Combining Inter-Subject Modeling with a Subject-Based Data Transformation to Improve Affect Recognition from EEG Signals

Identifying relevant asymmetry features of EEG for emotion processing

The left and right hemispheres of the brain process emotion differently. Neuroscientists have proposed two models to explain this difference. The first model states that the right hemisphere is dominant over the left to process all emotions. In contrast, the second model states that the left hemisphere processes positive emotions, whereas the right hemisphere processes negative emotions. Previous studies have used these asymmetry models to enhance the classification of emotions in machine learning models. However, little research has been conducted to explore how machine learning models can help identify associations between hemisphere asymmetries and emotion processing. To address this gap, we conducted two experiments using a subject-independent approach to explore how the asymmetry of the brain hemispheres is involved in processing happiness, sadness, fear, and neutral emotions. We analyzed electroencephalogram (EEG) signals from 15 subjects collected while they watched video clips evoking these four emotions. We derived asymmetry features from the recorded EEG signals by calculating the log ratio between the relative energy of symmetrical left and right nodes. Using the asymmetry features, we trained four binary logistic regressions, one for each emotion, to identify which features were more relevant to the predictions. The average AUC-ROC across the 15 subjects was 56.2, 54.6, 51.6, and 58.4% for neutral, sad, fear, and happy, respectively. We validated these results with an independent dataset, achieving comparable AUC-ROC values. Our results showed that brain lateralization was observed primarily in the alpha frequency bands, whereas for the other frequency bands, both hemispheres were involved in emotion processing. Furthermore, the logistic regression analysis indicated that the gamma and alpha bands were the most relevant for predicting emotional states, particularly for the lateral frontal, parietal, and temporal EEG pairs, such as FT7-FT8, T7-T8, and TP7-TP8. These findings provide valuable insights into which brain areas and frequency bands need to be considered when developing predictive models for emotion recognition.

Subject-level Normalization to Improve A-phase Detection of Cyclic Alternating Pattern in Sleep EEG

Automatic detection systems for activation phases (A-phase) of the cyclic alternating pattern (CAP) in electroencephalograms (EEG) are designed to automatically score A-phases in any individual but typically fail to factor in EEG signal variations between individuals, e.g. due to sleep disorders, recording site differences or equipment differences. Here, we investigate the effect of subject-level normalization on the performance of an automatic A-phase detection system consisting of a recurrent neural network. We compared the classification performance of various subject-level normalization methods to the standard training set normalization. Systems were trained and tested on subjects with different sleep disorders using the publicly available CAP Sleep Database on Physionet. Subject-level normalization using Zscore or median and interquartile range (IQR) increases the F1-score for A1-phases by +11-22% (Z-Score: +11-20%, Median/IQR: +16-22%), for A2-phases by +2-9% (Z-Score: +59%, Median/IQR: +2-7%), for A3-phases by -1 - +8% (Z-Score: +3-8%, Median/IQR: -1-+5%) as compared to the standard training data normalization when tested across sleep disorders. Our results show that subject-level normalization drastically improves the precision of A-phase detection in case the training population differs from the testing population.Clinical Relevance- Subject-level normalisation improves the automatic CAP scoring system performances for the general population by minimizing the effect of individual EEG differences.

An attention-based deep learning approach for the classification of subjective cognitive decline and mild cognitive impairment using resting-state EEG

Objective. This study aims to design and implement the first deep learning (DL) model to classify subjects in the prodromic states of Alzheimer's disease (AD) based on resting-state electroencephalographic (EEG) signals.Approach. EEG recordings of 17 healthy controls (HCs), 56 subjective cognitive decline (SCD) and 45 mild cognitive impairment (MCI) subjects were acquired at resting state. After preprocessing, we selected sections corresponding to eyes-closed condition. Five different datasets were created by extracting delta, theta, alpha, beta and delta-to-theta frequency bands using bandpass filters. To classify SCDvsMCI and HCvsSCDvsMCI, we propose a framework based on the transformer architecture, which uses multi-head attention to focus on the most relevant parts of the input signals. We trained and validated the model on each dataset with a leave-one-subject-out cross-validation approach, splitting the signals into 10 s epochs. Subjects were assigned to the same class as the majority of their epochs. Classification performances of the transformer were assessed for both epochs and subjects and compared with other DL models.Main results. Results showed that the delta dataset allowed our model to achieve the best performances for the discrimination of SCD and MCI, reaching an Area Under the ROC Curve (AUC) of 0.807, while the highest results for the HCvsSCDvsMCI classification were obtained on alpha and theta with a micro-AUC higher than 0.74.Significance. We demonstrated that DL approaches can support the adoption of non-invasive and economic techniques as EEG to stratify patients in the clinical population at risk for AD. This result was achieved since the attention mechanism was able to learn temporal dependencies of the signal, focusing on the most discriminative patterns, achieving state-of-the-art results by using a deep model of reduced complexity. Our results were consistent with clinical evidence that changes in brain activity are progressive when considering early stages of AD.

