In-depth performance analysis of an EEG based neonatal seizure detection algorithm

Abstract

HighlightsA novel method for in-depth analysis of neonatal seizure detection algorithms is proposed.The analysis estimated how seizure features are exploited by automated detectors.This method led to significant improvement of the ANSeR algorithm.

Characterisation of neonatal seizures and their treatment using continuous EEG monitoring: a multicentre experience

“…The most common artefacts seen in the EEG are those due to respiration, ECG, sweat and patting, and these can be difficult for clinicians to identify 17. Although experts generally agree regarding the interpretation of a formal EEG, some seizure patterns are very difficult even for experts to recognise 18…”

Phenobarbital reduces EEG amplitude and propagation of neonatal seizures but does not alter performance of automated seizure detection

“…The main purpose of this study was to examine the effect of phenobarbital on the morphology of seizures and on the performance of our seizure detection algorithm. In a previous study ( Mathieson et al, 2016a ) to investigate the features of seizures affecting ANSeR detection, using the same methodology for seizure analysis but multivariate analysis, it was found that an increase in 4 features; seizure duration, amplitude, rhythmicity and number of EEG channels involved in seizure peak (propagation), were independently associated with an increased likelihood of seizure detection by ANSeR. In the present study phenobarbital did not affect seizure duration or seizure rhythmicity but did reduce seizure amplitude and propagation.…”

“…Each seizure was then analysed by SM using the criteria outlined in Table 1 and features for pre- and post-phenobarbital seizures were compared statistically. The 10 features analysed were drawn from a previous in depth analysis of seizure factors affecting automated seizure detection by ANSeR ( Mathieson et al, 2016a ), and were chosen for the previous study specifically with the feature extraction criteria and general functioning of the algorithm in mind but were also deemed relevant as a general methodology for quantifying seizures for the present study to assess the effects of phenobarbital on seizures. The 10 features can be grouped into 3 broad categories: ‘seizure signature’ (1–5), ‘short-term temporal context or evolution’ of seizures (6–8) and ‘seizure spatial context’ (9–10).…”

“…ANSeR), engineers use general performance metrics such as seizure detection and false detection rates or sensitivity and specificity based on epoched data. End users also wish to know specifically what type of seizures the algorithm detects best and in a previous study ( Mathieson et al, 2016a ), a novel in-depth methodology, using 10 criteria to quantify seizures, was developed to determine the ‘features’ of detected and non-detected seizures. False detections were also categorized.…”

Bio-Signal Complexity Analysis in Epileptic Seizure Monitoring: A Topic Review

“…Moreover, in the field of machine learning, it is very common to combine multiple low-level features to solve a specific classification task. Several studies about epileptic seizure detection have shown that the fusion of multiple features almost always improves the classification accuracy [ 102 , 103 ]. The schematic diagram of the paradigm a typical machine learning algorithm abides is shown in Figure 3 .…”