Brain Informatics

Electrical recordings of neural mass activity, such as local field potentials (LFPs) and electroencephalograms (EEGs), have been instrumental in studying brain function. However, being aggregate signals that lack cellular resolution, these signals are not easy to interpret directly in terms of neural functions. Developing tools for a reliable estimation of key neural parameters from these signals, such as the interaction between excitation and inhibition or the level of neuromodulation, is important both for neuroscience and clinical applications. Over the years we have developed tools based on the combination of neural network modelling and computational analysis of empirical data to estimate neural parameters from aggregate neural signals. The purpose of this paper, which accompanies an Invited Plenary Lecture in this conference, is to review the main tools that we have developed to estimate neural parameters from mass signals, and to outline future challenges and directions for developing computational tools to invert aggregate neural signals in terms of neural circuit parameters.

In this paper, we propose a biologically plausible model for learning the decision-making sequence in an external environment with internal motivation. As a computational model, we propose a hierarchical architecture of an intelligent agent acquiring experience based on reinforcement learning. We use the basal ganglia model to aggregate a reward, and sparse distributed representation of states and actions in hierarchical temporal memory elements. The proposed architecture allows the agent to build a compact model of the environment and to form an effective strategy, which is experimentally demonstrated to search for resources in grid environments.

Tacit coordination games are games in which players need to coordinate with one another, for example, on how to divide resources, while they are not allowed to communicate with each other. In divergent interest tacit coordination games, their interests are not always aligned. For instance, player may need to choose between a solution that maximizes their individual profit or a solution that is perceptually more salient to both players, i.e., a focal point, that will increase the chances for successful coordination. The goal of this study was to examine the effect of two key variables, the Expected Revenue Proportions (ERP) and the player's Social Value Orientation (SVO) on the probability of realizing a focal point solution in divergent interest tacit coordination games. Our results show that there is an interaction between the expected payoff and the SVO. For example, prosocial players tend to implement a social point solution although the expected payoff is less than that of their opponent. Thus, the implementation of a focal point depends on other contextual variables such as the SVO and the expected payoff. The main contribution of this work is showing that the probability to choose a focal point solution is affected by the interaction between SVO and the expected revenue of the player. This finding may contribute to the construction of cognitive models for decision making in diverge interest tacit coordination problems.

The lateral geniculate nucleus (LGN) plays a vital role in visual information processing as an early stage in the visual system linking the retina with the visual cortex. Beyond its simple linear role, the LGN has been found to have a complex role in higher-order visual processing that is not fully understood. The aim of this study is to examine predicting high-level visual features from rat LGN firing activity. Extracellular neural activity of LGN neurons was recorded from 6 anesthetized rats in response to 4 × 8 checkerboard visual stimulation patterns using multi-electrode arrays. The first examined high-level feature is classifying the positions of the majority of white pixels in a visual pattern using the corresponding LGN activity. Three classes of patterns are identified in this task: majority in the top two rows, majority in the bottom two rows, or equal number of white pixels across the top and bottom halves of the pattern. The second examined high-level feature is estimating how far the white pixels are scattered in a visual stimulation pattern based on the corresponding LGN activity. Our results demonstrate that using LGN population activity achieves an $${F}_{1}$$ F 1 -score of 0.67 in the patterns classification and a root-mean-square error of 0.3 in the scatter estimation. Such performance outperforms that achieved using the visual stimulation patterns as inputs to the classification and scatter estimation methods. These results provide evidence that specific high-level visual features could be represented in the LGN; suggesting a critical role of the LGN in encoding visual information.

The analysis of the different varieties of language alterations from several causes has become an indicator to support tentative diagnoses, not only physical but degenerative, functional or cognitive. In this study, we explore fluency-disfluency in language of participants after suffering a traumatic brain injury. From a linguistic-computational approach, covering one-year of periodic post-recovery stages samples, candidate subsets of features were evaluated with a pool of learning methods until obtaining comparable scores to a baseline taken as the maximums achieved with the same evaluation, but on the full feature set. Starting in three-months recovery stage, this was extended to six, nine, and twelve months. After setting a global overview during this period of the fluency response based on F1-score of the learning algorithms, the identified feature was the basis to work on a model in a longitudinal sense of the disfluency-response with dichotomous global linear mixed effects model.

Understanding the brain function requires investigating information transfer across brain regions. Shannon began the remarkable new field of information theory in 1948. It basically can be divided into two categories: directed and undirected information-theoretical approaches. As we all know, neural signals are typically nonlinear and directed flow between brain regions. We can use directed information to quantify feed-forward information flow, feedback information, and instantaneous influence in the high-level visual cortex. Moreover, neural signals have bidirectional information flow properties and are not captured by the transfer entropy approach. Therefore, we used directed information to quantify bidirectional information flow in this study. We found that there has information flow between the scene-selective areas, e.g., OPA, PPA, RSC, and object-selective areas, e.g., LOC. Specifically, strong information flow exists between RSC and LOC. It explained that functionally coupled between RSC and LOC plays a vital role in visual scenes/object categories or recognition in our daily lives. Meanwhile, we also found weak reverse-directed information flow in the visual scenes and objects neural networks.

This paper describes a network model for mental processes making use of shared mental models (SMM) of team performance. The paper illustrates the value of adequate SMM’s for safe and efficient team performance. The addressed application context is that of a medical team performing a tracheal intubation executed by a nurse and a medical specialist. Simulations of successful and unsuccessful team performance have been performed, some of which are presented. The paper discusses potential further elaborations for future research as well as implications for other domains of team performance.

