INTERNATIONAL JOURNAL OF COMPUTERS & TECHNOLOGY

CaT-SleepNet: A Cross-Attention and Transformer-Based Hybrid Framework for EEG–EOG Sleep Stage Classification

Jin Peng — 2025-11-27

sleep disorders and understanding sleep mechanisms. However, traditional deep learning models often fail to effectively
capture both temporal dependencies within Electroencephalogram(EEG) signals and the semantic correlations between
multimodal inputs. In this study, we propose a dual-stream Transformer-based framework that integrates raw EEG and
electro-oculogram (EOG) signals and their corresponding time-frequency (TF) representations through a cross-attention
fusion mechanism. Each modality is first processed by independent feature extractors, followed by pre-trained channellevel Transformers to capture intra-channel temporal dependencies. Finally, a global Transformer module is used for feature extraction, and the learned representations are classified using an optimized XGBoost classifier to enhance
discrimination ability. Experimental results on the Sleep-EDF-20, Sleep-EDF-78, and ISRUC-S3 datasets show that
the proposed model achieves accuracies of 88.5%, 86.8%, and 84.0%, respectively, outperforming several state-of-the-art
baselines and confirming the effectiveness of the proposed multimodal fusion and hybrid Transformer-XGBoost design
for sleep stage classification.

Real-Time Positive Emotion Recognition Using the Positive Unlabeled Learning Method in a Brain Computer Interface System

Zizhu Li — 2025-11-04

Precise identification of emotional states is critical for affective computing applications—ranging from adaptive human–computer interfaces to clinical mental-health assessments. Traditional vision-based systems, however, lose effectiveness when facial expressivity is compromised (e.g., in Alzheimer’s or Bell’s palsy), driving interest in Electroencephalography (EEG)-based approaches. Yet, assembling large, reliably labeled EEG emotion datasets remains a major hurdle. To address this, we introduce a Brain–Computer Interface (BCI) framework that employs Positive–Unlabeled learning, training on a small, labeled subset alongside sufficient unlabeled data for preliminary evaluation. Coupled with a low-cost, portable EEG headset, our design minimizes equipment complexity without sacrificing performance. Validation shows an offline classification accuracy of 86.77% and a 86.20% success rate in real-time trials, confirming the method’s robustness and applicability.

Detection of Deep Sleep Stages in Multi-Channel EEG Signals Based on Spectral Feature Quantification Methods

Jinsha Liu — 2025-11-03

Deep sleep is essential for physical recovery and cognitive function. Accurate detection is crucial for evaluating sleep quality and diagnosing related disorders. Electroencephalography (EEG) remains the most reliable tool for assessing deep sleep. However, while deep learning methods have shown high performance, they often suffer from limited interpretability and require substantial computational resources, which can hinder real-time clinical applications. This study proposes a novel multi-channel EEG analysis approach combining Turning Tangent Empirical Mode Decomposition (2T-EMD) and Multi-Taper Power Spectral Density (MT-PSD) to extract physiologically meaningful spectral features, followed by classification using a Random Forest (RF) model. The method was validated on 61 recordings from 41 participants, achieving 95.57% accuracy, 97.35% recall, 97.58% F1-score, and 99.00% AUC. Compared with existing approaches, our method demonstrates (1) physiologically interpretable feature extraction; (2) favorable computational efficiency under our experimental setup; and (3) robust generalizability across different demographics, indicating its potential utility in clinical sleep monitoring scenarios.

Modelling the Employability of Management Graduates: Complementing Parametric Approaches with Machine Learning on Small Social Data

Ravelonahina Andrianjaka Hasina — 2025-09-29

This study investigates how supervised and unsupervised machine learning algorithms can complement

traditional statistical methods in the analysis of social survey data. Social science datasets are typically small,

noisy, and heterogeneous, which makes robustness and interpretability more important than computational

efficiency.

Using data from a 2024 survey on the employability of management graduates in Antananarivo, the study compares machine learning approaches with classical multivariate techniques. The objectives are to provide a statistical description of a social reality and to establish criteria for selecting algorithms suited to small-sample contexts.

