Parity FuseNet-Mi: Learning Spatio-Temporal Dynamics via Hierarchical Odd-Even Kernel Convolutions for Motor Imagery Decoding

At the forefront of contemporary developments is the development of brain-computer interface (BCI) technology, which has enormous potential for a wide range of applications. The ability to recognize motor imagery (MI) dynamics in electroencephalogram (EEG) stream is a pivotal component of this advancement. This technology is essential to improve human abilities, empower control over devices and surroundings, and help people with disabilities. Although promising, the limited accuracy of brain signal interpretation has prevented BCIs from being widely used in real-world scenarios. In order to classify EEG based MI tasks, this paper presents a novel model that uses odd and even kernel convolution for feature extraction. The independence of this model from artifact removal techniques and minimal preprocessing requirements are two of its main advantages. Using cross-validation with leave-one-subject-out (LOSO), proposed model was evaluated on three standard datasets. BCI contest IV-2a, IV-2b and III-IIIb. The accuracy results were 76. 80%, 78.34%, and 76. 21%, respectively. These findings demonstrate the strong performance of the model and the promise to further the study of MI decoding.