Neuroscience lies at the heart of a wide range of novel scientific and technological advances, from neurotechnology to computing and the cognitive sciences. Observing and studying the human brain allows glimpses into the most valuable resource resulting from evolution: the human mind. However abundant and widely available, neuroscience research data are knowledge and resource intensive to acquire and maintain, demanding a high level of skill and expertise to be produced, accessed and applied. As in all complex knowledge domains, reasoning with neuroscience relies on implicit metacognitive processes, and metacognition relies on high level cognitive models. This paper proposes a System Level Knowledge Representation schema as an abstraction for neuroscience data aimed at supporting explicit metacognition as well as knowledge organisation, understandability and to ease the cognitive load of processing systems neuroscience data.

Recent developments in neurotechnology effectively utilize the decades of neuroscientific findings of multiple meditation techniques. Meditation is linked to higher-order cognitive processes, which may function as a scaffold for cognitive control. In line with these developments, we analyze oscillatory brain activities of expert and non-expert meditators from the Himalayan Yoga tradition. We exploit four dimensions (Temporal, Spectral, Spatial and Pattern) of EEG data and present an analysis pipeline employing machine learning techniques. We discuss the significance of different frequency bands in relation with distinct primary 5 scalp brain regions. Functional connectivity networks (PLV) are utilized to generate features for classification between expert and non-expert meditators. We find (a) higher frequency $$\beta $$ β and $$\gamma $$ γ oscillations generate maximum discrimination over the parietal region whereas lower frequency $$\theta $$ θ and $$\alpha $$ α oscillations dominant over the frontal region; (b) maximum accuracy of over 90% utilizing features from all regions; (c) Quadratic Discriminant Analysis surpasses other classifiers by learning distribution for classification. Overall, this paper contributes a pipeline to analyze EEG data utilizing various properties and suggests potential neural markers for an expert meditative state. We discuss the implications of our research for the advancement of personalized headset design that rely on feedback on depth of meditation by learning from expert meditators.

Exploring information processing mechanisms in the human brain is of significant importance to the development of artificial intelligence and translational study. In particular, essential functions of the brain, ranging from perception to thinking, are studied, with the evolution of analytical strategies from a single aspect such as a single cognitive function or experiment to the increasing demands on the multi-aspect integration. Here we introduce a systematic approach to realize an integrated understanding of the brain mechanisms with respect to cognitive functions and brain activity patterns. Our approach is driven by a conceptual brain model, performs systematic experimental design and evidential type inference that are further integrated into the method of evidence combination and fusion computing, and realizes never-ending learning. It allows comparisons among various mechanisms on a specific brain-related disease by means of machine learning. We evaluate its ability from the brain functional connectivity perspective, which has become an analytical tool for exploring information processing of connected nodes between different functional interacting brain regions, and for revealing hidden relationships that link connectivity abnormalities to mental disorders. Results show that the potential relationships on clinical signs–cognitive functions–brain activity patterns have important implications for both cognitive assessment and personalized rehabilitation.

Electrical measurements of aggregate neural activity, such as local field potentials (LFPs) or electroencephalograms (EEGs), can capture oscillations of neural activity over a wide range of frequencies and are widely used to study brain function and dysfunction. However, relatively little is known about how to relate features of such aggregate neural recordings to the functional and anatomical configurations of the underlying neural circuits that produce them. An important neural circuit parameter which has profound effects on neural network dynamics and neural function is the ratio between excitation and inhibition (E:I), which has been found to be atypical in many neuropsychiatric conditions. Here we used simulations of recurrent networks of point-like leaky integrate-and-fire (LIF) neurons to study how to infer parameters such as the E:I ratio or the magnitude of the external input of the network from aggregate electrical measures. We used approximations (or proxies), validated in previous work, to generate realistic LFPs and EEGs from simulations of such networks. We computed different spectral features from simulated neural mass signals, such as the 1/f spectral power law or the Hurst exponent (H), and studied how these features changed when we changed the E:I ratio or the strength of the external input of the network model. We discuss how different spectral features of aggregate signals relate to the E:I ratio or the strength of the external input and outline our efforts to fit our model, in future work, to multiple measures extracted from empirical recordings of aggregate neural activity.

Considering the recent rapid advancements in digital technology, electroencephalogram (EEG) signal is a potential candidate for a robust human biometric authentication system. In this paper the focus of investigation is the use of brain activity as a new modality for identification. Univariate model biometrics such as speech, heart sound and electrocardiogram (ECG) require high-resolution computer system with special devices. The heart sound is obtained by placing the digital stethoscope on the chest, the ECG signals at the hands or chest of the client and speaks into a microphone for speaker recognition. It is challenging task when adapting these technologies to human beings. This paper proposed a series of tasks in a single paradigm rather than having users perform several tasks one by one. The advantage of using brain electrical activity as suggested in this work is its uniqueness; the recorded brain response cannot be duplicated, and a person’s identity is therefore unlikely to be forged or stolen. The disadvantage of applying univariate is that the process only includes correlation in time precedence of a signal, while the correlation between regions is ignored. The inter-regional could not be assessed directly from univariate models. The alternative to this problem is the generalization of univariate model to multivariate modeling, hypothesized that the inter-regional correlations could give additional information to discriminate between brain conditions where the models or methods can measure the synchronization between coupling regions and the coherency among them on brain biometrics. The key issue is to handle the single task paradigm proposed in this paper with multivariate signal EEG classification using Multivariate Autoregressive (MVAR) rather than univariate model. The brain biometric systems obtained a significant result of 95.33% for dynamic Vector autoregressive (VAR) time series and 94.59% for Partial Directed Coherence (PDC) and Coherence (COH) frequency domain features.

