A Class-Imbalance Aware and Explainable Spatio-Temporal Graph Attention Network for Neonatal Seizure Detection

Tiny Convolutional Neural Network with Supervised Contrastive Learning for Epileptic Seizure Prediction

Automatic seizure prediction based on ElectroEncephaloGraphy (EEG) ensures the safety of patients with epilepsy and mitigates anxiety. In recent years, significant progress has been made in this field. However, the predictive performance of existing methods encounters a bottleneck that is difficult to overcome. Moreover, there are certain limitations such as significant differences in prediction efficacy among patients or intricate model structures. Given these considerations, Siamese Network (SiaNet) and Triplet Network (TriNet) are proposed based on tiny convolutional neural network and supervised contrastive learning. Short-Time Fourier Transform (STFT) is first applied to the pre-processed data. Then data tuples are constructed and fed into the networks for training. Both networks try to minimize the interval between samples of the same class while maximize the interval between samples of different classes. The two networks consist of multiple branches with shared weights, which can learn from each other via contrastive learning. Promising results are obtained on the CHB-MIT and Siena datasets, with a total of 35 patients. Meanwhile, both models have only 19.351K parameters.

Scaling convolutional neural networks achieves expert level seizure detection in neonatal EEG

Neonatal seizures require urgent treatment, but often go undetected without expert EEG monitoring. We have developed and validated a seizure detection model using retrospective EEG data from 332 neonates. A convolutional neural network was trained and tested on over 50,000 hours (n = 202) of annotated single-channel EEG containing 12,402 seizure events. This model was then validated on two independent multi-reviewer datasets (n = 51 and n = 79). Increasing data and model size improved performance: Matthews correlation coefficient (MCC) and Pearson's correlation (r) increased by up to 50% (15%) with data (model) scaling. The largest model (21m parameters) achieved state-of-the-art on an open-access dataset (MCC = 0.764, r = 0.824, and AUC = 0.982). This model also attained expert-level performance on both validation sets, a first in this field, with no significant difference in inter-rater agreement when the model replaces an expert (∣Δκ∣ < 0.094, p > 0.05).

Analysis of the impact of deep learning know-how and data in modelling neonatal EEG

The performance gains achieved by deep learning models nowadays are mainly attributed to the usage of ever larger datasets. In this study, we present and contrast the performance gains that can be achieved via accessing larger high-quality datasets versus the gains that can be achieved from harnessing the latest deep learning architectural and training advances. Modelling neonatal EEG is particularly affected by the lack of publicly available large datasets. It is shown that greater performance gains can be achieved from harnessing the latest deep learning advances than using a larger training dataset when adopting AUC as a metric, whereas using AUC90 or AUC-PR as metrics greater performance gains are achieved from using a larger dataset than harnessing the latest deep learning advances. In all scenarios the best performance is obtained by combining both deep learning advances and larger datasets. A novel developed architecture is presented that outperforms the current state-of-the-art model for the task of neonatal seizure detection. A novel method to fine-tune the presented model towards site-specific settings based on pseudo labelling is also outlined. The code and the weights of the model are made publicly available for benchmarking future model performances for neonatal seizure detection.

TATPat based explainable EEG model for neonatal seizure detection

The most cost-effective data collection method is electroencephalography (EEG) to obtain meaningful information about the brain. Therefore, EEG signal processing is very important for neuroscience and machine learning (ML). The primary objective of this research is to detect neonatal seizures and explain these seizures using the new version of Directed Lobish. This research uses a publicly available neonatal EEG signal dataset to get comparative results. In order to classify these EEG signals, an explainable feature engineering (EFE) model has been proposed. In this EFE model, there are four essential phases and these phases: (i) automaton and transformer-based feature extraction, (ii) feature selection deploying cumulative weight-based neighborhood component analysis (CWNCA), (iii) the Directed Lobish (DLob) and Causal Connectome Theory (CCT)-based explainable result generation and (iv) classification deploying t algorithm-based support vector machine (tSVM). In the first phase, we have used a channel transformer to get channel numbers and these values have been divided into three levels and these levels are named (1) high, (2) medium and (3) low. By utilizing these levels, we have created an automaton and this automaton has three nodes (each node defines each level). In the feature extraction phase, transition tables of these nodes has been extracted. Therefore, the proposed feature extraction function is termed Triple Nodes Automaton-based Transition table Pattern (TATPat). The used EEG signal dataset contains 19 channels and there are 9 (= 32) connection in the defined automaton. Thus, the presented TATPat extracts 3249 (= 19 × 19 × 9) features from each EEG segment. To choose the most informative features of these 3249 features, a new feature selector which is CWNCA has been applied. By cooperating findings of this feature selector and the presented DLob, the explainable results have been obtained. The last phase is the classification phase and to get high classification performance from this phase, an ensemble classifier (tSVM) has been presented and the classification results have been obtained using two validation techniques which are 10-fold cross-validation (CV) and leave-one subject-out (LOSO) CV. The proposed EFE model generates a DLob string and by using this string, the explainable results have been obtained. Moreover, the presented EFE model attained 99.15% and 76.37% classification accuracy deploying 10-fold and LOSO CVs respectively. According to the classification performances, the recommended TATPat-based EFE is a good model at EEG signal classification. Also, the presented TATPat-based EFE model is a good model for explainable artificial intelligence (XAI) since TTPat-based EFE is cooperating by the DLob.

