Investigating the Impact of CNN Depth on Neonatal Seizure Detection Performance

Abstract

This study presents a novel, deep, fully convolutional architecture which is optimized for the task of EEG-based neonatal seizure detection. Architectures of different depths were designed and tested; varying network depth impacts convolutional receptive fields and the corresponding learned feature complexity. Two deep convolutional networks are compared with a shallow SVMbased neonatal seizure detector, which relies on the extraction of hand-crafted features. On a large clinical dataset, of over 800 hours of multichannel unedited EEG, containing 1389 seizure events, the deep 11-layer architecture significantly outperforms the shallower architectures, improving the AUC90 from 82.6% to 86.8%. Combining the end-to-end deep architecture with the feature-based shallow SVM further improves the AUC90 to 87.6%. The fusion of classifiers of different depths gives greatly improved performance and reduced variability, making the combined classifier more clinically reliable.

Deep learning-based electroencephalography analysis: a systematic review

“…While the answer depends on many factors such as the domain of application, we observed that in some cases raw EEG as input consistently outperformed baselines based using classically extracted features. For example, for seizure classification, recently proposed models using raw EEG data as input [64,185,156] achieved better performances than classical baseline methods, such as SVMs with frequency-domain features. For this particular task, we believe following the current trend of using raw EEG data is the best way to start exploring a new approach.…”

Comparison of different input modalities and network structures for deep learning-based seizure detection

“…This process clearly resembles the human inspection of EEG from the images www.nature.com/scientificreports www.nature.com/scientificreports/ on a computer screen. Our classifier for the raw EEG waveform images also showed better results compared to the three recent CNN-based seizure classifiers for the raw temporal EEG 26 , FFT results 25 , and STFT images 29 ( Table 1). More importantly, classifiers for the multi-channel human iEEG also demonstrated a clear performance advantage with the images of the raw EEG waveform (Fig.…”

“…These studies have been based on different deep neural network structures, such as a fully connected neural network (FCNN) 24 , convolutional neural network (CNN) 22,[25][26][27] , and recurrent neural network (RNN) 28 . These different neural networks can automatically learn discriminative features from various types of data input, including raw temporal EEG 26 , FFT results 25 , 2-dimensional (2D) representation of STFT results 29 , and 2D images of raw EEG 27 . The adoption of different input forms and network structures typically makes it difficult to directly compare performance among different deep learning methods.…”

Neonatal EEG Interpretation and Decision Support Framework for Mobile Platforms

“…4. It can be seen that the deeper 11-layer system consistently outperforms the 6layer alternative for the entire range of FDs/h [14]. The computational cost results presented in Table II indicate that the CPU, RAM and battery consumptions are larger for the deeper 11-layer architecture with the power consumption of the 11-layer architecture increased by 7.9mAh.…”

“…The 11layer architecture allows for more complex feature extraction but also increases processing time and power consumption. Performance comparison of the 6 and 11 layer algorithms is reported in [14]. Here, the two architectures are contrasted with more clinically relevant event-based metrics such as good detection rate (GDR) and the number of false detections per hour (FD/h).…”