Noman Naseer
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Noman Naseer has a total of 288 co-authors. Here are the top co-authors by number of shared publications:
- Keum-Shik Hong
Institute for Future, School of Automation, Qingdao University, Qingdao 266071, China
27 shared publications Hammad Nazeer
Department of Mechatronics and Biomedical Engineering, Air University
22 shared publications- Hammad Nazeer
Department of Mechatronics Engineering, Air University, Islamabad, Pakistan
16 shared publications Muhammad Jawad Khan
School of Mechanical and Manufacturing Engineering, National University of Sciences and Technology (NUST), Islamabad, Pakistan
16 shared publications- Rayyan Azam Khan
Department of Mechanical Engineering, University of Saskatchewan, Saskatoon, SK S7N 5A9, Canada
12 shared publications M. Raheel Bhutta
Department of Cogno-Mechatronics Engineering, Pusan National University, Busan 609-735, Korea
11 shared publications
Latest Publications
- Open Access
- 16 October 2026
- Published by Frontiers Media SA in Frontiers in Neuroergonomics
- Vol. 7, 1726286
Offline and simulated-online fNIRS decoding of ankle movements for prospective exo-ankle control: a pilot study in healthy adults and ankle-impaired patients
This study developed and evaluated a non-invasive brain–computer interface (BCI) framework that decodes functional near-infrared spectroscopy (fNIRS) signals from the motor cortex to generate offline and simulated-online exo-ankle control commands, as a proof-of-concept step toward future prospective exo-ankle control. fNIRS data were recorded from 20 healthy participants and seven patients with ankle motor impairment using a PsychoPy-based paradigm of rest, dorsiflexion, and plantar flexion. Following the literature, after preprocessing, an evaluation of different combinations of six statistical features and four classifiers, namely, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machine (SVM), and k-nearest neighbor (kNN) was conducted. Under leave-one-subject-out (LOSO) cross-validation across 20 healthy participants, SVM achieved the mean offline classification accuracy of 72.50% ± 13.8% followed by kNN (65.83% ± 13.5%), LDA (63.75% ± 16.1%), and QDA (60.42% ± 17.7%). A Friedman test indicated a significant overall difference among classifiers ( p = 0.0311). Simulated-online implementation used SVM with a sliding window approach with step sizes of 0.5, 1, and 1.5 s under participant-specific leave-one-trial-out validation. The results showed a highest mean accuracy of 62.2% ± 14.1% at a 1.5-s step size. The patient group showed a lower mean accuracy of 57.23% ± 5% using SVM under leave-one-trial-out cross-validation. Together, these offline and simulated-online results establish proof-of-concept feasibility for fNIRS-based exo-ankle control, providing a foundation for future prospective rehabilitation applications.
- 16 October 2026
- Published by Frontiers Media SA in Frontiers in Neuroergonomics
- Vol. 7, 1726286
Open AccessOffline and simulated-online fNIRS decoding of ankle movements for prospective exo-ankle control: a pilot study in healthy adults and ankle-impaired patients
This study developed and evaluated a non-invasive brain–computer interface (BCI) framework that decodes functional near-infrared spectroscopy (fNIRS) signals from the motor cortex to generate offline and simulated-online exo-ankle control commands, as a proof-of-concept step toward future prospective exo-ankle control. fNIRS data were recorded from 20 healthy participants and seven patients with ankle motor impairment using a PsychoPy-based paradigm of rest, dorsiflexion, and plantar flexion. Following the literature, after preprocessing, an evaluation of different combinations of six statistical features and four classifiers, namely, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machine (SVM), and k-nearest neighbor (kNN) was conducted. Under leave-one-subject-out (LOSO) cross-validation across 20 healthy participants, SVM achieved the mean offline classification accuracy of 72.50% ± 13.8% followed by kNN (65.83% ± 13.5%), LDA (63.75% ± 16.1%), and QDA (60.42% ± 17.7%). A Friedman test indicated a significant overall difference among classifiers ( p = 0.0311). Simulated-online implementation used SVM with a sliding window approach with step sizes of 0.5, 1, and 1.5 s under participant-specific leave-one-trial-out validation. The results showed a highest mean accuracy of 62.2% ± 14.1% at a 1.5-s step size. The patient group showed a lower mean accuracy of 57.23% ± 5% using SVM under leave-one-trial-out cross-validation. Together, these offline and simulated-online results establish proof-of-concept feasibility for fNIRS-based exo-ankle control, providing a foundation for future prospective rehabilitation applications.