Subject-independent EEG emotion recognition with hybrid spatio-temporal GRU-Conv architecture

Recently, various deep learning frameworks have shown excellent performance in decoding electroencephalogram (EEG) signals, especially in human emotion recognition. However, most of them just focus on temporal features and ignore the features based on spatial dimensions. Traditional gated recurrent unit (GRU) model performs well in processing time series data, and convolutional neural network (CNN) can obtain spatial characteristics from input data. Therefore, this paper introduces a hybrid GRU and CNN deep learning framework named GRU-Conv to fully leverage the advantages of both. Nevertheless, contrary to most previous GRU architectures, we retain the output information of all GRU units. So, the GRU-Conv model could extract crucial spatio-temporal features from EEG data. And more especially, the proposed model acquires the multi-dimensional features of multi-units after temporal processing in GRU and then uses CNN to extract spatial information from the temporal features. In this way, the EEG signals with different characteristics could be classified more accurately. Finally, the subject-independent experiment shows that our model has good performance on SEED and DEAP databases. The average accuracy of the former is 87.04%. The mean accuracy of the latter is 70.07% for arousal and 67.36% for valence.

Wearable electroencephalography and multi-modal mental state classification: A systematic literature review

BACKGROUND

Wearable multi-modal time-series classification applications outperform their best uni-modal counterparts and hold great promise. A modality that directly measures electrical correlates from the brain is electroencephalography. Due to varying noise sources, different key brain regions, key frequency bands, and signal characteristics like non-stationarity, techniques for data pre-processing and classification algorithms are task-dependent.

METHOD

Here, a systematic literature review on mental state classification for wearable electroencephalography is presented. Four search terms in different combinations were used for an in-title search. The search was executed on the 29th of June 2022, across Google Scholar, PubMed, IEEEXplore, and ScienceDirect. 76 most relevant publications were set into context as the current state-of-the-art in mental state time-series classification.

RESULTS

Pre-processing techniques, features, and time-series classification models were analyzed. Across publications, a window length of one second was mainly chosen for classification and spectral features were utilized the most. The achieved performance per time-series classification model is analyzed, finding linear discriminant analysis, decision trees, and k-nearest neighbors models outperform support-vector machines by a factor of up to 1.5. A historical analysis depicts future trends while under-reported aspects relevant to practical applications are discussed.

CONCLUSIONS

Five main conclusions are given, covering utilization of available area for electrode placement on the head, most often or scarcely utilized features and time-series classification model architectures, baseline reporting practices, as well as explainability and interpretability of Deep Learning. The importance of a 'test battery' assessing the influence of data pre-processing and multi-modality on time-series classification performance is emphasized.

An adversarial discriminative temporal convolutional network for EEG-based cross-domain emotion recognition

Domain adaptation (DA) tackles the problem where data from the source domain and target domain have different underlying distributions. In cross-domain (cross-subject or cross-dataset) emotion recognition based on EEG signals, traditional classification methods lack domain adaptation capabilities and have low performance. To address this problem, we proposed a novel domain adaptation strategy called adversarial discriminative-temporal convolutional networks (AD-TCNs) in this study, which can ensure the invariance of the representation of feature graphs in different domains and fill in the differences between different domains. For EEG data with specific temporal attributes, the temporal model TCN is used as the feature encoder. In the cross-subject experiment, our AD-TCN method achieved the highest accuracies of the valence and arousal dimensions in both the DREAMER and DEAP datasets. In the cross-dataset experiment, two of the eight task groups showed accuracies of 62.65% and 62.36%. Compared with the state-of-the-art performance in the same protocol, experimental results demonstrated that our method is an effective extension to realize EEG-based cross-domain emotion recognition.

Automatic subject-specific spatiotemporal feature selection for subject-independent affective BCI

The dimensionality of the spatially distributed channels and the temporal resolution of electroencephalogram (EEG) based brain-computer interfaces (BCI) undermine emotion recognition models. Thus, prior to modeling such data, as the final stage of the learning pipeline, adequate preprocessing, transforming, and extracting temporal (i.e., time-series signals) and spatial (i.e., electrode channels) features are essential phases to recognize underlying human emotions. Conventionally, inter-subject variations are dealt with by avoiding the sources of variation (e.g., outliers) or turning the problem into a subject-deponent. We address this issue by preserving and learning from individual particularities in response to affective stimuli. This paper investigates and proposes a subject-independent emotion recognition framework that mitigates the subject-to-subject variability in such systems. Using an unsupervised feature selection algorithm, we reduce the feature space that is extracted from time-series signals. For the spatial features, we propose a subject-specific unsupervised learning algorithm that learns from inter-channel co-activation online. We tested this framework on real EEG benchmarks, namely DEAP, MAHNOB-HCI, and DREAMER. We train and test the selection outcomes using nested cross-validation and a support vector machine (SVM). We compared our results with the state-of-the-art subject-independent algorithms. Our results show an enhanced performance by accurately classifying human affection (i.e., based on valence and arousal) by 16%–27% compared to other studies. This work not only outperforms other subject-independent studies reported in the literature but also proposes an online analysis solution to affection recognition.