The methodological framework integrates statistical tools such as Chi-square tests, analysis of variance, and multiple regression with exploratory approaches including association rules and clustering. It also incorporates supervised models such as neural networks trained via gradient descent and its variants. Beyond these models, ensemble methods based on decision trees—bagging, random forests, and gradient boosting—are evaluated to highlight their relative strengths.

Findings show that gradient boosting offers the most consistent predictive performance while remaining relatively simple to implement. This makes it particularly effective for analysing small and heterogeneous datasets, thereby providing practical value for applied social science research.

Joint Independent Component Analysis for Enhanced Preprocessing in Collaborative Multi-Brain Motor Imagery BCIs

Jiaxuan Qin — 2025-08-31

Collaborative brain–computer interfaces (cBCIs) extend single-user BCIs to multi-brain recordings, enabling the investigation of inter-brain synchrony and cooperative neural processing. Motor imagery (MI) paradigms are of particular interest but face challenges such as low signal-to-noise ratio, inter-subject variability, and artifact contamination in electroencephalography (EEG) data. Independent component analysis (ICA), though widely applied in single-user MI-BCIs, is suboptimal in cBCIs since it is performed separately for each participant, thereby overlooking shared neural dynamics and potentially distorting inter-brain coupling. To address these limitations, this work proposes a joint ICA-based preprocessing framework that jointly decomposes EEG data from paired participants to enhance artifact suppression while preserving cross-brain synchrony. The performance of joint ICA is evaluated against subject-wise ICA from two perspectives: (i) motor imagery decoding accuracy, examined using spatial filtering with subsequent linear discriminant classification and a convolutional neural network architecture, and (ii) inter-brain synchrony quantified by the phase locking value (PLV) across homologous electrode pairs. Experimental results show that joint ICA improves decoding accuracy by 4.8% and 3.67% in the respective pipelines, while also yielding significantly stronger inter-brain PLV. These findings demonstrate that joint ICA offers an effective preprocessing strategy to improve both neural fidelity and classification robustness in MI-driven cBCI systems.

A Portable EEG-Based Sleep Monitoring and Real-Time Feedback System Without Cloud Infrastructure

Jinsha Liu — 2025-07-28

This paper presents a mobile real-time sleepL. W. et al., 2023; S. Z. et al., 2024; X. M. et al., 2024; Y. E. et al., 2024;
Jirakittayakorn et al., 2024 staging system using EEG signalsA. A. et al., 2023; H. P. et al., 2022; J. K. L. et al.,
2023; P. J. et al., 2023; S. D. et al., 2023; T. L. et al., 2025; X. Z. et al., 2024a collected from the Muse headband. It
employs a lightweight deep neural network, EEGNetG. L. et al., 2024; V. J. L. et al., 2018a; W. C. et al., 2024, to
classify wakefulness, light sleep, and deep sleep. Designed for Android smartphonesS. B. et al., 2017; S. K. et al., 2023;
X. Z. et al., 2024b, EEG signals are transmitted via Bluetooth for local preprocessing and inference, reducing latency
and preserving privacy. Tests with five healthy subjects showed a classification accuracy of 89.4%, closely aligning with results from traditional polysomnography. The system also features sleep-stage-based interventions, such as adaptive white noise playback, enhancing user sleep experience. Compared to conventional EEG devices, the Muse-based system offers greater comfort, portability, and compliance for long-term use. Results highlight the potential of combining consumer-grade EEG and mobile deep learning for accurate real-time sleep monitoring and personalized sleep health managementM. S. et al., 2018; T. Z. et al., 2023; Lai et al., 2018.

Parallel Metaheuristic Algorithms for Task Scheduling Problems

Sirui Mao — 2025-03-31

This study addresses the task-scheduling optimization challenges in parallel computing systems using a novel meta-
heuristic framework. We analyze the differential evolution in task scheduling and propose an advanced shift-chain

methodology to improve the cooperation between scheduling components. The proposed framework introduces a wait-
ing time-based neighborhood exploration strategy for handling complex task dependencies, along with two parallel

implementation approaches: basic matching vector (MV) parallelization and an event-driven strategy. The experi-
mental results demonstrated superior solution quality and computational efficiency compared with existing methods,

particularly in large scale problems. The modular design of this framework enables practical applications in modern
computing environments.