Modeling and Optimization of Signals Using Machine Learning Techniques

16Kalman Filter-Based Seizure Prediction Using Concatenated Serial-Parallel Block Technique

Purnima P. S.^1* and Suresh M.²

¹SSAHE, Agalakote, Tumakuru, Karnataka, India

²Department of ECE, SSIT, Tumakuru, Karnataka, India

Abstract

Present state-of-the-art seizure prediction approaches use machine learning techniques and neural networks, which include K-means clustering, discriminant analysis, radial basis function neural networks (RBFNN), and backpropagation neural networks for efficient results, but most of them are not real-time approaches. We propose to implement an alternate real-time approach based on the Kalman filtering of electroencephalogram (EEG) time-series signal with a novel ‘serial-parallel block concatenation’ method, which predicts multiple signal samples ahead of time, and hence allows real-time processing. We also propose a novel probability density function-based (PDF-based) method to estimate signal data points for the adaptive filter updates. This technique applies to 250 different EEG records, each lasting 23.6 seconds. The receiver operating characteristic (ROC) analysis is used for evaluating the performance of the proposed recognition system. The proposed work achieves approximately a 94% success rate. The proposed concatenation approach, along with the PDF method, gave more accurate real-time predictions than a standard adaptive filter.

Keywords: Kalman filter, seizure detection, EEG signal, series-parallel concatenation

16.1 Introduction

Brain abnormal ...