- 30 August 2026
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 36 (5) , e70437
Enhancing fNIRS ‐Based Decoding: A Heterogeneous Ensemble Network for Six‐Class Movement Classification
Neurological upper‐limb impairments severely affect functional independence, motivating the development of brain‐computer interface (BCI) systems for decoding motor intention. Although functional near‐infrared spectroscopy (fNIRS) is a promising modality, its low signal‐to‐noise ratio and large inter‐subject variability make multi‐class motor decoding difficult. In this study, a Heterogeneous Controlled Temporal Ensemble Network (HCTE‐Net) is proposed to classify six different upper‐limb motor tasks based on fNIRS signals. Twenty healthy participants were included for the data acquisition from the motor cortex region of the brain and standardized hemodynamic modeling was used for pre‐processing of signals to include task‐related dynamics. HCTE‐Net combines temporal representations of LSTM, Bi‐LSTM, and Peephole LSTM by probabilistic ensemble fusion. Among conventional machine learning models, XGBoost achieved the highest average accuracy of 69.68%, outperforming SVM (39.09%), LR (37.64%), DT (36.87%), RF (48.54%), LDA (39.39%), and NB (34.58%). The proposed HCTE‐Net achieved a maximum accuracy of 95.46%, exceeding LSTM (84.35%), Bi‐LSTM (87.17%), and Peephole LSTM (73.58%). The results of statistical analysis revealed significant differences among all methods (F = 659.324, p = 4.234 × 10 −152 ), and Tukey's HSD test showed significant improvement ( p < 0.001). Experimental results demonstrate that HCTE‐Net consistently outperforms single models across multiple evaluation metrics.
- 30 August 2026
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 36 (5) , e70437
Enhancing fNIRS ‐Based Decoding: A Heterogeneous Ensemble Network for Six‐Class Movement Classification
Neurological upper‐limb impairments severely affect functional independence, motivating the development of brain‐computer interface (BCI) systems for decoding motor intention. Although functional near‐infrared spectroscopy (fNIRS) is a promising modality, its low signal‐to‐noise ratio and large inter‐subject variability make multi‐class motor decoding difficult. In this study, a Heterogeneous Controlled Temporal Ensemble Network (HCTE‐Net) is proposed to classify six different upper‐limb motor tasks based on fNIRS signals. Twenty healthy participants were included for the data acquisition from the motor cortex region of the brain and standardized hemodynamic modeling was used for pre‐processing of signals to include task‐related dynamics. HCTE‐Net combines temporal representations of LSTM, Bi‐LSTM, and Peephole LSTM by probabilistic ensemble fusion. Among conventional machine learning models, XGBoost achieved the highest average accuracy of 69.68%, outperforming SVM (39.09%), LR (37.64%), DT (36.87%), RF (48.54%), LDA (39.39%), and NB (34.58%). The proposed HCTE‐Net achieved a maximum accuracy of 95.46%, exceeding LSTM (84.35%), Bi‐LSTM (87.17%), and Peephole LSTM (73.58%). The results of statistical analysis revealed significant differences among all methods (F = 659.324, p = 4.234 × 10 −152 ), and Tukey's HSD test showed significant improvement ( p < 0.001). Experimental results demonstrate that HCTE‐Net consistently outperforms single models across multiple evaluation metrics.
