Investigating the Impact of CNN Depth on Neonatal Seizure Detection Performance

A method for AI assisted human interpretation of neonatal EEG

The study proposes a novel method to empower healthcare professionals to interact and leverage AI decision support in an intuitive manner using auditory senses. The method’s suitability is assessed through acoustic detection of the presence of neonatal seizures in electroencephalography (EEG). Neurophysiologists use EEG recordings to identify seizures visually. However, neurophysiological expertise is expensive and not available 24/7, even in tertiary hospitals. Other neonatal and pediatric medical professionals (nurses, doctors, etc.) can make erroneous interpretations of highly complex EEG signals. While artificial intelligence (AI) has been widely used to provide objective decision support for EEG analysis, AI decisions are not always explainable. This work developed a solution to combine AI algorithms with a human-centric intuitive EEG interpretation method. Specifically, EEG is converted to sound using an AI-driven attention mechanism. The perceptual characteristics of seizure events can be heard using this method, and an hour of EEG can be analysed in five seconds. A survey that has been conducted among targeted end-users on a publicly available dataset has demonstrated that not only does it drastically reduce the burden of reviewing the EEG data, but also the obtained accuracy is on par with experienced neurophysiologists trained to interpret neonatal EEG. It is also shown that the proposed communion of a medical professional and AI outperforms AI alone by empowering the human with little or no experience to leverage AI attention mechanisms to enhance the perceptual characteristics of seizure events.

A new hybrid model of deep learning ResNeXt-SVM for Weed Detection

A set of experiments has shown that deep learning as well as traditional learning can be used in the weed detection process and perform well, although sometimes these models cannot fully exploit and utilize the long-term dependency relationship between some key features of images and image labels. To remedy this known problem in the field of image classification we have introduced a classifier known as the Linear Support Vector Machine (SVM). Specifically, we have combined a ResNeXt and SVM network to provide the ResNeXt-SVM framework that can deepen the exploitation of the structured features of images and the understanding of their content. The experimental results show that compared to other algorithm models such as ResNet, ResNeXt and VGG, our proposed solution is more precise and efficient in classifying weeds.

Ensemble Learning Using Individual Neonatal Data for Seizure Detection

Objective: Sharing medical data between institutions is difficult in practice due to data protection laws and official procedures within institutions. Therefore, most existing algorithms are trained on relatively small electroencephalogram (EEG) data sets which is likely to be detrimental to prediction accuracy. In this work, we simulate a case when the data can not be shared by splitting the publicly available data set into disjoint sets representing data in individual institutions. Methods and procedures: We propose to train a (local) detector in each institution and aggregate their individual predictions into one final prediction. Four aggregation schemes are compared, namely, the majority vote, the mean, the weighted mean and the Dawid-Skene method. The method was validated on an independent data set using only a subset of EEG channels. Results: The ensemble reaches accuracy comparable to a single detector trained on all the data when sufficient amount of data is available in each institution. Conclusion: The weighted mean aggregation scheme showed best performance, it was only marginally outperformed by the Dawid–Skene method when local detectors approach performance of a single detector trained on all available data. Clinical impact: Ensemble learning allows training of reliable algorithms for neonatal EEG analysis without a need to share the potentially sensitive EEG data between institutions.

A High Accuracy Electrographic Seizure Classifier Trained Using Semi-Supervised Labeling Applied to a Large Spectrogram Dataset

The objective of this study was to explore using ECoG spectrogram images for training reliable cross-patient electrographic seizure classifiers, and to characterize the classifiers’ test accuracy as a function of amount of training data. ECoG channels in ∼138,000 time-series ECoG records from 113 patients were converted to RGB spectrogram images. Using an unsupervised spectrogram image clustering technique, manual labeling of 138,000 ECoG records (each with up to 4 ECoG channels) was completed in 320 h, which is an estimated 5 times faster than manual labeling without ECoG clustering. For training supervised classifier models, five random folds of data were created; with each fold containing 72, 18, and 23 patients’ data for model training, validation and testing respectively. Five convolutional neural network (CNN) architectures, including two with residual connections, were trained. Cross-patient classification accuracies and F₁ scores improved with model complexity, with the shallowest 6-layer model (with ∼1.5 million trainable parameters) producing a class-balanced seizure/non-seizure classification accuracy of 87.9% on ECoG channels and the deepest ResNet50-based model (with ∼23.5 million trainable parameters) producing a classification accuracy of 95.7%. The trained ResNet50-based model additionally had 93.5% agreement in scores with an independent expert labeller. Visual inspection of gradient-based saliency maps confirmed that the models’ classifications were based on relevant portions of the spectrogram images. Further, by repeating training experiments with data from varying number of patients, it was found that ECoG spectrogram images from just 10 patients were sufficient to train ResNet50-based models with 88% cross-patient accuracy, while at least 30 patients’ data was required to produce cross-patient classification accuracies of >90%.

