Data augmentation for deep-learning-based electroencephalography

Abstract

Highlights • Data augmentation (DA) is increasingly used with deep learning (DL) on EEG • It enhances decoding accuracy left unexplained by 29% on average on the datasets we review • We analyze which specific DA techniques appear to work best for which EEG tasks • We tested various DA techniques on an open motor-imagery task and compared the accuracy gains to demonstrate the usefulness of DA for DL-based EEG analysis • We propose guidelines for reporting parameters for different DA techniques Abstract-Background Data augmentation (DA) has recently been demonstrated to achieve considerable performance gains for deep learning (DL)-increased accuracy and stability and reduced overfitting. Some electroencephalography (EEG) tasks suffer from low samples-to-features ratio, severely reducing DL effectiveness. DA with DL thus holds transformative promise for EEG processing, possibly like DL revolutionized computer vision, etc.-New method We review trends and approaches to DA for DL in EEG to address: Which DA approaches exist and are common for which EEG tasks? What input features are used? And, what kind of accuracy gain can be expected?-Results DA for DL on EEG begun 5 years ago and is steadily used more. We grouped DA techniques (noise addition, generative adversarial networks, sliding windows, sampling, Fourier transform, recombination of segmentation, and others) and EEG tasks (into seizure detection, sleep stages, motor imagery, mental workload, emotion recognition, motor tasks, and visual tasks). DA efficacy across techniques varied considerably. Noise addition and sliding windows provided the highest accuracy boost; mental workload most benefitted from DA. Sliding window, noise addition, and sampling methods most common for seizure detection, mental workload, and sleep stages, respectively.-Comparing with existing methods Percent of decoding accuracy explained by DA beyond unaugmented accuracy varied between 8% for recombination of segmentation and 36% for noise addition and from 14% for motor imagery to 56% for mental workload-29% on average.-Conclusions DA increasingly used and considerably improved DL decoding accuracy on EEG. Additional publications-if adhering to our reporting guidelines-will facilitate more detailed analysis.

Data Augmentation: Using Channel-Level Recombination to Improve Classification Performance for Motor Imagery EEG

“…We only used one DA method in each experiment instead of experimenting with multiple DA methods' additive effects. Two DA methods (noiseadded and flipping) were implemented as reference methods (Lashgari et al, 2020). The flipping DA method flipped each sample along the time axis to generate a new sample.…”

“…Methods like GANs and VAEs having massive parameters to be learned are not suitable for this situation. Moreover, considering that the no-parameter (or few parameters) methods will often achieve better results for this situation, we selected noise-added and flipping methods for comparative investigation (Lashgari et al, 2020). Concerning the noise-added method, a Gaussian noise matrix with SNR (signal to noise ratio) of 5 is calculated, and then this noise matrix is added to the original sample.…”

A New Deep Stacked Architecture for Multi-Fault Machinery Identification With Imbalanced Samples

“…• Sliding Window [50]: this model is mainly based on sliding a window of a specified size over the sample and applying some modification while sliding to produce different samples. • Fourier transformation [51]: this is a popular method for augmenting signals. It is focused on imbalanced datasets in transitional sleep stages.…”

Evaluation of a Sliding Window mechanism as DataAugmentation over Emotion Detection on Speech

“…Overfitting limits the generalization of the algorithm. Some techniques used during the construction of the neural network algorithm can avoid this problem, one of which is Data Augmentation, which is related to the generation of new samples from the existing data, in order to increase the size of the dataset [16]. Thus, the neural network algorithm receives a larger amount of data, both for training and for testing, and consequently, its generalization capacity will be greater, since it will be able to recognize more data samples.…”