In this paper, we introduce a novel model for establishing brain functional connectivity based on noninvasive Electroencephalogram (EEG) data sources. We reviewed the main methods used in EEG brain functional connectivity, and the current research progress of analyzing EEG datasets. In this paper, we proposed a new model for bridging the missing link between human brain functions and real time brain wave activities. The proposed model combines graph theory/complex network methods with fuzzy logic method to deliver an explicit connection in a real time environment. We conducted the EEG data preprocessing experiments for our new model.

In online interactions, users frequently add emojis (e.g., smileys, hearts, angry faces) to text for expressing the emotions behind the communication context, aiming at a better interpretation to text especially of polysemous short expressions. Emotion recognition refers to the automated process of identifying and classifying human emotions. If text-based emoticons (i.e., emojis created by textual symbols and characters) can be directly understood by semantic-based context recognition tools used in the Web and Artificial Intelligence and robotics, image-based emojis need instead image recognition for a complete semantic context interpretation. This study aims to explore and compare systematically different classification models of emoticon pictograms collected from the Internet, with different labels according to the Ekman model of six basic emotions. A first comparison involves supervised machine learning classifiers trained on features extracted through neural networks. In the second phase, the comparison is extended to different deep learning models. Results indicate that deep learning models performed excellent, and traditional supervised algorithms also achieve very promising outcomes.

Face detection and sparse facial feature analysis is popular as a non-invasive approach to diagnosis special disease. In futuristic intelligent healthcare system, the confined way of preliminary computer aided diagnosis of diseases becoming more inclusive and faster than usual time. Therefore, face spacial feature analysis can be an elegant way of measuring attempt in tele-medicine industry. In this research paper, we investigate thorough review on disease diagnosis techniques, healthcare management and, data security features being used currently. Moreover, this work propose a i-health care monitoring and examining system of neuronal/brain disorder in layer base approach. Overall, this paper reviews about diseases which have already been detected by spacial feature of face using deep learning algorithm or feature based learning with a proposal of a monitoring system with its research area and challenges in smart intelligent healthcare system in society 5.0.

Over the past year, distance learning has become an integral part of our lives and a rapidly developing field. Compared to traditional classroom learning, online learning has advantages such as no place restrictions and a wide range of interactions. At the same time, distance learning lacks interaction between teachers and learners, and teachers cannot observe learners face to face. Identifying students’ emotions during distance learning can have a positive impact on learning, improve learning outcomes and improve the effectiveness, quality of learning. This paper describes a theoretical model for determining sentiment/emotion from audio data based on speech recognition. The recognition model based on generalized transcription was proposed. This method can be used to determine the student’s emotions during the distance learning and online exam.

Emotion recognition has been extensively studied in a single modality in the last decade. However, humans express their emotions usually through multiple modalities like voice, facial expressions, or text. This paper proposes a new method to learn a joint emotion representation for multimodal emotion recognition. Emotion-based feature for speech audio is learned by an unsupervised triplet-loss objective, and a text-to-text transformer network is used to extract text embedding for latent emotional meaning. Transfer learning provides a powerful and reusable technique to help fine-tune emotion recognition models trained on mega audio and text datasets respectively. The extracted emotional information from speech audio and text embedding are processed by dedicated transformer networks. The alternating co-attention mechanism is used to construct a deep transformer network. Multimodal fusion is implemented by a deep co-attention transformer network. Experimental results show the proposed method for learning a joint emotion representation achieves good performance in multimodal emotion recognition.

From 20 March to 10 May 2020, the “stay at home” countermeasures for the Covid-19 emergency lockdown were defined in the United Kingdom (UK) as leaving home for only the following reasons: “Key worker travelling to work”, “Shopping for basic necessities”, “Any medical need” or “Exercise once a day”. Data collected from the UK Office for National Statistics through online and telephone questionnaires are an exceptional baseline data set on people behaviour during the Covid-19 pandemics. In this paper, data from demographic surveys from the UK are compared to statistical and feedback data from the Virtual Reality app called TRIPP for meditation in the experiences called Focus and Calm. Our data analysis shows that during lockdown the psychological and emotional mindset, severely challenged, has been successfully enhanced with the use of Virtual Reality.

As one of the types of mobile social media, short video has a huge number of users in China, which is over 873 million, accounting for 88.3% of the total number of Internet users by December 2020. The purpose of this study is to understand the emotional and psychological state of users in process of using short video and improve the subjective well-being level of the users. Methods: The Study conducted in-depth interviews with 12 short video users to explore the factors that affect the subjective well-being of short video users. Results: (1) The influential factors affecting the subjective well-being of short video users in the process of using social media usage were information overload and social pressure in environmental factors; physical condition, emotional state, aesthetic fatigue, social comparison, fear of missing out, self-expectation and self-control in personal factors; passive use and diving factors in behavioral factors; (2) It was further generalized that boredom proneness and time management proneness are the two most important factors that affect users’ subjective well-being. Conclusion: The most important factors influencing the subjective well-being of short video users are boredom proneness and time management proneness.

Recent studies have shown that brain lesions following stroke can be probabilistically mapped onto disconnections of white matter tracts, and that the resulting “disconnectome” is predictive of the patient’s behavioral deficits. Disconnectome maps are sparse, high-dimensional 3D matrices that require unsupervised dimensionality reduction followed by supervised learning for prediction of the associated behavioral data. However, the optimal machine learning pipeline for disconnectome data still needs to be identified. We examined four dimensionality reduction methods at varying levels of compression and used the extracted features as input for cross-validated regularized regression to predict the associated language and motor deficits. Features extracted by Principal Component Analysis and Non-Negative Matrix Factorization were found to be the best predictors, followed by Independent Component Analysis and Dictionary Learning. Optimizing the number of extracted features improved predictive accuracy and greatly reduced model complexity. Moreover, the choice of dimensionality reduction technique was found to optimally combine with a specific type of regularized regression (ridge vs. LASSO). Overall, our findings represent an important step towards an optimal pipeline that yields high prediction accuracy with a small number of features, which can also improve model interpretability.