Detection of mind wandering using EEG: Within and across individuals

Mind wandering is often characterized by attention oriented away from an external task towards our internal, self-generated thoughts. This universal phenomenon has been linked to numerous disruptive functional outcomes, including performance errors and negative affect. Despite its prevalence and impact, studies to date have yet to identify robust behavioral signatures, making unobtrusive, yet reliable detection of mind wandering a difficult but important task for future applications. Here we examined whether electrophysiological measures can be used in machine learning models to accurately predict mind wandering states. We recorded scalp EEG from participants as they performed an auditory target detection task and self-reported whether they were on task or mind wandering. We successfully classified attention states both within (person-dependent) and across (person-independent) individuals using event-related potential (ERP) measures. Non-linear and linear machine learning models detected mind wandering above-chance within subjects: support vector machine (AUC = 0.715) and logistic regression (AUC = 0.635). Importantly, these models also generalized across subjects: support vector machine (AUC = 0.613) and logistic regression (AUC = 0.609), suggesting we can reliably predict a given individual's attention state based on ERP patterns observed in the group. This study is the first to demonstrate that machine learning models can generalize to "never-seen-before" individuals using electrophysiological measures, highlighting their potential for real-time prediction of covert attention states.

The Effects of Individual Differences, Non-Stationarity, and the Importance of Data Partitioning Decisions for Training and Testing of EEG Cross-Participant Models

EEG-based deep learning models have trended toward models that are designed to perform classification on any individual (cross-participant models). However, because EEG varies across participants due to non-stationarity and individual differences, certain guidelines must be followed for partitioning data into training, validation, and testing sets, in order for cross-participant models to avoid overestimation of model accuracy. Despite this necessity, the majority of EEG-based cross-participant models have not adopted such guidelines. Furthermore, some data repositories may unwittingly contribute to the problem by providing partitioned test and non-test datasets for reasons such as competition support. In this study, we demonstrate how improper dataset partitioning and the resulting improper training, validation, and testing of a cross-participant model leads to overestimated model accuracy. We demonstrate this mathematically, and empirically, using five publicly available datasets. To build the cross-participant models for these datasets, we replicate published results and demonstrate how the model accuracies are significantly reduced when proper EEG cross-participant model guidelines are followed. Our empirical results show that by not following these guidelines, error rates of cross-participant models can be underestimated between 35% and 3900%. This misrepresentation of model performance for the general population potentially slows scientific progress toward truly high-performing classification models.

Upper Limb Rehabilitation Tools in Virtual Reality Based on Haptic and 3D Spatial Recognition Analysis: A Pilot Study

With aging, cerebrovascular diseases can occur more often. Stroke cases involve hemiplegia, which causes difficulties in performing activities of daily living. Existing rehabilitation treatments are based on the subjective evaluation of the therapist as the need for non-contact care arises; it is necessary to develop a system that can self-rehabilitate and offer objective analysis. Therefore, we developed rehabilitation tools that enable self-rehabilitation exercises in a virtual space based on haptics. Thirty adults without neurological damage were trained five times in a virtual environment, and the time, number of collisions, and coordinates were digitized and stored in real time. An analysis of variance (ANOVA) of the time and distance similarity changes revealed that as the number of rounds increased, no changes or increases occurred (p ≥ 0.05), and the collisions and paths were stable as the training progressed (p < 0.05). ANOVA showed a high correlation (0.90) with a decrease in the number of crashes and time required. It was meaningful to users when performing rehabilitation training more than four times and significantly impacted the analysis. This study analyzed the upper limb and cognitive rehabilitation of able-boded people in three-dimensional space in a virtual environment; the performance difficulty could be controlled through variations in rehabilitation models.

