(A) 95% highly accurate EEG-connectome processor for mental health monitoring systems

Abstract

An electroencephalography (EEG)-connectome processor for mental health monitoring systems is proposed. From 19-channel EEG signals, the proposed processor extracts features and they are used to determine whether the mental state is healthy or not. As it deals with the mental health, the most important thing is high accuracy. Connectome is known as the most accurate approach for complicated mental health because it contains information about inter-channel connectivity. However, it requires more memory and computational costs. As a result, the proposed processor is implemented to calculate connectome with memory reduction and computational cost reduction schemes. Before calculating connectome, EEG reconstruction has to be proceeded. Reconstruction optimizer (ReOpt) block compensates the reconstruction parameters named embedding delay and embedding dimension. Optimal selection of delay leads to the highest accuracy up to 95%, and optimal selection of dimension averagely halves the computational cost. Synchronization likelihood extractor (SLE) block calculates synchronization likelihoods (SLs) as a connectome feature. Sparse matrix inscription (SMI) scheme is proposed in SLE to reduce the required memory size to 1/24. Totally 171 SLs are calculated as the outputs of SLE. Small world feature extractor (SWFE) block converts those 171 SLs into 6 small world features to reduce the data dimension. 171 SLs or 6 small world features become the input vectors for the classifier, radial basis function (RBF) kernel-based support vector machine (SVM). Look-up-tables (LUTs) are adopted to replace the floating-point operations in SWFE and RBF, resulting in 54% decrease of required operations. The validity of the proposed processor is verified with 19-channel EEG measurement for controls and Alzheimer’s disease (AD) patients. Thanks to SLE, the proposed processor achieves 95% of diagnosis accuracy. The proposed processor occupies $3.8 mm^2$, consumes 1.71mW, and has 0.6 seconds of latency in $0.18\mum$ CMOS technology.