Efficient EEG Feature Learning Model Combining Random Convolutional Kernel with Wavelet Scattering for Seizure Detection

Automatic seizure detection has significant value in epilepsy diagnosis and treatment. Although a variety of deep learning models have been proposed to automatically learn electroencephalography (EEG) features for seizure detection, the generalization performance and computational burden of such deep models remain the bottleneck of practical application. In this study, a novel lightweight model based on random convolutional kernel transform (ROCKET) is developed for EEG feature learning for seizure detection. Specifically, random convolutional kernels are embedded into the structure of a wavelet scattering network instead of original wavelet transform convolutions. Then the significant EEG features are selected from the scattering coefficients and convolutional outputs by analysis of variance (ANOVA) and minimum redundancy-maximum relevance (MRMR) methods. This model not only preserves the merits of the fast-training process from ROCKET, but also provides insight into seizure detection by retaining only the helpful channels. The extreme gradient boosting (XGboost) classifier was combined with this EEG feature learning model to build a comprehensive seizure detection system that achieved promising epoch-based results, with over 90% of both sensitivity and specificity on the scalp and intracranial EEG databases. The experimental comparisons showed that the proposed method outperformed other state-of-the-art methods for cross-patient and patient-specific seizure detection.

Combining EEG Features and Convolutional Autoencoder for Neonatal Seizure Detection

Neonatal epilepsy is a common emergency phenomenon in neonatal intensive care units (NICUs), which requires timely attention, early identification, and treatment. Traditional detection methods mostly use supervised learning with enormous labeled data. Hence, this study offers a semi-supervised hybrid architecture for detecting seizures, which combines the extracted electroencephalogram (EEG) feature dataset and convolutional autoencoder, called Fd-CAE. First, various features in the time domain and entropy domain are extracted to characterize the EEG signal, which helps distinguish epileptic seizures subsequently. Then, the unlabeled EEG features are fed into the convolutional autoencoder (CAE) for training, which effectively represents EEG features by optimizing the loss between the input and output features. This unsupervised feature learning process can better combine and optimize EEG features from unlabeled data. After that, the pre-trained encoder part of the model is used for further feature learning of labeled data to obtain its low-dimensional feature representation and achieve classification. This model is performed on the neonatal EEG dataset collected at the University of Helsinki Hospital, which has a high discriminative ability to detect seizures, with an accuracy of 92.34%, precision of 93.61%, recall rate of 98.74%, and F1-score of 95.77%, respectively. The results show that unsupervised learning by CAE is beneficial to the characterization of EEG signals, and the proposed Fd-CAE method significantly improves classification performance.

Improving the performance of deep learning models in predicting and classifying gamma passing rates with discriminative features and a class balancing technique: a retrospective cohort study

BACKGROUND

The purpose of this study was to improve the deep learning (DL) model performance in predicting and classifying IMRT gamma passing rate (GPR) by using input features related to machine parameters and a class balancing technique.

METHODS

A total of 2348 fields from 204 IMRT plans for patients with nasopharyngeal carcinoma were retrospectively collected to form a dataset. Input feature maps, including fluence, leaf gap, leaf speed of both banks, and corresponding errors, were constructed from the dynamic log files. The SHAP framework was employed to compute the impact of each feature on the model output for recursive feature elimination. A series of UNet++ based models were trained on the obtained eight feature sets with three fine-tuning methods including the standard mean squared error (MSE) loss, a re-sampling technique, and a proposed weighted MSE loss (WMSE). Differences in mean absolute error, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were compared between the different models.