- Open Access
- 5 December 2025
- Published by Frontiers Media SA in Frontiers in Robotics and AI
- Vol. 12, 1695169
Open-access fNIRS dataset for motor imagery of lower-limb knee and ankle joint tasks
Brain-Computer Interface (BCI) is an advanced system that enables direct communication between the human brain and external devices, bypassing the need for muscle activity. BCI can translate brain signals into commands to control assistive technologies, offering support for individuals with severe motor impairments, such as those caused by stroke or spinal cord injury. BCI hold applications in healthcare, especially for restoring movement and interaction in patients with limited voluntary motor control. By leveraging neural activity, BCIs open pathways for motor rehabilitation, assistive device control, and even cognitive assessment, enhancing quality of life and promoting greater independence.A widely used control strategy in BCI research is motor imagery (MI), which involves mentally simulating specific physical movements without any actual muscle activation. MI is considered a cognitively driven process that mimics the planning stages of real movements and activates neural networks involved in motor execution. This includes regions such as the primary motor cortex, supplementary motor area (SMA), and premotor cortex, supporting the "functional equivalence" hypothesis (Naseer and Hong, 2015). Because MI can be performed without physical movement, it is especially suitable for patients who are paralyzed, amputated, or recovering from neurological trauma, making it highly relevant for neurorehabilitation and assistive technology development.Over the past decades, a variety of neuroimaging modalities have been employed to capture the neural correlates of MI for BCI applications. These include invasive techniques like intracranial EEG (iEEG) and non-invasive modalities such as electroencephalography (EEG), magnetoencephalography (MEG), functional magnetic resonance imaging (fMRI), and functional near-infrared spectroscopy (fNIRS) (Cutini and Brigadoi, 2014). EEG is widely favoured for its high temporal resolution and affordability; however, it is highly susceptible to electrical interference and motion artefacts. On the other hand, fMRI offers excellent spatial resolution but is often limited by its high cost, immobility, and operational complexity. These limitations restrict their practical use, especially in real-world or bedside BCI applications.Among non-invasive techniques, fNIRS has emerged as a promising neuroimaging modality for BCI systems. fNIRS uses near-infrared light to measure cerebral hemodynamic responses, specifically tracking changes in oxygenated (ΔHbO) and deoxygenated hemoglobin (ΔHbR) concentrations in the cortex (Khan et al., 2020). When a brain region becomes active, local blood flow increases, leading to a rise in ΔHbO and a decrease in ΔHbR. fNIRS offers a balance between spatial and temporal resolution, is portable, non-invasive, relatively affordable, and less sensitive to movement artifacts compared to EEG or fMRI (Nazeer, Naseer, Mehboob, et al., 2020). Its quiet, mobile nature also makes it more comfortable for users and suitable for long-duration BCI sessions. These advantages make fNIRS a particularly suitable tool for developing MI-based BCI systems aimed at real-world rehabilitation scenarios.While MI has been extensively studied for upper limb movements, there is increasing interest in lower limb motor imagery tasks due to their clinical importance. Lower limb movements such as ankle dorsiflexion/plantarflexion and knee flexion/extension are fundamental to essential motor functions like walking, balance, and postural control (Brockett and Chapman, 2016). Understanding how these movements are represented in the brain during MI can help in developing effective rehabilitation protocols and control strategies for assistive devices. MI of lower limb tasks engages similar motor areas to actual movement, supporting its use in BCIs that aim to stimulate recovery or control robotic assistance.Datasets focusing on lower limb MI are essential for advancing research in rehabilitation and assistive robotics. They provide the foundation for developing and training machine learning models that can accurately decode movement intent from brain signals. Several studies have shown that MI of knee or ankle movements can generate meaningful patterns suitable for decoding and can be used to control prosthetic limbs or lower limb exoskeletons (Colucci et al., 2022). For instance, joint EEG-fNIRS studies have demonstrated the effectiveness of decoding ankle dorsiflexion and its relevance to balance recovery in stroke patients (Liang et al., 2022). Similarly, several studies has been conducted to improve classification accuracy of fNIRS (Li et al., 2024;Li, Li, et al., 2025;Li, Sun, et al., 2025) and hybrid EEG-fNIRS signals (Xu et al., 2023;Wang et al., 2025) for motor imagery/execution tasks. Such insights are crucial for improving the responsiveness and reliability of BCI-driven assistive devices.To support this line of research, we introduce a novel fNIRS dataset focused on lower limb MI tasks, specifically ankle dorsiflexion/plantarflexion and knee flexion/extension. The dataset is designed to facilitate the development of BCI systems intended for controlling lower limb exoskeletons in rehabilitation applications (Minhas, Nazeer, Naseer, Khan, et al., 2024). By capturing the hemodynamic responses associated with these specific MI tasks, the dataset enables researchers to explore neural patterns linked to lower limb motor control and develop machine learning models for real-time BCI applications. The unique focus on lower limb motor tasks, combined with the use of fNIRS, distinguishes this dataset and highlights its potential for impactful contributions to neurorehabilitation, exoskeleton design, and assistive technology research.This section details the study's methodology, including participant selection, data acquisition 80 procedures, and the experimental paradigm for MI tasks. 81The study included 21 healthy participants (12 males and 9 females)...