Vast amounts of knowledge resources are emerging from Neuroscience research, thanks to increasingly widely available imaging and analysis technologies and open data sets. Learning, processing and keeping up to date with developments however imply steep learning curves. Advances in Neuroscience research provide insights into the human brain and the mind for brain specialists, but also inform other scientific disciplines such as to complexity science, cognitive computing, medicine in general and general technology and policy sectors. With a few exceptions however, the majority of Neuroscience research outcomes can still be accessed and leveraged mostly only by highly trained neuroscientists and brain informatics researchers in their respective specialisations, as the knowledge and skill sets required to query and manipulate such data is only meaningful for individuals with specific training. This paper identifies and addresses the need to lower the cognitive barriers to accessing Neuroscience research, as it is becoming very relevant to other fields, and to widen its accessibility to a broader range of scholars of other disciplines - such as Computer Science and Information Technology for example, reducing the efforts required in tracking innovation and facilitate knowledge acquisition. Keeping up with the state of the art is a challenge in any field, yet increasing number of researchers, students and practitioners from diverse professions and with a wide range of interests and goals, and multi disciplinary and linguistic backgrounds. The approach proposed here leverages a combination of elementary core methods from semantic technology, including simple corpus and linguistic analysis techniques, and devises a low tech instrument that can be adopted irrespective of the availability of software and level of English language proficiency to acquire the necessary familiarity to handle Neuroscience topics The same method can also be leveraged in other complex knowledge domains. A set of experiments to evaluate the effectiveness of the method is described with preliminary results.

Study about brain connectivity provides important bio-markers for predicting brain related disorders and also for analyzing normal human functions. Findings of this study are reported in the form of neuroscience research articles. We propose a tool, ConnExt-BioBERT, to mine relevant scientific literature for curating a large resource of reported connections between regions of the brain. We have utilized the popular transfer learning technique that has been trained on large datasets, the Bidirectional Encoder Representations for Transformers (BERT) to apply it to a narrowband subject area of extracting brain regions and potential connection mentions from a set of 53,000 full-text neuroscience articles (53kNeuroFullText) indexed on PubMed. Evaluation of ConnExt-BioBERT has been performed on a benchmark dataset of abstracts and on a dataset of seven full-text articles annotated by a domain expert. Additionally, connections retrieved by the tool on 53kNeuroFullText have been evaluated using a manually curated resource, Brain Architecture Management System (BAMS). A web-application has been developed for search over extracted brain region connections on 53kNeuroFullText. This application is currently being used by neuroscience researchers to quickly retrieve brain connectivity information reported by various authors. Large scale text mining of brain-connectivity information reported in neuroscience literature, aids in progressing research in the area of neurological disorders and further helps diagnosis and treatment of the same.

In this digital age, social media is an essential part of life. People share their moments and emotions through it. Consequently, detecting emotions in their behavior can be an effective way to determine their emotional disposition, which can then be used to control their negative thinking by making them see the positive aspects of the world. This study proposes an emotion detection-based mood control framework that reorganizes social media posts to match the user’s mental state. An emotion detection model based on Attention mechanism, Bidirectional Long Short Term Memory (LSTM), and Convolutional Neural Network (CNN) has been proposed which can detect six emotions from Bangla text with 66.98% accuracy. It also demonstrates how emotion detection frameworks can be implemented in other languages as well.

Neuroscience is an important area of research due to the nature of the brain and its diseases. Scientists in this field tend to ask complicated questions which are time-consuming to answer and need several resources. Analysing, representing and finally, classifying these questions assist question resolution systems to be able to tackle them more easily.To achieve its objectives, this study contains three different tasks, including an ontology-based question analysis approach to find question dimensions for representing questions and shaping categories for them; and two approaches in classifying questions, including one ontology-based and a set of statistical approaches.

Visual, audio, and emotional perception by human beings have been an interesting research topic in the past few decades. Electroencephalography (EEG) signals are one of the ways to represent human brain activity. It has been shown, that different brain networks correspond to processes corresponding to varieties of emotional stimuli. In this paper, we demonstrate a deep learning architecture for the movie identification task from the EEG response using Convolutional Neural Network (CNN). The dataset includes nine movie clips that span across different emotional states. The EEG time series data has been collected for 20 participants. Given one second EEG response of particular participant, we tried to predict its corresponding movie ID. We have also discussed the various pre-processing steps for data cleaning and data augmentation process. All the participants have been considered in both train and test data. We obtained 80.22% test accuracy for this movie classification task. We also tried cross participant testing using the same model and the performance was poor for the unseen participants. Our result gives insight toward the creation of identifiable patterns in the brain during audiovisual perception.

Identifying unique descriptors for primary colours in EEG signals will open the way to Brain-Computer Interface (BCI) systems that can control devices by exposure to primary colours. This study is aimed to identify such unique descriptors in visual evoked potentials (VEPs) elicited in response to the exposure to primary colours (RGB: red, green, and blue) from 31 subjects. For that, we first created a classification method with integrated transfer learning that can be suitable for an online setting. The method classified between the three RGB classes for each subject, and the obtained average accuracy over 23 subjects was 74.48%. 14 out of 23 subjects were above the average level and the maximum accuracy was 93.42%. When cross-session transfer learning was evaluated, 86% of the subjects tested showed an average variation of 5.0% in the accuracy comparing with the source set.

