Pipeline to detect spike-and-wave EEG patterns based on polynomial regression modeling and Taylor series feature selection

Abstract

Epilepsy is a common neurological disorder diagnosed and monitored through EEG recordings. Accurate spike-and-wave (SW) pattern classification is crucial for distinguishing this epileptic seizure disorder from normal brain wave activity (NW). However, mathematically modeling SW remains challenging, affecting classification accuracy. This study proposes a pipeline in two stages combining polynomial regression techniques, and data processing, in a machine-learning classification scheme. At the first stage of decision-making, the idea is to create a generalized waveform mother that represents all the waveforms of the EEG patterns, such as SW and NW. This waveform is derived from a polynomial regression model that is assessed by the truncation error of the Taylor series. In the second stage, a feature selection algorithm based on a vector that includes the coefficients from Taylor and the statistical properties of the SW and NW waveforms was designed for the machine learning classifier. This algorithm uses the confidence interval to extract the Taylor series points that do not represent the generalized mother equation. This yields a dimensional reduction of this vector, which can be used in a classification and detection scheme. Three polynomial regression models, such as Fourier, Gaussian, and sums-of-sines were evaluated using the pipeline methodology. The best model was the Fourier regression, which achieved an accuracy of 96.2% using the SVM classifier with a Gaussian kernel to detect spike-and-wave patterns.