- 5 December 2025
- Published by Frontiers Media SA in Frontiers in Robotics and AI
- Vol. 12, 1695169
Open AccessOpen-access fNIRS dataset for motor imagery of lower-limb knee and ankle joint tasks
Brain-Computer Interface (BCI) is an advanced system that enables direct communication between the human brain and external devices, bypassing the need for muscle activity. BCI can translate brain signals into commands to control assistive technologies, offering support for individuals with severe motor impairments, such as those caused by stroke or spinal cord injury. BCI hold applications in healthcare, especially for restoring movement and interaction in patients with limited voluntary motor control. By leveraging neural activity, BCIs open pathways for motor rehabilitation, assistive device control, and even cognitive assessment, enhancing quality of life and promoting greater independence.A widely used control strategy in BCI research is motor imagery (MI), which involves mentally simulating specific physical movements without any actual muscle activation. MI is considered a cognitively driven process that mimics the planning stages of real movements and activates neural networks involved in motor execution. This includes regions such as the primary motor cortex, supplementary motor area (SMA), and premotor cortex, supporting the "functional equivalence" hypothesis (Naseer and Hong, 2015). Because MI can be performed without physical movement, it is especially suitable for patients who are paralyzed, amputated, or recovering from neurological trauma, making it highly relevant for neurorehabilitation and assistive technology development.Over the past decades, a variety of neuroimaging modalities have been employed to capture the neural correlates of MI for BCI applications. These include invasive techniques like intracranial EEG (iEEG) and non-invasive modalities such as electroencephalography (EEG), magnetoencephalography (MEG), functional magnetic resonance imaging (fMRI), and functional near-infrared spectroscopy (fNIRS) (Cutini and Brigadoi, 2014). EEG is widely favoured for its high temporal resolution and affordability; however, it is highly susceptible to electrical interference and motion artefacts. On the other hand, fMRI offers excellent spatial resolution but is often limited by its high cost, immobility, and operational complexity. These limitations restrict their practical use, especially in real-world or bedside BCI applications.Among non-invasive techniques, fNIRS has emerged as a promising neuroimaging modality for BCI systems. fNIRS uses near-infrared light to measure cerebral hemodynamic responses, specifically tracking changes in oxygenated (ΔHbO) and deoxygenated hemoglobin (ΔHbR) concentrations in the cortex (Khan et al., 2020). When a brain region becomes active, local blood flow increases, leading to a rise in ΔHbO and a decrease in ΔHbR. fNIRS offers a balance between spatial and temporal resolution, is portable, non-invasive, relatively affordable, and less sensitive to movement artifacts compared to EEG or fMRI (Nazeer, Naseer, Mehboob, et al., 2020). Its quiet, mobile nature also makes it more comfortable for users and suitable for long-duration BCI sessions. These advantages make fNIRS a particularly suitable tool for developing MI-based BCI systems aimed at real-world rehabilitation scenarios.While MI has been extensively studied for upper limb movements, there is increasing interest in lower limb motor imagery tasks due to their clinical importance. Lower limb movements such as ankle dorsiflexion/plantarflexion and knee flexion/extension are fundamental to essential motor functions like walking, balance, and postural control (Brockett and Chapman, 2016). Understanding how these movements are represented in the brain during MI can help in developing effective rehabilitation protocols and control strategies for assistive devices. MI of lower limb tasks engages similar motor areas to actual movement, supporting its use in BCIs that aim to stimulate recovery or control robotic assistance.Datasets focusing on lower limb MI are essential for advancing research in rehabilitation and assistive robotics. They provide the foundation for developing and training machine learning models that can accurately decode movement intent from brain signals. Several studies have shown that MI of knee or ankle movements can generate meaningful patterns suitable for decoding and can be used to control prosthetic limbs or lower limb exoskeletons (Colucci et al., 2022). For instance, joint EEG-fNIRS studies have demonstrated the effectiveness of decoding ankle dorsiflexion and its relevance to balance recovery in stroke patients (Liang et al., 2022). Similarly, several studies has been conducted to