An Intra-Subject Approach Based on the Application of HMM to Predict Concentration in Educational Contexts from Nonintrusive Physiological Signals in Real-World Situations

Previous research has proven the strong influence of emotions on student engagement and motivation. Therefore, emotion recognition is becoming very relevant in educational scenarios, but there is no standard method for predicting students’ affects. However, physiological signals have been widely used in educational contexts. Some physiological signals have shown a high accuracy in detecting emotions because they reflect spontaneous affect-related information, which is fresh and does not require additional control or interpretation. Most proposed works use measuring equipment for which applicability in real-world scenarios is limited because of its high cost and intrusiveness. To tackle this problem, in this work, we analyse the feasibility of developing low-cost and nonintrusive devices to obtain a high detection accuracy from easy-to-capture signals. By using both inter-subject and intra-subject models, we present an experimental study that aims to explore the potential application of Hidden Markov Models (HMM) to predict the concentration state from 4 commonly used physiological signals, namely heart rate, breath rate, skin conductance and skin temperature. We also study the effect of combining these four signals and analyse their potential use in an educational context in terms of intrusiveness, cost and accuracy. The results show that a high accuracy can be achieved with three of the signals when using HMM-based intra-subject models. However, inter-subject models, which are meant to obtain subject-independent approaches for affect detection, fail at the same task.

Cross-Subject EEG-Based Emotion Recognition Through Neural Networks With Stratified Normalization

Due to a large number of potential applications, a good deal of effort has been recently made toward creating machine learning models that can recognize evoked emotions from one's physiological recordings. In particular, researchers are investigating the use of EEG as a low-cost, non-invasive method. However, the poor homogeneity of the EEG activity across participants hinders the implementation of such a system by a time-consuming calibration stage. In this study, we introduce a new participant-based feature normalization method, named stratified normalization, for training deep neural networks in the task of cross-subject emotion classification from EEG signals. The new method is able to subtract inter-participant variability while maintaining the emotion information in the data. We carried out our analysis on the SEED dataset, which contains 62-channel EEG recordings collected from 15 participants watching film clips. Results demonstrate that networks trained with stratified normalization significantly outperformed standard training with batch normalization. In addition, the highest model performance was achieved when extracting EEG features with the multitaper method, reaching a classification accuracy of 91.6% for two emotion categories (positive and negative) and 79.6% for three (also neutral). This analysis provides us with great insight into the potential benefits that stratified normalization can have when developing any cross-subject model based on EEG.

Application of Transfer Learning in EEG Decoding Based on Brain-Computer Interfaces: A Review

The algorithms of electroencephalography (EEG) decoding are mainly based on machine learning in current research. One of the main assumptions of machine learning is that training and test data belong to the same feature space and are subject to the same probability distribution. However, this may be violated in EEG processing. Variations across sessions/subjects result in a deviation of the feature distribution of EEG signals in the same task, which reduces the accuracy of the decoding model for mental tasks. Recently, transfer learning (TL) has shown great potential in processing EEG signals across sessions/subjects. In this work, we reviewed 80 related published studies from 2010 to 2020 about TL application for EEG decoding. Herein, we report what kind of TL methods have been used (e.g., instance knowledge, feature representation knowledge, and model parameter knowledge), describe which types of EEG paradigms have been analyzed, and summarize the datasets that have been used to evaluate performance. Moreover, we discuss the state-of-the-art and future development of TL for EEG decoding. The results show that TL can significantly improve the performance of decoding models across subjects/sessions and can reduce the calibration time of brain-computer interface (BCI) systems. This review summarizes the current practical suggestions and performance outcomes in the hope that it will provide guidance and help for EEG research in the future.

CNN and LSTM-Based Emotion Charting Using Physiological Signals

Novel trends in affective computing are based on reliable sources of physiological signals such as Electroencephalogram (EEG), Electrocardiogram (ECG), and Galvanic Skin Response (GSR). The use of these signals provides challenges of performance improvement within a broader set of emotion classes in a less constrained real-world environment. To overcome these challenges, we propose a computational framework of 2D Convolutional Neural Network (CNN) architecture for the arrangement of 14 channels of EEG, and a combination of Long Short-Term Memory (LSTM) and 1D-CNN architecture for ECG and GSR. Our approach is subject-independent and incorporates two publicly available datasets of DREAMER and AMIGOS with low-cost, wearable sensors to extract physiological signals suitable for real-world environments. The results outperform state-of-the-art approaches for classification into four classes, namely High Valence-High Arousal, High Valence-Low Arousal, Low Valence-High Arousal, and Low Valence-Low Arousal. Emotion elicitation average accuracy of 98.73% is achieved with ECG right-channel modality, 76.65% with EEG modality, and 63.67% with GSR modality for AMIGOS. The overall highest accuracy of 99.0% for the AMIGOS dataset and 90.8% for the DREAMER dataset is achieved with multi-modal fusion. A strong correlation between spectral- and hidden-layer feature analysis with classification performance suggests the efficacy of the proposed method for significant feature extraction and higher emotion elicitation performance to a broader context for less constrained environments.