A Novel LSTM-based Data Synthesis Approach for Performance Improvement in Detecting Epileptic Seizures

Abstract

Bio-electrical time signals play a significant role in assisting non-invasive observational procedures in healthcare. These bioelectrical signals are weak signals with inherently low voltage and low frequency, hidden mostly under relatively large high-voltage noise signals. Hence it is extra challenging to analyze them. In modern clinical data analysis, these signals could be further analyzed using conventional machine learning (ML) methods. Also, in the recent past, two-dimensional spectrum-based classification, predominantly with Convolutional Neural Networks (CNN), has been tried with time-series data. One of the objectives of this study is to find which approach would suit better for biomedical signal analysis when data are scarce and signals are weak. Also, in bio-medical signal analysis data is scarce. Yet, to effectively train either an ML or a deep learning (DL) model, a sample clinical dataset of a significant size is required. Hence, the second objective of this research is to present a novel data synthesis method to address data scarcity. With these objectives, the study compares the performance of the time-series-based classification with traditional ML approaches, against the 2D spectrum-based classification for bio-electrical signal classification. For this purpose the study utilizes learning models; Multi-layer Perceptron (MLP), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long Short-Term Memory Networks (LSTMs), Auto Encoder (AE), and Convolutions Neural Network (CNN). Also, the authors propose a novel data synthesis method based on LSTMs to improve the sample size of the standard CHB-MIT Scalp EEG dataset. The results show that with the expanded dataset, the two-dimensional spectrum-based classification architecture was able to achieve a precision level of 85% at the classification. The conventional ML-based methods showed on average a precision level of 82%. In conclusion with the proposed virtual sample generation approach, 2D spectrum-based classification with Convolutional Neural Networks showed promising performances.