improve classification accuracy of fNIRS (Li et al., 2024;Li, Li, et al., 2025;Li, Sun, et al., 2025) and hybrid EEG-fNIRS signals (Xu et al., 2023;Wang et al., 2025) for motor imagery/execution tasks. Such insights are crucial for improving the responsiveness and reliability of BCI-driven assistive devices.To support this line of research, we introduce a novel fNIRS dataset focused on lower limb MI tasks, specifically ankle dorsiflexion/plantarflexion and knee flexion/extension. The dataset is designed to facilitate the development of BCI systems intended for controlling lower limb exoskeletons in rehabilitation applications (Minhas, Nazeer, Naseer, Khan, et al., 2024). By capturing the hemodynamic responses associated with these specific MI tasks, the dataset enables researchers to explore neural patterns linked to lower limb motor control and develop machine learning models for real-time BCI applications. The unique focus on lower limb motor tasks, combined with the use of fNIRS, distinguishes this dataset and highlights its potential for impactful contributions to neurorehabilitation, exoskeleton design, and assistive technology research.This section details the study's methodology, including participant selection, data acquisition 80 procedures, and the experimental paradigm for MI tasks. 81The study included 21 healthy participants (12 males and 9 females)...
- 17 September 2025
- Published by Elsevier in Journal of the Mechanical Behavior of Biomedical Materials
- Vol. 173, 107187
Bio-inspired auto-adaptive framework for optimized movement of passive knee prosthesis
- 17 September 2025
- Published by Elsevier in Journal of the Mechanical Behavior of Biomedical Materials
- Vol. 173, 107187
Bio-inspired auto-adaptive framework for optimized movement of passive knee prosthesis
- 1 September 2025
- Published by Institution of Engineering and Technology (IET) in IET Conference Proceedings
- Vol. 2025 (3) , 520-525
Design and fNIRS based control of 2-DOF exo-ankle for gait rehabilitation
- 1 September 2025
- Published by Institution of Engineering and Technology (IET) in IET Conference Proceedings
- Vol. 2025 (3) , 520-525
Design and fNIRS based control of 2-DOF exo-ankle for gait rehabilitation
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 222-226
Improving Navigation Precision of Autonomous Robots with Adaptive Monte Carlo Localization in Dynamic Scenarios
Autonomous navigation robots face substantial challenges in enhancing the accuracy and efficiency of movement localization, particularly in dynamic and unstructured environments. Localization accuracy is highly dependent on the methods employed, making continuous innovation essential to address obstacles in complex conditions. Various techniques, such as the Kalman filter, have been widely adopted to improve localization performance. However, while these methods can deliver satisfactory results under certain conditions, their effectiveness often declines in the presence of uncertainty—especially when dealing with rapid or unpredictable environmental changes. An alternative approach with greater adaptability is Adaptive Monte Carlo Localization (AMCL), which dynamically estimates the robot’s position while accounting for real-time environmental variations. Recent experiments comparing actual trajectories with those estimated using AMCL achieved accuracy levels of 89.2% on the x-axis and 91.2% on the y-axis, demonstrating the method’s effectiveness in improving robot localization precision.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 222-226
Improving Navigation Precision of Autonomous Robots with Adaptive Monte Carlo Localization in Dynamic Scenarios
Autonomous navigation robots face substantial challenges in enhancing the accuracy and efficiency of movement localization, particularly in dynamic and unstructured environments. Localization accuracy is highly dependent on the methods employed, making continuous innovation essential to address obstacles in complex conditions. Various techniques, such as the Kalman filter, have been widely adopted to improve localization performance. However, while these methods can deliver satisfactory results under certain conditions, their effectiveness often declines in the presence of uncertainty—especially when dealing with rapid or unpredictable environmental changes. An alternative approach with greater adaptability is Adaptive Monte Carlo Localization (AMCL), which dynamically estimates the robot’s position while accounting for real-time environmental variations. Recent experiments comparing actual trajectories with those estimated using AMCL achieved accuracy levels of 89.2% on the x-axis and 91.2% on the y-axis, demonstrating the method’s effectiveness in improving robot localization precision.