A person’s state of attentiveness can be affected by various outside factors. Having energy, feeling tired, or even simply being distracted all play a role in someone’s level of attention. The task at hand can potentially affect the person’s attention or concentration level as well. In terms of students who take online courses, constantly watching lectures and conducting these courses solely online can cause lack of concentration or attention. Attention can be considered in two categories: passive or active. Conducting active and passive attention-based trials can reveal different states of attentiveness. This paper compares active and passive attention trial results of the two states, wide awake and tired. This has been done in order to uncover a difference in results between the two states. The data analyzed throughout this paper was collected from DSI 24 EEG equipment, and the generated EEG is processed through a 3D Convolutional Neural Network (CNN) to produce results. Three passive attention trials and three active attention trials were performed on seven subjects, while they were wide awake and when they were tired. The experiments on the preprocessed data results in accuracies as high as 81.78% for passive attention detection accuracy and 63.67% for active attention detection accuracy, which shown a clear ability to separate between the two attention categories.

Topic modeling based on neuroimaging literature is an important approach to aggregate world-wide research findings for decoding brain cognitive mechanism, as well as diagnosis and treatment of brain and mental diseases, artificial intelligence researches, etc. However, existing neuroimaging literature mining only focused on texts and neglects brain images which contain a large amount of topic information. Following the writing and reading habits combining images with texts, we present in this paper an image-enhanced LDA (Latent Dirichlet Allocation), which extracts literature topics from both neuroimaging images and full texts. Combining topics from fMRI brain regions activation images with topics from full texts to model neuroimaging literatures more accurately. On the one hand, topics related brain cognitive mechanism can be pertinently extracted from activated brain images and their descriptions. On the other hand, topics from activated brain images can be integrated with topics from full text to model neuroimaging literature more accurately. The experiments based on actual data has preliminarily proved effectiveness of proposed method.

Individualized treatment is crucial for epileptic patients with different types of seizures. The difference among patients impacts the drug choice as well as the surgery procedure. With the advance in machine learning, automatic seizure detection could ease the manual time-consuming and labour-intensive procedure for diagnose seizure in the clinical setting. In this paper, we propose a electroencephalography frequency bands selection method that exploits Natural Language Processing (NLP) features from individual’s condition and patients with same seizure types. We used Temple University Hospital (TUH) EEG seizure corpus and conducted experiments with various input data for different seizure types classified using Random Forest (RF) and Support Vector Machine (SVM). The results show that with reduced frequency bands the performance slightly deviates from the whole frequency bands, thus leading to possible resource-efficient implementation for seizure detection.

Programmers are the most important part of software production and individual developers are hard to substitute. The essential part of the knowledge intensive development process is the developers mind state. Understanding the mental states of software developers has become a main interest of software production companies since it is the most valuable resource for software development. However the main challenge in analysing the software developers mental states is that most precise equipment, such as fMRI, is extremely expensive and not portable. Thus, fMRI approximation from EEG readings tools such as MNE, have been developed over the years. The idea of recreating the fMRI based on EEG signal is the main motivation for the current work. This research explains how we used this tool in our studies.

Although automatic prediction of Alzheimer’s disease (AD) from Magnetic Resonance Imaging (MRI) showed excellent performance, Machine Learning (ML) algorithms often provide high accuracy at the expense of interpretability of findings. Indeed, building ML models that can be understandable has fundamental importance in clinical context, especially for early diagnosis of neurodegenerative diseases. Recently, a novel interpretability algorithm has been proposed, the Explainable Boosting Machine (EBM), which is a glassbox model based on Generative Additive Models plus Interactions GA2Ms and designed to show optimal accuracy while providing intelligibility. Thus, the aim of present study was to assess – for the first time – the EBM reliability in predicting the conversion to AD and its ability in providing the predictions explainability. In particular, two-hundred brain MRIs from ADNI of Mild Cognitive Impairment (MCI) patients equally divided into stable (sMCI) and progressive (pMCI) were processed with Freesurfer for extracting twelve hippocampal subfields volumes, which already showed good AD prediction power. EBM models with and without pairwise interactions were built on training set (80%) comprised of these volumes, and global explanations were investigated. The performance of classifiers was evaluated with AUC-ROC on test set (20%) and local explanations of four randomly selected test patients (sMCIs and pMCIs correctly classified and misclassified) were given. EBMs without and with pairwise interactions showed accuracies of respectively 80.5% and 84.2%, thus demonstrating high prediction accuracy. Moreover, EBM provided practical clinical knowledge on why a patient was correctly or incorrectly predicted as AD and which hippocampal subfields drove such prediction.

The neural recordings in the form of local field potentials offer useful insights on higher-level neural functions by providing information about the activation and deactivation of neural circuits. But often these recordings are contaminated by multiple internal and external sources of noise from nearby electronic systems and body movements. However, to facilitate knowledge extraction from these recordings, identification and removal of the artefacts are empirical, and various computational techniques have been applied for this purpose. Here we report a new module for artefact removal, an extension of the toolbox named SANTIA (SigMate Advanced: a Novel Tool for Identification of Artefacts in Neuronal Signals) which allows for fast application of deep learning techniques to remove said artefacts without relying on data from other channels.

