Mathieson, SR;
(2016)
Neurophysiology Based Performance Analysis of a Novel Automated Seizure Detection Algorithm for Neonatal EEG.
Doctoral thesis , UCL (University College London).
Preview |
Text
Mathieson_E-thesis + articles.pdf Download (29MB) | Preview |
Preview |
Text (Supporting digital documentation on CD: Study 5: ANSeR educational support document)
SUPPORTING DIGITAL DOCUMENTATION. Study 5. ANSeR educational support document.pdf - Supplemental Material Download (10MB) | Preview |
Preview |
Text (Supporting digital documentation on CD: Study 5: ANSeR educational support document (Powerpoint version))
SUPPORTING DIGITAL DOCUMENTATION. Study 5.ANSeR educational support document powerpoint.pdf - Supplemental Material Download (5MB) | Preview |
Preview |
Text (Supporting digital documentation on CD: Study 5: ANSeR educational support document Test package 1)
SUPPORTING DIGITAL DOCUMENTATION. Study 5.Test package 1.pdf - Supplemental Material Download (15MB) | Preview |
Preview |
Text (Supporting digital documentation on CD: Study 5: ANSeR educational support document Test package 2)
SUPPORTING DIGITAL DOCUMENTATION. Study 5.Test package 2.pdf - Supplemental Material Download (16MB) | Preview |
Abstract
To aid seizure detection in sick neonates, our group has developed an automated seizure detection algorithm (ANSeR) and published initial performance results. In this thesis a validation study of the performance of ANSeR on a large unedited, unseen dataset of 70 EEGs from 2 institutions is presented. Results indicate that ANSeR sensitivity thresholds between 0.5-0.3 provide performance considered acceptable for clinical use with seizure detection rates of 52.6-75% and false detection rates of 0.04 -0.36 FD/h respectively. To determine the features of seizures affecting automated detection, a subset of 20 EEGs from the validation study were selected and seizures were manually analysed using a novel set of 10 criteria. Using multivariate analysis, 4 seizure features were found to affect automated detection including; seizure amplitude, duration, rhythmicity and propagation. The main causes of false detection were also characterised and quantified leading to an adaptation of the algorithm with improved performance. Observations suggest that phenobarbitone, a first line anticonvulsant, may affect seizure morphology and potentially the performance of ANSeR. Using similar seizure analysis criteria to compare pre and post phenobarbitone seizures, it was shown that post phenobarbitone seizures were both lower amplitude and showed reduced propagation but the performance of ANSeR was unaffected. As well as an ongoing quantification of seizure burden, it is important that automated algorithms detect seizures soon after onset to facilitate prompt treatment. Results of retrospective analysis of seizure detection by ANSeR here suggests that the use of ANSeR may reduce the latency of first treatment after seizure onset by 30-40 minutes at clinically relevant ANSeR thresholds, compared to current practice. As rhythmic artefacts and other sources may result in ANSeR false detections, clinicians reviewing the EEG as a result of ANSeR detections must decide if the detection is a true seizure or a false detection. A training resource developed by the author improved the ability of clinical staff to discriminate true seizures from false detections in a randomised study.
Archive Staff Only
 |
View Item |
