首页|Automated identification of inter-ictal discharges using residual deep learning neural network amidst of various artefacts
Automated identification of inter-ictal discharges using residual deep learning neural network amidst of various artefacts
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
Background: Visual analysis to identify inter-ictal activity in scalp EEG to support the diagnosis of epilepsy is a challenging task, which is embarked on by an experienced neurologist. Inter-Ictal state is a phase between convolutions (seizures) that are a feature of epilepsy disorder. The objective of this work is to automate the process of identification of inter-ictal activity and to distinguish it from the activity of a controlled patient with and without presence of artifacts Methods: In this work, we have used two-second scalp EEG data. The novel data is collected from Max Super Speciality Hospital, Saket, New Delhi. Expert neurologists mark the data according to the exclusion and inclusion criterion presented and approved by the scientific and ethical committee. Under our archi-tecture, we have first divided the EEG data collected from the patients into two-second segments. The two-second EEG signal is converted to scalograms used as input to fourteen layer novel Residual neu-ral network architecture. For comparison we have created fourteen layer convolution neural network and sixteen layer model where CNN and LSTM models are stacked. For this work we have worked on two cases, the first group is a comparison between intect-ictal and controlled, while the second group is a classification between inte-ictal vs (different artifacts and controlled). Results: We have evaluated our model based on six parameters Accuracy, Sensitivity, Specificity, Precision, Recall, and AUC. Under this architecture, we have divided the complete data set into two parts 80% of data is training data on which k-fold validation is being applied. The value of k is set to 10. The rest, 20%, is used as testing data on which the performance of the model is evaluated. The developed model (RNN) has provided outstanding results in identifying the inter-ictal activity, detecting test dataset with 97.98% accuracy, and has achieved an AUC value of .9974 without the presence of artifacts accuracy of 91.42% and AUC value of 0.9698, has been acheived.Conclusion: Residual neural network in its two-dimensional implementation with fourteen layers has outperformed the two other models developed on similar lines. This research suggests that the proposed architecture has the potential to be utilized in the real-time clinical setup.(c) 2022 Elsevier Ltd. All rights reserved.
ScalogramResidual Deep Neural NetworkEEGEpilepsyClassificationArtifactsWAVELET TRANSFORMCLASSIFICATIONEPILEPSYSEIZURESENTROPYRECOGNITIONPATTERN
NSTL
Elsevier