Alzheimer’s disease, an incurable chronic neurological disorder (NLD) that affects human memory and demises cognitive thinking ability with shrinkage of the brain area. Early detection of Alzheimer’s disease (AD) is the only hope to delay its effect. This study designed a computer-aided automated detection method that can detect mild cognitive impairment for AD from magnetic resonance image scans. The data-driven solution approach requires an extensive quantity of annotated images for diagnosis. However, obtaining a large amount of annotated data for medical application is a challenging task. We have exploited a deep convolutional generative adversarial network (DCGAN) for synthesizing high-quality images to increase dataset size. A fine-tuned CNN (VGG16 architecture) model works on images to extract the intuitive features for early diagnosis. The extracted features of images by VGG16 feed into the support vector machine for classification. This research has conducted copious experiments to validate the proposed method outperformed relative baselines on public datasets.

Mental health is an important aspect of an individual’s well-being which still continues to remain unaddressed. With the rise of the COVID-19 pandemic, mental health has far continued to decline, especially amongst the younger generation. The aim of this research is to raise awareness about mental health while simultaneously working towards removing the societal stigma surrounding it. Thus, in this paper, we have created an integrated chatbot that is specifically geared towards mentally ill individuals. The chatbot responds empathetically which is built using a Sequence-to-Sequence (Seq2Seq) encoder-decoder architecture. The encoder uses Bi-directional Long Short Term Memory (BiLSTM). To compare the performance, we used Beam Search and Greedy Search. We found Beam Search decoder performs much better, providing empathetic responses to the user with greater precision in terms of BLEU score.

Parkinson’s disease is a neurological disorder. It affects the structures of the central and peripheral nervous system that control movement. One of the symptoms of Parkinson’s disease is difficulty in speaking. Hence, analysis of speech signal of patients may provide valuable features for diagnosing. Previous works on diagnosis based on speech data have employed machine learning and deep learning techniques. However, these approaches do not address the various uncertainties in data. Belief rule based expert system (BRBES) is an approach that can reason under various forms of data uncertainty. Thus, the main objective of this research is to compare the potential of BRBES on various speech signal features of patients of parkinson’s disease. The research took into account various types of standard speech signal features such MFCCs, TQWTs etc. A BRBES was trained on a dataset of 188 patients of parkinson’s disease and 64 healthy candidates with 5-fold cross validation. It was optimized using an exploitive version of the nature inspired optimization algorithm called BRB-based adaptive differential evolution (BRBaDE). The optimized model performed better than explorative BRBaDE, genetic algorithm and MATLAB’s FMINCON optimization on most of these features. It was also found that for speech based diagnosis of Parkinson’s disease under uncertainty, the features such as Glottis Quotient, Jitter variants, MFCCs, RPDE, DFA and PPE are relatively more suitable.

Autism spectrum disorder (ASD) is a neuro-developmental disease that has a lifetime impact on a person’s ability to interact and communicate with others. Early discovery of autism can assist to prepare a plan for suitable therapy and reduce its impact on patients at an appropriate time. The aim of this work is to propose a machine learning model which generates autism subtypes and identifies discriminatory factors among them. In this work, we use Quantitative Checklist for Autism in Toddlers-10 (Q-CHAT-10) of toddler and Autism Spectrum Quotient-10 (AQ-10) datasets of child, adolescent, and adult screening datasets respectively. Then, only autism records are merged and implemented k-means algorithm to extract various autism subtypes. According to Silhoutte score, we select the best autism dataset and balance its subtypes using random oversampling (ROS) and synthetic minority oversampling technique for numeric and categorical values (SMOTENC). Afterwards, various classifiers are employed into both primary dataset and its balanced subtypes. In this work, logistic regression shows the highest result for primary dataset. Also, it achieves the greatest results for ROS and SMOTENC datasets. Hence, shapely adaptive explanation (SHAP) technique is used to rank features and scrutinized discriminatory factors of these autism subtypes.

A handheld device (such as a smartphone/wearable) can be used for tracking and delivering navigation within a building using a wireless interface (such as WiFi or Bluetooth Low Energy), in situations when a traditional navigation system (such as a global positioning system) is unable to function effectively. In this paper, we present an indoor navigation system based on a combination of wall-mounted wireless sensors, a mobile health application (mHealth app), and WiFi/Bluetooth beacons. Such a system can be used to track and trace people with neurological disorders, such as Alzheimer’s disease (AD) patients, throughout the hospital complex. The Contact tracing is accomplished by using Bluetooth low-energy beacons to detect and monitor the possibilities of those who have been exposed to communicable diseases such as COVID-19. The communication flow between the mHealth app and the cloud-based framework is explained elaborately in the paper. The system provides a real-time remote monitoring system for primary medical care in cases where relatives of Alzheimer’s patients and doctors are having complications that may demand medical care or hospitalization. The proposed indoor navigation system has been found to be useful in assisting patients with Alzheimer’s disease (AD) while in the hospital building.

This paper proposes a parallel machine learning framework for detecting Alzheimer’s disease through T1-weighted MRI scans localised to the hippocampus, segmented between the left and right hippocampi. Feature extraction is first performed by 2 separately trained, unsupervised learning based AutoEncoders, where the left and right hippocampi are fed into their respective AutoEncoder. Classification is then performed by a pair of classifiers on the encoded data from the AutoEncoders, to which each pair of the classifiers are aggregated together using a soft voting ensemble process. The best averaged aggregated model results recorded was with the Gaussian Naïve Bayes classifier where sensitivity/specificity achieved were 80%/81% respectively and a balanced accuracy score of 80%.

