MS-HyFS: A novel multiscale hybrid framework with Scalable electrodes for motor imagery classification

Abstract

Hybrid deep neural networks have been developed to enrich features in spatial–temporal domains from electroencephalogram (EEG) based on motor imagery (MI) classification. However, these networks primarily focus on forming a subject-independent model that disregards individual variations/difference caused by various reaction time associated with fixed time window and insufficient spatial information due to different brain functional connectivity. Additionally, analyzing such complex networks may incur significant computational costs. This study proposes a novel Multiscale Hybrid Framework with Scalable electrodes (MS-HyFS), which includes of a multiscale filter bank CSP (MS-FBCSP) algorithm to deal with fixed time window issue by extracting multiscale CSP features, followed by a combination of a multiscale Hybrid network with a 1D-CNN and LSTM (MS-HyCaL) to enrich the spatial–temporal features from local and global perspectives. We reduce the computational costs by selecting critical electrodes based on the brain’s asymmetric properties and neural activity areas. MS-HyFS was evaluated across two publicly available EEG datasets [BCIIV-2a and BCIIV-2b]. These are divided into training and test datasets using an 8:2 ratio, and the training data are further divided into training and validation sets using a fivefold cross-validation (CV) method, in which the model with the highest accuracy among the five was selected. The model is trained once more with the full training set, and the test data were then used to evaluate its performance. This approach achieved average classification accuracies of 84.3% and 64.0% for the BCIIV-2a and BCIIV-2b datasets, respectively. Experimental results showed MS-HyFS was superior to state of-the-art algorithms.