RESULTS

The models trained with feature sets including leaf speed and leaf gap features predicted GPR for failed fields more accurately than the other models (F(7, 147) = 5.378, p < 0.001). The WMSE loss had the highest accuracy in predicting GPR for failed fields among the three fine-tuning methods (F(2, 42) = 14.149, p < 0.001), while an opposite trend was observed in predicting GPR for passed fields (F(2, 730) = 9.907, p < 0.001). The WMSE_FS5 model achieved a superior AUC (0.92) and more balanced sensitivity (0.77) and specificity (0.89) compared to the other models.

CONCLUSIONS

Machine parameters can provide discriminative input features for GPR prediction in DL. The novel weighted loss function demonstrates the ability to balance the prediction and classification accuracy between the passed and failed fields. The proposed approach is able to improve the DL model performance in predicting and classifying GPR, and can potentially be integrated into the plan optimization process to generate higher deliverability plans.

TRIAL REGISTRATION

This clinical trial was registered in the Chinese Clinical Trial Registry on March 26th, 2020 (registration number: ChiCTR2000031276). https://clinicaltrials.gov/ct2/show/ChiCTR2000031276.

Attention-guided graph structure learning network for EEG-enabled auditory attention detection

Objective: Decoding auditory attention from brain signals is essential for the development of neuro-steered hearing aids. This study aims to overcome the challenges of extracting discriminative feature representations from electroencephalography (EEG) signals for auditory attention detection (AAD) tasks, particularly focusing on the intrinsic relationships between different EEG channels.Approach: We propose a novel attention-guided graph structure learning network, AGSLnet, which leverages potential relationships between EEG channels to improve AAD performance. Specifically, AGSLnet is designed to dynamically capture latent relationships between channels and construct a graph structure of EEG signals.Main result: We evaluated AGSLnet on two publicly available AAD datasets and demonstrated its superiority and robustness over state-of-the-art models. Visualization of the graph structure trained by AGSLnet supports previous neuroscience findings, enhancing our understanding of the underlying neural mechanisms.Significance: This study presents a novel approach for examining brain functional connections, improving AAD performance in low-latency settings, and supporting the development of neuro-steered hearing aids.

Multi-Semantic Decoding of Visual Perception with Graph Neural Networks

Constructing computational decoding models to account for the cortical representation of semantic information plays a crucial role in understanding visual perception. The human visual system processes interactive relationships among different objects when perceiving the semantic contents of natural visions. However, the existing semantic decoding models commonly regard categories as completely separate and independent visually and semantically and rarely consider the relationships from prior information. In this work, a novel semantic graph learning model was proposed to decode multiple semantic categories of perceived natural images from brain activity. The proposed model was validated on the functional magnetic resonance imaging data collected from five normal subjects while viewing 2750 natural images comprising 52 semantic categories. The results showed that the Graph Neural Network-based decoding model achieved higher accuracies than other deep neural network models. Moreover, the co-occurrence probability among semantic categories showed a significant correlation with the decoding accuracy. Additionally, the results suggested that semantic content organized in a hierarchical way with higher visual areas was more closely related to the internal visual experience. Together, this study provides a superior computational framework for multi-semantic decoding that supports the visual integration mechanism of semantic processing.

Epileptic Seizure Detection with an End-to-End Temporal Convolutional Network and Bidirectional Long Short-Term Memory Model

Automatic seizure detection plays a key role in assisting clinicians for rapid diagnosis and treatment of epilepsy. In view of the parallelism of temporal convolutional network (TCN) and the capability of bidirectional long short-term memory (BiLSTM) in mining the long-range dependency of multi-channel time-series, we propose an automatic seizure detection method with a novel end-to-end TCN-BiLSTM model in this work. First, raw EEG is filtered with a 0.5-45 Hz band-pass filter, and the filtered data are input into the proposed TCN-BiLSTM network for feature extraction and classification. Post-processing process including moving average filtering, thresholding and collar technique is then employed to further improve the detection performance. The method was evaluated on two EEG database. On the CHB-MIT scalp EEG database, our method achieved a segment-based sensitivity of 94.31%, specificity of 97.13%, and accuracy of 97.09%. Meanwhile, an event-based sensitivity of 96.48% and an average false detection rate (FDR) of 0.38/h were obtained. On the SH-SDU database we collected, the segment-based sensitivity of 94.99%, specificity of 93.25%, and accuracy of 93.27% were achieved. In addition, an event-based sensitivity of 99.35% and a false detection rate of 0.54/h were yielded. The total detection time consumed for 1[Formula: see text]h EEG data was 5.65[Formula: see text]s. These results demonstrate the superiority and promising potential of the proposed method in real-time monitoring of epileptic seizures.