Parkinson’s disease (PD) is a neurological illness that occurs by the degeneration of cells in the nervous system. Early symptoms include tremors or involuntary movements of the hands, arms, legs, and jaw. Currently, the only method to diagnose PD involves the observation of its prodromal symptoms. Moreover, detecting handwriting will work as a variable for clinitians to understand PD in patients better. With the advancement of technology, it is possible to build applications that will aid in diagnosing PD without any clinical intervention. The majority suffering from PD have handwriting abnormalities (referred to as micrographia), which is the most reported among earlier signs of the disease. So this research is undertaken by focusing on the implication of micrographia. For this purpose, handwritten images are collected from a group of 136 PD patients and 36 healthy patients. These images form a dataset of 800 images that are used to train a model which will accurately classify PD patients. To achieve this transfer learning is chosen because of its ability to produce accurate results regardless of the limited size of the dataset. Here, different models of transfer learning are trained to figure out the well-fitting model. It was observed that VGG-16 performed adequately with a training accuracy of 90.63% while a testing accuracy of 91.36%.

With the rapid growth of the Internet of Healthcare Things, a massive amount of data is generated by a broad variety of medical devices. Because of the complex relationship in large-scale healthcare data, researchers who bring a revolution in the healthcare industry embrace Artificial Intelligence (AI). In certain cases, it has been reported that AI can do better than humans at performing healthcare tasks. The data-driven black-box model, on the other hand, does not appeal to healthcare professionals as it is not transparent, and any biasing can hamper the performance the prediction model for the real-life operation. In this paper, we proposed an AI model for early detection of autism in children. Then we showed why AI with explainability is important. This paper provides examples focused on the Autism Spectrum Disorder dataset (Autism screening data for toddlers by Dr Fadi Fayez Thabtah) and discussed why explainability approaches should be used when using AI systems in healthcare.

Epilepsy is a common neurological disease characterized by recurrent seizures. Electroencephalography (EEG), which records neural activity, is commonly used to diagnose epilepsy. This paper proposes an Empirical Mode Decomposition (EMD) and Deep Convolutional Neural Network epileptic seizure prediction method. First, the original EEG signals are segmented using 30 s sliding windows, and the segmented EEG signal is decomposed into Intrinsic Mode Functions (IMF) and residuals. Then, the entropy features which can better express the signal are extracted from the decomposed components. Finally, a deep convolutional neural network is used to construct the epileptic seizure prediction model. This experiment was conducted on the CHB-MIT Scalp EEG dataset to evaluate the performance of our proposed EMD-CNN epileptic EEG seizure detection model. The experimental results show that, compared with some previous EEG classification models, this model is helpful to improving the accuracy of epileptic seizure prediction.

Electroencephalography (EEG) classification is an important part in brain-computer interface system. Motor imagery is a novel experimental paradigm that has been proved effective clinically in recognizing EEG from different limb motions. Our object is to finish motor imagery based EEG classification. Due to EEG signals followed by some features, e.g. noisy, weak signal, personalization and so on, traditional methods could encounter limit from the single feature extraction. In this work, we propose a multi-scale spatio-temporal features fusion deep learning model. Given raw EEG signals, we calculate mutual information matrix among different channels. It incorporates spatio-temporal feature extraction and mutual information matrix. We deploy experiments on two datasets that consists of the High Gamma Dataset and BCI IV 2A dataset. Experiment results show that the proposed temporal spatial convolution neural network fused with mutual information model outperform other methods.

Magnetic Resonance Imaging (MRI) is used extensively for the diagnosis of Alzheimer’s Disease (AD). Early detection of AD can help people with early intervention and alleviate the progression of disease symptoms. Previous studies have applied deep learning methods for computer-aided diagnosis of AD. In this present study, an efficient architecture has been proposed, composed of a 2D Convolutional neural network with batch normalization for the classification of AD using MRI images. The proposed model was created using 11 layers, which was obtained by experimenting with different combinations of batch normalization and activation functions. All the experiments are performed using the Alzheimer Disease Neuroimaging Initiative (ADNI) data. The novelty of our approach was that different slices of the brain, such as axial, coronal, and sagittal, were used to classify brain slices into three classes: Cognitively Normal (NC), Mild Cognitive Impairment (MCI), and AD. The proposed model achieved a sensitivity (SEN) of 99.73% for NC, 99.79% for MCI, and 99.96% for AD, a specificity (SPE) of 99.80% for NC, 99.90% for MCI, and 99.74% for AD, and accuracy of 99.82%. The contribution of our proposed method’s classification accuracy was better than that of the recent state-of-the-art methods.

Parkinson’s disease (PD) is a kind of neurodegenerative disorder characterized by the loss of dopamine-producing cells in the brain. The disruption of brain cells that create dopamine, a chemical that allows brain cells to connect with one another, causes Parkinson’s disease. Control, adaptability, and rapidity of movement are all controlled by dopamine-producing cells in the brain. Researchers have been investigating for techniques to identify non-motor symptoms that show early in the disease as soon as possible, slowing the disease’s progression. A machine learning-based detection of Parkinson’s disease is proposed in this research. Feature selection and classification techniques are used in the proposed detection technique. Boruta, Recursive Feature Elimination (RFE) and Random Forest (RF) Classifier have been used for the feature selection process. Four classification algorithms are considered to detect Parkinson disease which are gradient boosting, extreme gradient boosting, bagging and Extra Tree Classifier. Bagging with recursive feature elimination was found to outperform the other methods. The lowest number of voice characteristics for the diagnosis in Parkinson attained 82.35% accuracy.

