A Tensor PCA to Derive a Spectrum Domain LSTM Architecture for Mild Cognitive Impairment Detection by EEG

The Electroencephalogram (EEG) is a cost-effective and highly sensitive biomarker for accurate and timely diagnosis of Mild Cognitive Impairment (MCI) and essential for halting the progression of dementia. Due to high dimensionality, non-stationarity and nonlinearity of the EEG signal, as well as the influence of a large number of background waveforms and artifacts, the automatic detection of MCI with this technique is a challenging problem. Among various Deep Learning (DL) techniques that have been proposed during the latest years to address MCI detection from EEG signals, architectures based on Long Short-Term Memory (LSTM) networks, are the best suited to classify sequential data like EEG. However, in order to obtain the best performance with such networks is of paramount importance to reduce the dimensionality of data. The aim of this chapter is to derive an architecture called Spectrum-Domain LSTM (SD-LSTM), in which a tensor PCA transforms pre-processed data in high-dimensional time-domain to data in low-dimensional spectrum-domain. The core of the proposed approach is the development of a closed-form tensor PCA that is able to overcome the limitations of techniques commonly used for dimensionality reduction of tensor data, thus enabling the implementation of a high-performance low-complexity SD-LSTM architecture for EEG signal classification. Several experimental results showing the performance achieved with the SD-LSTM architecture prove the relevance of the proposed approach in solving typical MCI detection problems by EEG signal.