A protocol for trustworthy EEG decoding with neural networks DOI Creative Commons
Davide Borra, Elisa Magosso, Mirco Ravanelli

и другие.

Neural Networks, Год журнала: 2024, Номер 182, С. 106847 - 106847

Опубликована: Ноя. 2, 2024

Deep learning solutions have rapidly emerged for EEG decoding, achieving state-of-the-art performance on a variety of decoding tasks. Despite their high performance, existing do not fully address the challenge posed by introduction many hyperparameters, defining data pre-processing, network architecture, training, and augmentation. Automatic hyperparameter search is rarely performed limited to network-related hyperparameters. Moreover, pipelines are highly sensitive fluctuations due random initialization, hindering reliability. Here, we design comprehensive protocol that explores hyperparameters characterizing entire pipeline includes multi-seed initialization providing robust estimates. Our validated 9 datasets about motor imagery, P300, SSVEP, including 204 participants 26 recording sessions, different deep models. We accompany our with extensive experiments main aspects influencing it, such as number used search, split into sequential simpler searches (multi-step search), use informed vs. non-informed algorithms, seeds obtaining stable performance. The best included 2-step via an algorithm, final training evaluation using 10 initializations. optimal trade-off between computational time was achieved subset 3-5 search. consistently outperformed baseline pipelines, widely across models, could represent standard approach neuroscientists in trustworthy reliable way.

Язык: Английский

Graph-informed convolutional autoencoder to classify brain responses during sleep DOI Creative Commons
Sahar Zakeri, Somayeh Makouei, Sebelan Danishvar

и другие.

Frontiers in Neuroscience, Год журнала: 2025, Номер 19

Опубликована: Апрель 28, 2025

Automated machine-learning algorithms that analyze biomedical signals have been used to identify sleep patterns and health issues. However, their performance is often suboptimal, especially when dealing with imbalanced datasets. In this paper, we present a robust state (SlS) classification algorithm utilizing electroencephalogram (EEG) signals. To aim, pre-processed EEG recordings from 33 healthy subjects. Then, functional connectivity features recurrence quantification analysis were extracted sub-bands. The graphical representation was calculated phase locking value, coherence, phase-amplitude coupling. Statistical select p-values of less than 0.05. These compared between four states: wakefulness, non-rapid eye movement (NREM) sleep, rapid (REM) during presenting auditory stimuli, REM without stimuli. Eighteen types different stimuli including instrumental natural sounds presented participants REM. selected significant train novel deep-learning classifiers. We designed graph-informed convolutional autoencoder called GICA extract high-level the features. Furthermore, an attention layer based on rate EEGs incorporated into classifier enhance dynamic ability model. proposed model assessed by comparing it baseline systems in literature. accuracy SlS-GICA 99.92% feature set. This achievement could be considered real-time automatic applications develop new therapeutic strategies for sleep-related disorders.

Язык: Английский

Процитировано

0
A protocol for trustworthy EEG decoding with neural networks DOI Creative Commons
Davide Borra, Elisa Magosso, Mirco Ravanelli

и другие.

Neural Networks, Год журнала: 2024, Номер 182, С. 106847 - 106847

Опубликована: Ноя. 2, 2024

Deep learning solutions have rapidly emerged for EEG decoding, achieving state-of-the-art performance on a variety of decoding tasks. Despite their high performance, existing do not fully address the challenge posed by introduction many hyperparameters, defining data pre-processing, network architecture, training, and augmentation. Automatic hyperparameter search is rarely performed limited to network-related hyperparameters. Moreover, pipelines are highly sensitive fluctuations due random initialization, hindering reliability. Here, we design comprehensive protocol that explores hyperparameters characterizing entire pipeline includes multi-seed initialization providing robust estimates. Our validated 9 datasets about motor imagery, P300, SSVEP, including 204 participants 26 recording sessions, different deep models. We accompany our with extensive experiments main aspects influencing it, such as number used search, split into sequential simpler searches (multi-step search), use informed vs. non-informed algorithms, seeds obtaining stable performance. The best included 2-step via an algorithm, final training evaluation using 10 initializations. optimal trade-off between computational time was achieved subset 3-5 search. consistently outperformed baseline pipelines, widely across models, could represent standard approach neuroscientists in trustworthy reliable way.

Язык: Английский

Процитировано

0