Brain-Computer Interfaces (BCI) is one of the alluring breakthroughs for mankind as it provides a new way of communication for the patients of neuro-muscular disorders. Electroencephalography (EEG) signals are the most studied type of signals to detect brain activities because of its non-invasive and portable nature. The major problem in the identification of neural activities from EEG signals and the presence of non-task related artifacts in the signal data. These artifacts affect the classification of feature set. With these effective techniques, BCI classifier can efficiently classify EEG signals. The proposed research deals with different motor imagery datasets for the detection of movements. An EEG based BCI system is proposed that implement a linear regression based artifact removal method for EOG processing, feature construction and recursive feature elimination with cross-validation. It achieved promising results with relatively fewer data used for training than the original competition’s data, that shows the significance as compared to top leaderboard entries. The results obtained show that our approach tackles noise and artifacts in EEG signals which provides reliable features for BCI classification.

The information analysis of the electroencephalogram (EEG) signal is carried out by granulation and reciprocal entropy (PeEn). The analysis of the EEG signal is obtained by experimental activity. Due to its complexity and multichannel characteristic, together with granular computing (GrC) and PeEn are used to analyze the EEG signal. The EEG signal consists of 32 channels of data and the experimental data are used to discriminate patterns, with experimental focus on considering real and thinking actions. The time-series EEG signals were granularized according to the changes in the signal and analyzed by PeEn coding and Fuzzy C-Means (FCM) algorithm. Because there are two main actions, i.e., left-handed, and right-handed actions were clearly delineated. In addition, we provide the GrC algorithm to prove the boundary problem with the help of Hilbert-Huang transform. The obtained results show an advanced approach for analyzing EEG signals, which can be the basis for solving complex multichannel data analysis.

Brainstorming has been used for many years and have a very good efforts, however, the brainstorm also mainly relies on human’s capability of brains, especially experience and knowledge they possessed. Based on the new discipline called Extenics, we proposed a testing method to explore the process of how ideas are bring out of brains, and help people think in multi dimensions and put forward more ideas. Extenics has been applied to study the extension and transformation of things in formalized models and obtain systematic creatives to solve contradictory problems intelligently since 1983. Support with Information technology and artificial Intelligence, we collect more Information and knowledge systematically to form basic elements based on Extenics using Human-Computer Interaction model to help people find more characteristics and its value of objectives. This will compensate the limited Information and knowledge in human brains. Also, we provide a methodology to help people thinking in multi dimensions positively according to the instruction of our methods based on extension innovation method. The case study proves its effectiveness in improving the capability of innovation of college students by EGG and Data Statistics.

Perkinson’s disease is a progressive degenerative disorder that comes from a recognized clinical parkinsonian syndrome. The manifestations of Parkinson’s disease include both motor and nonmotor symptoms identified as tremor, bradykinesia (slowed movements), rigidity, and postural instability. PD is marked as one of the most prevalent disorders from various researches and surveys because it has been observed in 90% of people out of 100. It is imperative to design CAD to develop an advanced model for the determination of this disease with accuracy since up to date there is no accurate clinical intervention for the diagnosis of PD. In contrast to conventional methods. Deep learning convolutional neural network tools are implied for the faster and accurate identification of PD through MRI. The purpose of this research is to contribute to the development of an accurate PD detection method. To conduct the research a public dataset NTU (National Technical University of Athens) is used. The data samples are categorized into three sets (Training, Test, and Validation). A DenseNet integrated with LSTM is applied to the MRI data samples. DenseNet is used to strengthen the feature selection ability, as each layer selects features depending on the temporal closeness of the image. The output is then fed into the LSTM layer, for discovering the significant dependencies in temporal features. The performance of the proposed DenseNet-LSTM is compared to other CNN state-of-the-art models. The proposed model outputs a training accuracy of 93.75%, testing accuracy of 90%, and validation accuracy of 93.8% respectively.

Automatic pose and shape initialization is the first step to conquer the problem of human tracking, acquiring prior knowledge about the tracking subject. It is crucial for accomplishing a successful tracking. In this paper, we present a simple and effective framework to automatically calibrate the human pose and shape by integrating a data driven shape parameterization into the skeletal animation pipeline and optimizing the template body model against multiview silhouette images to acquire the tracking subject shape and pose information. A PCA based approach is proposed to summarize the space of human body variations in height, weight, muscle tone, gender, body shape. Multiview analysis by synthesis optimization in the hierarchial order is employed to realize pose and shape calibration. Finally, experiments on HumanEvaII and Human3.6m dataset demonstrate our approach is very effective and robust to real world situations.

High variability between individual subjects and recording sessions is a known fact about scalp recorded EEG signal. While some do, the majority of the EEG based machine learning studies do not attempt to assess performance of algorithms across recording sessions or across subjects, instead studies use the whole data-set available for training and testing, using an established k-fold cross validation technique and thus missing performance in a real-life setting on an unseen subject. This study primarily aimed to show how important is to have a leave-one-subject-out (LOSO) evaluation done for any scalp recorded EEG based machine learning. This study also demonstrates effectiveness of a Multilayer Perceptron (MLP) in getting good LOSO accuracy from balanced, clean EEG data, without any pre-processing in comparison with traditional machine learning algorithms. The study used data from participants diagnosed with schizophrenia, as well as a group of participants with no known neurological disorder. Classification was done using traditional methods and MLP to classify the participants as belonging to disease or control subjects. Results shows that 85% accuracy on unseen subject was achievable from a clean data-set. MLP is seen to be effective in finding features by which schizophrenia could be detected from clean EEG data. LOSO evaluation done with this proven MLP configuration using carefully and intentionally corrupted data clearly indicate that for disease diagnosis, the k-fold classification result is misleading. Therefore, evaluation of any scalp recorded EEG based disease classification method must use a LOSO style cross-validation.