Non-invasive wearable seizure detection using long-short-term memory networks with transfer learning

Objective. The detection of seizures using wearable devices would improve epilepsy management, but reliable detection of seizures in an ambulatory environment remains challenging, and current studies lack concurrent validation of seizures using electroencephalography (EEG) data.Approach. An adaptively trained long-short-term memory deep neural network was developed and trained using a modest number of seizure data sets from wrist-worn devices. Transfer learning was used to adapt a classifier that was initially trained on intracranial electroencephalography (iEEG) signals to facilitate classification of non-EEG physiological datasets comprising accelerometry, blood volume pulse, skin electrodermal activity, heart rate, and temperature signals. The algorithm's performance was assessed with and without pre-training on iEEG signals and transfer learning. To assess the performance of the seizure detection classifier using long-term ambulatory data, wearable devices were used for multiple months with an implanted neurostimulator capable of recording iEEG signals, which provided independent electrographic seizure detections that were reviewed by a board-certified epileptologist.Main results. For 19 motor seizures from 10 in-hospital patients, the algorithm yielded a mean area under curve (AUC), a sensitivity, and an false alarm rate per day (FAR/day) of 0.98, 0.93, and 2.3, respectively. Additionally, for eight seizures with probable motor semiology from two ambulatory patients, the classifier achieved a mean AUC of 0.97 and an FAR of 2.45 events/day at a sensitivity of 0.9. For all seizure types in the ambulatory setting, the classifier had a mean AUC of 0.82 with a sensitivity of 0.47 and an FAR of 7.2 events/day.Significance. The performance of the algorithm was evaluated using motor and non-motor seizures during in-hospital and ambulatory use. The classifier was able to detect multiple types of motor and non-motor seizures, but performed significantly better on motor seizures.

Hybrid neonatal EEG seizure detection algorithms achieve the benchmark of visual interpretation of the human expert

Neonatal EEG seizure detection algorithms (NSDAs) have an upper bound of performance related to the agreement between visual interpretation of human experts. No published algorithms have reported performance that has reached this upper bound. In this paper, we combined two recently developed NSDAs in order to improve detection performance. An outlier detection stage was also added to improve robustness in the presence of unseen data. A large database of EEG from 79 term infants labeled by three independent human experts was used to develop and test the sufficiency of the hybrid NSDA. The inter-observer agreement (IOA) between experts was high (κ = 0.757, 95%CI: 0.665-0.836, n=79). The area under the receiver operator characteristic of the NSDA compared to the consensus annotation of the human experts was 0.952 (95%CI: 0.0927-0.971). The IOA of seizure detection between the NSDA and human experts was not significantly less than the IOA among human experts (Δκ = 0.022, 95%CI: -0.20 to 0.072) and was further improved by increasing the minimum seizure duration from 10s to 30s (Δκ = -0.002, 95%CI: -0.073 to 0.055). Automated methods of neonatal EEG seizure detection have sufficient accuracy to replace human interpretation, potentially, providing reliable interpretations for clinicians in the neonatal intensive care unit.

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

The manual review of an electroencephalogram (EEG) for seizure detection is a laborious and error-prone process. Thus, automated seizure detection based on machine learning has been studied for decades. Recently, deep learning has been adopted in order to avoid manual feature extraction and selection. In the present study, we systematically compared the performance of different combinations of input modalities and network structures on a fixed window size and dataset to ascertain an optimal combination of input modalities and network structures. The raw time-series EEG, periodogram of the EEG, 2D images of short-time Fourier transform results, and 2D images of raw EEG waveforms were obtained from 5-s segments of intracranial EEGs recorded from a mouse model of epilepsy. A fully connected neural network (FCNN), recurrent neural network (RNN), and convolutional neural network (CNN) were implemented to classify the various inputs. The classification results for the test dataset showed that CNN performed better than FCNN and RNN, with the area under the curve (AUC) for the receiver operating characteristics curves ranging from 0.983 to 0.984, from 0.985 to 0.989, and from 0.989 to 0.993 for FCNN, RNN, and CNN, respectively. As for input modalities, 2D images of raw EEG waveforms yielded the best result with an AUC of 0.993. Thus, CNN can be the most suitable network structure for automated seizure detection when applied to the images of raw EEG waveforms, since CNN can effectively learn a general spatially-invariant representation of seizure patterns in 2D representations of raw EEG.