The electroencephalogram (EEG), for measuring the electrophysiological activity of the brain, has been widely applied in automatic detection of epilepsy seizures. Various EEG-based seizure detection algorithms have already yielded high sensitivity, but training those algorithms requires a large amount of labelled data. Data labelling is often done with a lot of human efforts, which is very time-consuming. In this study, we propose a hybrid system integrating an unsupervised learning (UL) module and a supervised learning (SL) module, where the UL module can significantly reduce the workload of data labelling. For preliminary seizure screening, UL synthesizes amplitude-integrated EEG (aEEG) extraction, isolation forest-based anomaly detection, adaptive segmentation, and silhouette coefficient-based anomaly detection evaluation. The UL module serves to quickly locate the determinate subjects (seizure segments and seizure-free segments) and the indeterminate subjects (potential seizure candidates). Afterwards, more robust seizure detection for the indeterminate subjects is performed by the SL using an EasyEnsemble algorithm. EasyEnsemble, as a class-imbalance learning method, can potentially decrease the generalization error of the seizure-free segments. The proposed method can significantly reduce the workload of data labelling while guaranteeing satisfactory performance. The proposed seizure detection system is evaluated using the Children’s Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) scalp EEG dataset, and it achieves a mean accuracy of 92.62%, a mean sensitivity of 95.55%, and a mean specificity of 92.57%. To the best of our knowledge, this is the first epilepsy seizure detection study employing the integration of both the UL and the SL modules, achieving a competitive performance superior or similar to that of the state-of-the-art methods.

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring well-trained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two state-of-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the source domain, four small sleep databases as the target domain. Thereinto, clinical sleep records acquired in Huashan Hospital, Shanghai, were used. The results show that both DSA and AdaDSA could significantly improve the performance of source models on target domains, providing novel insights into the domain generalization problem in sleep staging tasks.

Sleep posture, as a crucial index for sleep quality assessment, has been widely studied in sleep analysis. In this paper, an unobtrusive smart mat system based on a dense flexible sensor array and printed electrodes along with an algorithmic framework for sleep posture recognition is proposed. With the dense flexible sensor array, the system offers a comfortable and high-resolution solution for long-term pressure sensing. Meanwhile, compared to other methods, it reduces production costs and computational complexity with a smaller area of the mat and improves portability with fewer sensors. To distinguish the sleep posture, the algorithmic framework that includes preprocessing and Deep Residual Networks (ResNet) is developed. With the ResNet, the proposed system can omit the complex hand-crafted feature extraction process and provide compelling performance. The feasibility and reliability of the proposed system were evaluated on seventeen subjects. Experimental results exhibit that the accuracy of the short-term test is up to 95.08% and the overnight sleep study is up to 86.35% for four categories (supine, prone, right, and left) classification, which outperform the most of state-of-the-art studies. With the promising results, the proposed system showed great potential in applications like sleep studies, prevention of pressure ulcers, etc.

Surface electromyogram (sEMG) based hand gesture recognition for prosthesis or armband is an important application of the human-machine interface. However, the measurement location of sensors greatly influences the hand gesture performance, especially with the inter-day or inter-subject validation protocols. Therefore, we acquired two-day hand gesture data of 41 subjects with a 256 (16×16) channel high-density sEMG electrode array. With the acquired data, we initially compared the support vector machine (SVM) and other four state-of-art classifiers under three validation protocols, i.e., intra-day, inter-day and inter-subject validation protocols. Then, we screened 14 feature optimization techniques, including 5 feature-projection methods and 9 feature-ranking approaches. To present the accuracy tendency with varying measure locations, we systematically explored the 10-hand gesture performance using data of 16 prosthesis measurement locations (PMLs) and 15 armband measurement locations (AMLs). As a result, the SVM classifier was suitable for the intra-day and inter-day validation protocols and the 2-dimensional convolutional neural network was selected for the inter-subject validation protocol. The mean accuracies of the hand gesture classification ranged from 95.68% to 99.12% (intra-day validation), from 68.41% to 88.02% (inter-day validation) and from 63.39% to 86.33% (inter-subject validation) for the prosthesis application. In addition, for the armband application, the mean accuracies ranged from 96.25% to 97.43% (intra-day validation), from 67.44% to 75.83% (inter-day validation) and from 65.53% to 75.40% (inter-subject validation). The accuracy is greatly correlated with the measurement location, which is highly associated with the neuromuscular structures of human bodies. In summary, our work can serve as a factor-screening tool for users customizing their systems according to their physical conditions and requirements.

Commonly used sensors like accelerometers, gyroscopes, surface electromyography sensors, etc., which provide a convenient and practical solution for human activity recognition (HAR), have gained extensive attention. However, which kind of sensor can provide adequate information in achieving a satisfactory performance, or whether the position of a single sensor would play a significant effect on the performance in HAR are sparsely studied. In this paper, a comparative study to fully investigate the performance of the aforementioned sensors for classifying four activities (walking, tooth brushing, face washing, drinking) is explored. Sensors are spatially distributed over the human body, and subjects are categorized into three groups (able-bodied people, stroke survivors, and the union of both). Performances of using accelerometer, gyroscope, sEMG, and their combination in each group are evaluated by adopting the Support Vector Machine classifier with the Leave-One-Subject-Out Cross-Validation technique, and the optimal sensor position for each kind of sensor is presented based on the accuracy. Experimental results show that using the accelerometer could obtain the best performance in each group. The highest accuracy of HAR involving stroke survivors was 95.84 ± 1.75% (mean ± standard error), achieved by the accelerometer attached to the extensor carpi ulnaris. Furthermore, taking the practical application of HAR into consideration, a novel approach to distinguish various activities of stroke survivors based on a pre-trained HAR model built on healthy subjects is proposed, the highest accuracy of which is 77.89 ± 4.81% (mean ± standard error) with the accelerometer attached to the extensor carpi ulnaris.