- Open Access
- 8 August 2025
- Published by Elsevier in SSRN Electronic Journal
Analysis of Fnirs Signals Corresponding to Lower Limb Motor Imagery for Improvement of Classification Accuracy
- 8 August 2025
- Published by Elsevier in SSRN Electronic Journal
Open AccessAnalysis of Fnirs Signals Corresponding to Lower Limb Motor Imagery for Improvement of Classification Accuracy
- Open Access
- 6 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 140118-140138
Mental Workload Detection and Assessment Through Statistical Features Extraction and Optimization Using GEL-RF Method for EEG Signals Using N-Back Dataset
Mental workload (MW) assessment measures the cognitive effort required to perform tasks and is crucial in fields such as aviation, clinical medicine, and healthcare. This study addresses the critical need to optimize cognitive load and resource allocation by evaluating MW through EEG-based analysis. Recently, MW assessment using EEG signals has gained importance, as they show a high correlation with specific cognitive processes and MW. Although previous studies have employed various techniques for MW assessment using EEG, detailed statistical feature selection remains underexplored. This work analyzes the N-back dataset, a standard benchmark for cognitive load, using four combinations of binary and multiclass classification to assess three MW levels. We propose a novel model, which integrates genetic, evolutionary and linear (GEL) feature selection techniques with a Random Forest (RF) classifier as GEL-RF model. Experimental results demonstrate that the GEL-RF model achieves classification accuracies of 97% for binary tasks and 96.3% for multiclass tasks MW assessment, outperforming existing methods. These results shows that our proposed model can improve MW assessment accuracy, helping to increase safety and efficiency in demanding mental tasks.
- 6 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 140118-140138
Open AccessMental Workload Detection and Assessment Through Statistical Features Extraction and Optimization Using GEL-RF Method for EEG Signals Using N-Back Dataset
Mental workload (MW) assessment measures the cognitive effort required to perform tasks and is crucial in fields such as aviation, clinical medicine, and healthcare. This study addresses the critical need to optimize cognitive load and resource allocation by evaluating MW through EEG-based analysis. Recently, MW assessment using EEG signals has gained importance, as they show a high correlation with specific cognitive processes and MW. Although previous studies have employed various techniques for MW assessment using EEG, detailed statistical feature selection remains underexplored. This work analyzes the N-back dataset, a standard benchmark for cognitive load, using four combinations of binary and multiclass classification to assess three MW levels. We propose a novel model, which integrates genetic, evolutionary and linear (GEL) feature selection techniques with a Random Forest (RF) classifier as GEL-RF model. Experimental results demonstrate that the GEL-RF model achieves classification accuracies of 97% for binary tasks and 96.3% for multiclass tasks MW assessment, outperforming existing methods. These results shows that our proposed model can improve MW assessment accuracy, helping to increase safety and efficiency in demanding mental tasks.
