EEG-Oriented Self-Supervised Learning With Triple Information Pathways Network.

Recently, deep learning-based electroencephalogram (EEG) analysis and decoding have attracted widespread attention for monitoring the clinical condition of users and identifying their intention/emotion. Nevertheless, the existing methods generally model EEG signals with limited viewpoints or restricted concerns about the characteristics of the EEG signals, and thus represent complex spectro-/spatiotemporal patterns and suffer from high variability. In this work, we propose the novel EEG-oriented self-supervised learning methods and a novel deep architecture to learn rich representation, including information about the diverse spectral characteristics of neural oscillations, the spatial properties of electrode sensor distribution, and the temporal patterns of both the global and local viewpoints. Along with the proposed self-supervision strategies and deep architectures, we devise a feature normalization strategy to resolve the intra-/inter-subject variability problem. We demonstrate the validity of our proposed deep learning framework on the four publicly available datasets by conducting comparisons with the state of the art baselines. It is also noteworthy that we exploit the same network architecture for the four different EEG paradigms and outperform the comparison methods, thereby meeting the challenge of the task-dependent network architecture engineering in EEG-based applications.

Gender-Related and Hemispheric Effects in Cortical Thickness-Based Hemispheric Brain Morphological Network.

OBJECTIVE: The current study examined gender-related differences in hemispheric asymmetries of graph metrics, calculated from a cortical thickness-based brain structural covariance network named hemispheric morphological network. METHODS: Using the T1-weighted magnetic resonance imaging scans of 285 participants (150 females, 135 males) retrieved from the Human Connectome Project (HCP), hemispheric morphological networks were constructed per participant. In these hemispheric morphologic networks, the degree of similarity between two different brain regions in terms of the distributed patterns of cortical thickness values (the Jensen-Shannon divergence) was defined as weight of network edge that connects two different brain regions. After the calculation and summation of global and local graph metrics (across the network sparsity levels K = 0.10-0.36), asymmetry indexes of these graph metrics were derived. RESULTS: Hemispheric morphological networks satisfied small-worldness and global efficiency for the network sparsity ranges of K = 0.10-0.36. Between-group comparisons (female versus male) of asymmetry indexes revealed opposite directionality of asymmetries (leftward versus rightward) for global metrics of normalized clustering coefficient, normalized characteristic path length, and global efficiency (all p < 0.05). For the local graph metrics, larger rightward asymmetries of cingulate-superior parietal gyri for nodal efficiency in male compared to female, larger leftward asymmetry of temporal pole for degree centrality in female compared to male, and opposite directionality of interhemispheric asymmetry of rectal gyrus for degree centrality between female (rightward) and male (leftward) were shown (all p < 0.05). CONCLUSION: Patterns of interhemispheric asymmetries for cingulate, superior parietal gyrus, temporal pole, and rectal gyrus are different between male and female for the similarities of the cortical thickness distribution with other brain regions. Accordingly, possible effect of gender-by-hemispheric interaction has to be considered in future studies of brain morphology and brain structural covariance networks.

Dynamic activity classification based on automatic adaptation of postural orientation.

We propose a dynamic activity classification system with tri-axial accelerometer sensor using adaptation of user's postural orientation. In general, the sensor module is worn at a fixed position such as waist, head, wrist, thigh, and so on. However, in reality, the tilt of the attached sensor could be changed from time to time in actions such as sitting down, standing up, lying, walking or running. Moreover, most of the users want to wear the sensor at their own favorite positions instead of a recommended position. In these cases, the activity detection methods based on fixed tilt value may produce serious problem in their performance. Therefore, we propose a user adapted activity classification method which enables users to freely wear the sensor everywhere on their torso. In order to decide tilt values corresponding user's postural orientation, we focused on tilt-free activities such as walking and running. While walking, the algorithm tries to modify the predefined reference tilt values for the three axes, X, Y and Z. From an experiment, we have achieved 88% of the activity classification accuracy even though the tilt angle is changed while wearing sensors.

Toward real time detection of the basic living activity in home using a wearable sensor and smart home sensors.

As the elderly people living alone are enormously increasing recently, we need the system inferring activities of daily living (ADL) for maintaining healthy life and recognizing emergency. The system should be constructed with sensors, which are used to associate with people's living while remaining as non intrusive views as possible. To do this, the proposed system use a triaxial accelerometer sensor and environment sensors indicating contact with subject in home. Particularly, in order to robustly infer ADLs, we present component ADL, which is decided with conjunction of human motion together, not just only contacted object identification. It is an important component in inferring ADL. In special, component ADL decision firstly refines misclassified initial activities, which improves the accuracy of recognizing ADL. Preliminary experiments results for proposed system provides overall recognition rate of over 97% over 8 component ADLs, which can be effectively applicable to recognize the final ADLs.

An effective method for component activity classification supporting location awareness and user identification.

We present an effective method for component activity classification supporting location awareness and user identification at the same time. The system is comprised of three modules: Pressure Sensing Module (PSM), Activity Detecting Module (ADM), and Receiving Station (RS). The ADM having a unique id is a wearable module putting on one's waist-belt, which classifies component activity such as sitting chair, lying bed, sitting sofa, etc. utilizing both user's interaction with household furniture and atomic activities like lie, sit, and stand. We limit transmission range of RF chip in PSM to around 1 m so that we can find the most adjacent furniture to the ADM. It makes possible to find the user's relative location to the PSM, so we can aware of both who and where the acting person is while recognizing his/her activities. We obtained 92.5% of average precision of the activity classification.

Semantic tagging for medical knowledge tracking.

We propose a semantic tagger that provides high level concept information for phrases in clinical documents, which enriches medical information tracking system that support decision making or quality assurance of medical treatment. In this paper, we have tried to deal with patient records written by doctors rather than well-formed documents such as Medline abstracts. In addition, annotating clinical text on phrases semantically rather than syntactically has been attempted, which are at higher level granularity than words that have been the target for most tagging work.