The iterative multiblock framework using high-frequency EEG signals achieved 94% accuracy, 92% sensitivity, and 93% specificity in detecting neurological disorders.
Does an iterative multiblock framework using MS-CRNN and advanced signal processing improve the detection accuracy of neurological disorders from high frequency EEG signals?
An iterative multiblock framework using advanced signal processing and MS-CRNN provides highly accurate and interpretable detection of neurological disorders from high-frequency EEG signals.
It has become pertinent to develop early and accurate diagnosis tools for these neurological diseases, such as Alzheimer's and Parkinson's. The diagnosis may be high frequency electroencephalogram (EEG) signal based. These techniques promise good results but fail to obtain the desired clinically relevant features because of the intrinsically non-stationary and noisy nature of high frequency EEG components. Limitations of existing methods include suboptimal signal processing, ineffective strategies for feature selection, lack of robustness in feature fusion mechanisms, and limited explainability for clinical adoptions. This work, therefore, proposes a holistic framework in the context of clinical detection of neurological disorders using high frequency EEG signals which are enhanced as a pipeline of multi-blocks. The combination of Hilbert-Huang transform (HHT) with a modified empirical mode decomposition ensures that the decomposition is adaptive in nature and effective noise reduction leads to preprocessing of the data. Wavelet Packets transform (WPT) in conjunction with shannon entropy-based feature selection reduces the dimensions of the data without information loss, which aids in meaningful extraction of temporal and frequency domain features. Canonical correlation analysis with multi-view representation learning allows integration of EEG features along with clinical metadata as auxiliary information to create a common feature space for increased sensitivity in diagnosis. A new multi-scale convolutional recurrent neural network (MS-CRNN) uses an attention mechanism to process the combined features and find spatiotemporal dependencies while focusing on patterns that are important for diagnosis. The method is demonstrated through grad-cam and integrated gradient techniques that help in visualizing and quantitatively attributing feature extraction. This method was 94% accurate; 92% sensitive; and 93% specific when identifying issues early on. The high accuracy in making clinical interpretation and diagnosis has set a new bar for clinicians and has encouraged public policy to support early intervention.
Agrawal et al. (Thu,) conducted a other in Neurological disorders (Alzheimer's and Parkinson's disease). Iterative multiblock framework (MS-CRNN) was evaluated on Diagnostic accuracy. The iterative multiblock framework using high-frequency EEG signals achieved 94% accuracy, 92% sensitivity, and 93% specificity in detecting neurological disorders.