- 14 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Annual International Conference of the IEEE Engineering in Medicine and Biology Society
- Vol. 2025, 1-6
fNIRS Based Comparative Study of Classifiers and Feature Selection Techniques for Finger Tapping
This study seeks to classify five-finger movements using machine learning (ML) algorithms. It also examines how feature optimization methods affect classification performance. The signals of functional near-infrared spectroscopy (fNIRS) were acquired from 20 healthy participants as they performed five different finger movements. The recorded signals are represented by a total of 17 spatial features such as kurtosis, variance, mean, skewness and others. The ML classifiers used in the beginning are Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost). Their performance parameters including precision, accuracy, F1-score, recall and processing time are recorded initially for the dataset comprising of all the features. Afterwards, three population-based metaheuristic algorithms Genetic Algorithm (GA), Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO) are used to determine the top features from the dataset. The same ML classifiers are then applied to the selected feature datasets. Classification performance is significantly improved by optimized features, with GA and PSO outperforming ACO. SVM is beaten by XGBoost, while its accuracy (94.94%) is greatest when adopting GA-optimized features. The study also shows the role played by feature selection in improving the efficiency and accuracy of ML models in neuroimaging applications. It also suggests optimized classification pipelines for brain-computer interface systems.
- 14 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Annual International Conference of the IEEE Engineering in Medicine and Biology Society
- Vol. 2025, 1-6
fNIRS Based Comparative Study of Classifiers and Feature Selection Techniques for Finger Tapping
This study seeks to classify five-finger movements using machine learning (ML) algorithms. It also examines how feature optimization methods affect classification performance. The signals of functional near-infrared spectroscopy (fNIRS) were acquired from 20 healthy participants as they performed five different finger movements. The recorded signals are represented by a total of 17 spatial features such as kurtosis, variance, mean, skewness and others. The ML classifiers used in the beginning are Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost). Their performance parameters including precision, accuracy, F1-score, recall and processing time are recorded initially for the dataset comprising of all the features. Afterwards, three population-based metaheuristic algorithms Genetic Algorithm (GA), Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO) are used to determine the top features from the dataset. The same ML classifiers are then applied to the selected feature datasets. Classification performance is significantly improved by optimized features, with GA and PSO outperforming ACO. SVM is beaten by XGBoost, while its accuracy (94.94%) is greatest when adopting GA-optimized features. The study also shows the role played by feature selection in improving the efficiency and accuracy of ML models in neuroimaging applications. It also suggests optimized classification pipelines for brain-computer interface systems.
- 24 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Enhancing Alcohol Use Disorder Diagnosis Using EEG and Hybrid CNN–BiLSTM Deep Learning Models
Alcohol misuse remains a major global health concern, contributing to approximately 3.3 million deaths each year. Conventional methods for diagnosing Alcohol Use Disorder (AUD), such as self-reported questionnaires, are often subjective and susceptible to bias. This study investigates the application of electroencephalography (EEG) combined with machine learning techniques to improve the accuracy and objectivity of AUD diagnosis. Several traditional classifiers—including Decision Tree, Random Forest, Naïve Bayes, and K-Nearest Neighbors—were employed on EEG spectral power features, yielding varying levels of classification accuracy. Among the approaches tested, deep learning models, particularly a combination of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), achieved the highest performance, with a training accuracy of $99.87 \%$ and a test accuracy of $99.37 \%$. These findings underscore the promise of EEG-based deep learning frameworks as precise and reliable tools for diagnosing AUD.- 24 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Enhancing Alcohol Use Disorder Diagnosis Using EEG and Hybrid CNN–BiLSTM Deep Learning Models
Alcohol misuse remains a major global health concern, contributing to approximately 3.3 million deaths each year. Conventional methods for diagnosing Alcohol Use Disorder (AUD), such as self-reported questionnaires, are often subjective and susceptible to bias. This study investigates the application of electroencephalography (EEG) combined with machine learning techniques to improve the accuracy and objectivity of AUD diagnosis. Several traditional classifiers—including Decision Tree, Random Forest, Naïve Bayes, and K-Nearest Neighbors—were employed on EEG spectral power features, yielding varying levels of classification accuracy. Among the approaches tested, deep learning models, particularly a combination of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), achieved the highest performance, with a training accuracy of $99.87 \%$ and a test accuracy of $99.37 \%$. These findings underscore the promise of EEG-based deep learning frameworks as precise and reliable tools for diagnosing AUD.