Investigating the Use of Raw EEG Data Augmentation Methods in Neuropsychiatric Disorder Diagnosis

1670 

Abstract Submission 

Charles Ellis¹, Robyn Miller¹, Vince Calhoun²

¹Center for Translational Research in Neuroimaging and Data Science (TReNDS), Atlanta, GA, ²Center for Translational Research in Neuroimaging and Data Science (TReNDS), Decatur, GA

Charles Ellis  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Robyn Miller, PhD  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Vince Calhoun  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Decatur, GA

While the application of deep learning methods to raw electroencephalogram (EEG) data is becoming more common, the development of reliable models can be challenging due to the small size of many EEG datasets. One potential solution is data augmentation, wherein synthetic training data is generated by duplicating and modifying the training data. While multiple data augmentation approaches have been developed for raw EEG data, their utility for the diagnosis of neuropsychiatric disorders remains underexplored. Moreover, it is unclear if existing approaches boost model performance because they increase the number of training samples or because they introduce augmentations that help the model learn better representations. In this study, we train a baseline convolutional neural network for automated major depressive disorder (MDD) diagnosis, train a second baseline model with duplicate training data (i.e., two identical copies), and compare performance to 6 models that are each trained with a different data augmentation approach that doubles the training set size (i.e., original data + augmented data).

We used a publicly available resting-state EEG dataset (Mumtaz et al., 2017) with 5-10-minute recordings from 30 individuals with MDD and 28 healthy controls (HCs). We trained our model with 19 channels: Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2. We downsampled the data to 200 Hz, channel-wise z-scored each recording separately, and separated the data into 25-second samples with a sliding window approach (2.5-second step size) that yielded 2,942 MDD samples and 2,950 HC samples. We trained 8 models: a baseline model with no duplicate data (M1.1), a baseline model with 2 copies of identical unmodified training data (M1.2), 6 models trained with 1 copy of unmodified training data and a copy of data augmented with Gaussian noise (M2) (Ellis et al., 2022), time reverse (M3) (Rommel et al., 2022), smooth time masking (M4) (Mohsenvand et al., 2020), Fourier surrogate (M5) (Schwabedal et al., 2019), frequency shift (M6) (Rommel et al., 2022), or channel dropout (M7) (Saeed et al., 2021) augmentation. For M1.1, we began with an architecture from (Oh et al., 2019) and optimized it with the Hyperband algorithm in Keras-Tuner. We used M1.1 as a starting point for M1.2-7 and optimized the learning rate of each model with Hyperband and the data augmentation parameters. We trained for 25 folds with subject-wise cross-validation and tested performance for each fold via accuracy, balanced accuracy, sensitivity, and specificity. We lastly compared the performance of each model with pair-wise t-tests and false discovery rate correction.

Table 1 shows our model test performance, which is fairly high. Most data augmentation approaches improved model performance by 2-2.5% compared to M1.1 but not M1.2. Also, our statistical testing did not identify any significant changes in accuracy, balanced accuracy, and sensitivity and only a couple significant changes in specificity.

Our results suggest that the benefits of existing EEG data augmentation approaches may mainly result from increased training set size and not necessarily from the introduction of helpful augmentations. As such, it may be beneficial to compare new augmentation methods to a baseline model with duplicate training data to see if their augmentations are helpful. It is also possible that future studies might benefit from an "augmentation by duplication" approach in which models are simply trained on two identical copies of the training data. We hope that our findings will provide helpful guidance as the field seeks to develop more robust models on small EEG datasets. Moreover, we would also like to stress that absolute rules cannot be derived from our findings. It is possible that other architectures or datasets might yield different conclusions on the efficacy of the data augmentation approaches we examined.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ²

EEG/MEG Modeling and Analysis ¹

Electroencephaolography (EEG)

Machine Learning

Psychiatric Disorders

Other - Data Augmentation