A Graph-Based Neural Approach to Linear Sum Assignment Problems

Linear assignment problems are well-known combinatorial optimization problems involving domains such as logistics, robotics and telecommunications. In general, obtaining an optimal solution to such problems is computationally infeasible even in small settings, so heuristic algorithms are often used to find near-optimal solutions. In order to attain the right assignment permutation, this study investigates a general-purpose learning strategy that uses a bipartite graph to describe the problem structure and a message passing Graph Neural Network (GNN) model to learn the correct mapping. Comparing the proposed structure with two existing DNN solutions, simulation results show that the proposed approach significantly improves classification accuracy, proving to be very efficient in terms of processing time and memory requirements, due to its inherent parameter sharing capability. Among the many practical uses that require solving allocation problems in everyday scenarios, we decided to apply the proposed approach to address the scheduling of electric smart meters access within an electricity distribution smart grid infrastructure, since near-real-time energy monitoring is a key element of the green transition that has become increasingly important in recent times. The results obtained show that the proposed graph-based solver, although sub-optimal, exhibits the highest scalability, compared with other state-of-the-art heuristic approaches. To foster the reproducibility of the results, we made the code available at https://github.com/aircarlo/GNN_LSAP.

A Sequential End-to-End Neonatal Sleep Staging Model with Squeeze and Excitation Blocks and Sequential Multi-Scale Convolution Neural Networks

Automatic sleep staging offers a quick and objective assessment for quantitatively interpreting sleep stages in neonates. However, most of the existing studies either do not encompass any temporal information, or simply apply neural networks to exploit temporal information at the expense of high computational overhead and modeling ambiguity. This limits the application of these methods to multiple scenarios. In this paper, a sequential end-to-end sleep staging model, SeqEESleepNet, which is competent for parallelly processing sequential epochs and has a fast training rate to adapt to different scenarios, is proposed. SeqEESleepNet consists of a sequence epoch generation (SEG) module, a sequential multi-scale convolution neural network (SMSCNN) and squeeze and excitation (SE) blocks. The SEG module expands independent epochs into sequential signals, enabling the model to learn the temporal information between sleep stages. SMSCNN is a multi-scale convolution neural network that can extract both multi-scale features and temporal information from the signal. Subsequently, the followed SE block can reassign the weights of features through mapping and pooling. Experimental results exhibit that in a clinical dataset, the proposed method outperforms the state-of-the-art approaches, achieving an overall accuracy, F1-score, and Kappa coefficient of 71.8%, 71.8%, and 0.684 on a three-class classification task with a single channel EEG signal. Based on our overall results, we believe the proposed method could pave the way for convenient multi-scenario neonatal sleep staging methods.

Deep Learning-Based Classification of Epileptic Electroencephalography Signals Using a Concentrated Time-Frequency Approach

ConceFT (concentration of frequency and time) is a new time-frequency (TF) analysis method which combines multitaper technique and synchrosqueezing transform (SST). This combination produces highly concentrated TF representations with approximately perfect time and frequency resolutions. In this paper, it is aimed to show the TF representation performance and robustness of ConceFT by using it for the classification of the epileptic electroencephalography (EEG) signals. Therefore, a signal classification algorithm which uses TF images obtained with ConceFT to feed the transfer learning structure has been presented. Epilepsy is a common neurological disorder that millions of people suffer worldwide. Daily lives of the patients are quite difficult because of the unpredictable time of seizures. EEG signals monitoring the electrical activity of the brain can be used to detect approaching seizures and make possible to warn the patient before the attack. GoogLeNet which is a well-known deep learning model has been preferred to classify TF images. Classification performance is directly related to the TF representation accuracy of the ConceFT. The proposed method has been tested for various classification scenarios and obtained accuracies between 95.83% and 99.58% for two and three-class classification scenarios. High results show that ConceFT is a successful and promising TF analysis method for non-stationary biomedical signals.