Brain-computer interface provides a new communication bridge between the human mind and devices, depending largely on the accurate classification and identification of non-invasive EEG signals. Recently, the deep learning approaches have been widely used in many fields to extract features and classify various types of data successfully. However, the deep learning approach requires massive data to train its neural networks, and the amount of data impacts greatly on the quality of the classifiers. This paper proposes a novel approach that combines deep learning and data augmentation for EEG classification. We applied the empirical mode decomposition on the EEG frames and mixed their intrinsic mode functions to create new artificial EEG frames, followed by transforming all EEG data into tensors as inputs of the neural network by complex Morlet wavelets. We proposed two neural networks-convolutional neural network and wavelet neural network-to train the weights and classify two classes of motor imagery signals. The wavelet neural network is a new type of neural network using wavelets to replace the convolutional layers. The experimental results show that the artificial EEG frames substantially improve the training of neural networks, and both two networks yield relatively higher classification accuracies compared to prevailing approaches. Meanwhile, we also verified the performance of our new proposed wavelet neural network model in the classification of steady-state visual evoked potentials.

Brain-computer interface (BCI) technologies have been widely used in many areas. In particular, non-invasive technologies such as electroencephalography (EEG) or near-infrared spectroscopy (NIRS) have been used to detect motor imagery, disease, or mental state. It has been already shown in literature that the hybrid of EEG and NIRS has better results than their respective individual signals. The fusion algorithm for EEG and NIRS sources is the key to implement them in real-life applications. In this research, we propose three fusion methods for the hybrid of the EEG and NIRS-based brain-computer interface system: linear fusion, tensor fusion, and p th-order polynomial fusion. Firstly, our results prove that the hybrid BCI system is more accurate, as expected. Secondly, the p th-order polynomial fusion has the best classification results out of the three methods, and also shows improvements compared with previous studies. For a motion imagery task and a mental arithmetic task, the best detection accuracy in previous papers were 74.20% and 88.1%, whereas our accuracy achieved was 77.53% and 90.19%. Furthermore, unlike complex artificial neural network methods, our proposed methods are not as computationally demanding.

Previous studies made progress in the early diagnosis of Alzheimer’s disease (AD) using electroencephalography (EEG) without considering EEG connectivity. To fill this gap, we explored significant differences between early AD patients and controls based on frequency domain and spatial properties using functional connectivity in mild cognitive impairment (MCI) and mild AD datasets. Four global metrics, network resilience, connection-level metrics and node versatility were used to distinguish between controls and patients. The results show that the main frequency bands that are different between MCI patients and controls are the $\theta $ and low $\alpha $ bands, and the differently affected brain areas are the frontal, left temporal and parietal areas. Compared to MCI patients, in patients with mild AD, the main frequency bands that are different are the low and high $\alpha $ bands, and the main differently affected brain region is a larger right temporal area. Four LOFC bands were used as input to train the ResNet-18 model. For the MCI dataset, the average accuracy of 20 runs was 93.42% and the best accuracy was 98.33%, while for the mild AD dataset, the average accuracy was 98.54% and the best accuracy was 100%. To determine the timing of early treatment and discovering the susceptible patients, and to slow the progression of the disease, we assume that the occurrence of MCI and mild AD and their progression to more serious AD and dementia could be inferred by analyzing the topological structure of the brain network generated by EEG. Our findings provide a novel solution for connectome-based biomarker analysis to improve personalized medicine.

Treadmills are widely used to recover walking function in the rehabilitation field for those patients with gait disorders. Nevertheless, the ultimate goal of walking function recovery is to walk on the ground rather than on the treadmill. This study aims to determine the effect of treadmill walking on gait and upper trunk movement characteristics using wearable sensors. Eight healthy male subjects are recruited to perform 420-m straight overground walking (OW) and 5 min treadmill walking (TW), wearing 3 inertial measurement units and a pair of insole sensors. In addition to common linear features, nonlinear features, which contains sample entropy, maximal Lyapunov exponent and fractal dynamic of stride intervals (detrended fluctuation analysis), are used to compare the difference between TW and OW condition. Canonical correlation analysis is also used to indicate the correlation between upper trunk movement characteristics and gait features in the aspects of spatiotemporal parameters and gait dynamic features. The experimental results show that the treadmill can cause a shorter stride length, less stride time and worsen long-range correlation of stride intervals. And the treadmill can significantly increase the stability for both gait and upper trunk, while it can significantly reduce gait regularity during swing phase. Canonical correlation analysis results show that treadmill can reduce the correlation between gait and upper trunk features. One possible interpretation of these results is that people tend to walk more cautiously to prevent the risk of falling and neglect the coordination between gait and upper trunk when walking on the treadmill. This study can provide fundamental insightful information about the effect of treadmill walking on gait and upper trunk to support future similar studies.

In this article, we present a collection of fifteen novel contributions on machine learning methods with low-quality or imperfect datasets, which were accepted for publication in the special issue “Machine Learning Methods with Noisy, Incomplete or Small Datasets”, Applied Sciences (ISSN 2076-3417). These papers provide a variety of novel approaches to real-world machine learning problems where available datasets suffer from imperfections such as missing values, noise or artefacts. Contributions in applied sciences include medical applications, epidemic management tools, methodological work, and industrial applications, among others. We believe that this special issue will bring new ideas for solving this challenging problem, and will provide clear examples of application in real-world scenarios.

Great progress has been made in diagnosing medical diseases based on deep learning. Large-scale medical data are expected to improve deep learning performance further. It is almost impossible for a single institution to collect so much data due to the time-consuming and costly collection and labeling of medical data. Many studies have turned attention to data sharing among multiple medical institutions. However, due to different data acquiring and processing procedures, multiple institutions' medical data is characterized by distribution heterogeneity. Besides, the protection of patient privacy in medical data sharing has also been a common concern. To simultaneously address the problems of heterogeneous data distribution and privacy protection, we propose a novel multi-source source free domain adaptation. When aligning distributed heterogeneous data, our method only requires the pre-trained source models rather than direct access to the source domain data, thus protecting patients' privacy. In addition, it has the advantages of being efficient and less costly in network resources. The proposed method is evaluated on the multi-site fMRI database Autism Brain Imaging Data Exchange (ABIDE) and yields an average accuracy of 69.54%. We also analyzed its effectiveness on network resource-saving and conducted additional experiments on Camelyon17 to validate the generalization.