Modeling Neonatal EEG Using Multi-Output Gaussian Processes

Victor Caro,Jou-Hui Ho,Felipe Tobar,Scarlet Witting

doi:10.1109/access.2022.3159653

Victor Caro, Jou-Hui Ho + Show 2 more

Open Access

https://doi.org/10.1109/access.2022.3159653

Copy DOI

Abstract

Neonatal seizures are sudden events in brain activity with detrimental effects in neurological functions usually related to epileptic fits. Though neonatal seizures can be identified from electroencephalography (EEG), this is a challenging endeavour since expert visual inspection of EEG recordings is time consuming and prone to errors due the data’s nonstationarity and low signal-to-noise ratio. Towards the greater aim of automatic clinical decision making and monitoring, we propose a multi-output Gaussian process (MOGP) framework for neonatal EEG modelling. In particular, our work builds on the multi-output spectral mixture (MOSM) covariance kernel and shows that MOSM outperforms other commonly-used covariance functions in the literature when it comes to data imputation and hyperparameter-based seizure detection. To the best of our knowledge, our work is the first attempt at modelling and classifying neonatal EEG using MOGPs. Our main contributions are: i) the development of an MOGP-based framework for neonatal EEG analysis; ii) the experimental validation of the MOSM covariance kernel on real-world neonatal EEG for data imputation; and iii) the design of features for EEG based on MOSM hyperparameters and their validation for seizure detection (classification) in a patient specific approach.

Highlights

D ETECTING neonatal seizures [1], [2] is crucial, as they may affect the development of the brain during the first four weeks of a child’s life [3]
As a first step towards this objective, in this article we show that multi-output Gaussian processes (MOGP) [8], and the multi-output spectral mixture (MOSM) kernel [9], are suitable for modelling noise-corrupted EEG signals in terms of imputation of missing data, consistency of model selection and interpretation
Though our approach does not aim to beat the state of the art in terms of classification performance, the fact that MOSM hyperparameters are interpretable in terms of signal power, correlations and delays as explained in Sec

Summary

Introduction

D ETECTING neonatal seizures [1], [2] is crucial, as they may affect the development of the brain during the first four weeks of a child’s life [3]. A. EEG MODELLING Previous approaches to model-based analysis of neonatal EEG have considered parametric, e.g., autoregressive (AR) [10] or nonlinear [11] models. [10] developed an autoregressive model for neonatal EEG, [11] built on a mechanical analogy to model EEG using concepts from nonlinear dynamic systems called Duffing oscillators [12], while [13], [14] characterised multi-channel neonatal EEG from a spectral perspective. These methods struggle to properly account for the dynamic features of seizurerelated EEG comprising fast and repeating patterns [14]. In [15], the authors used Gaussian processes [16] to model neonatal EEG, which allowed them to classify seizure and nonseizure data using the magnitude of the learnt variance of the noise: seizure segments are known to be more repetitive and deterministic having a reduced noise variance than nonseizure ones

Methods

Results

Discussion

Conclusion