diagnosis management disorders. However, manual evaluation EEG not only time-consuming but also requires specialized skills. This problem exacerbated by scarcity trained neurologists healthcare sector, especially low- middle-income countries. These factors emphasize necessity automated diagnostic processes. With advancement machine learning algorithms, there great interest automating process early diagnoses using EEGs. Therefore, this paper presents novel deep model consisting two distinct paths, WaveNet-Long Short-Term Memory (LSTM) LSTM, automatic detection abnormal raw data. Through multiple ablation experiments, we demonstrated effectiveness importance all parts our proposed model. The performance was evaluated TUH Corpus V.2.0.0. (TUAB) achieved high classification accuracy 88.76%, which higher than existing state-of-the-art research studies. Moreover, generalization evaluating it another independent dataset, TUEP, without any hyperparameter tuning or adjustment. obtained 97.45% between normal recordings, confirming robustness

Language: Английский

---

An extended clinical EEG dataset with 15,300 automatically labelled recordings for pathology decoding

NeuroImage Clinical, Journal Year: 2023, Volume and Issue: 39, P. 103482 - 103482

Published: Jan. 1, 2023

Automated clinical EEG analysis using machine learning (ML) methods is a growing research area. Previous studies on binary pathology decoding have mainly used the Temple University Hospital (TUH) Abnormal Corpus (TUAB) which contains approximately 3,000 manually labelled recordings. To evaluate and eventually even improve generalisation performance of for pathology, larger, publicly available datasets required. A number addressed automatic labelling large open-source as an approach to create new decoding, but little known about extent training automatically dataset affects performances established deep neural networks. In this study, we created additional labels (TUEG) based medical reports rule-based text classifier. We generated 15,300 newly recordings, call TUH Expansion (TUABEX), five times larger than TUAB. Since TUABEX more pathological (75%) non-pathological (25%) then selected balanced subset 8,879 Balanced (TUABEXB). investigate how networks, applied four convolutional networks (ConvNets) task versus classification compared each architecture after different datasets. The results show that TUABEXB rather TUAB increases accuracies itself some architectures. argue can be efficiently utilise massive amount data stored in archives. make proposed open source thus offer research.

Language: Английский

---