18 November 2024 in IEEE Open Journal of the Computer Society
Epilepsy has been analyzed through uni-modality non-invasive brain measurements such as electroencephalogram (EEG) signal, but identifying seizure patterns is more challenging due to the non-stationary nature of the brain activity and various non-brain artifacts. In this paper, we leverage a vision transformer model (ViT) to classify three types of seizure patterns based on multimodal EEG and functional near-infrared spectroscopy (fNIRS) recordings. We used spectral encoding techniques to capture temporal and spatial relationships for brain signals as feature map inputs to the transformer architecture. We evaluated model performance using the receiver operating characteristic (ROC) curves and the area under the curve (AUC), demonstrating that multimodal EEG-fNIRS signals improved the classification accuracy of seizure patterns. Our work showed that power spectral density (PSD) features often led to better results than features derived from dynamic mode decomposition (DMD), particularly for seizures with high-frequency oscillations (HFO) and generalized spike-and-wave discharge (GSWD) patterns, with an accuracy of 93.14% and 91.69%, respectively. Low-voltage fast activity (LVFA) seizures achieved consistently high performance in EEG, fNIRS, and multimodal EEG-fNIRS setups. Overall, our findings suggest the effectiveness of using the ViT architecture with multimodal brain data accompanied by appropriate spectral features to classify the neural activity of epileptic seizure patterns.
16 January 2026 in Applied Computing and Geosciences
3 July 2025 in IEEE Access
The use of electroencephalography (EEG)-based analysis in educational research offers valuable insight into the dynamics of virtual- and robot-based tutoring systems. This study aims to classify neural patterns associated with two distinct tutoring modalities (virtual tutor vs. robot tutor) to understand how different educational interfaces affect brain activity and engagement. Such classification enables adaptive learning systems that can optimize educational environments based on real-time neural responses, potentially enhancing learning outcomes by matching tutoring modalities to individual neural preferences. We examine the performance of multiple EEG features such as power spectral density, Fast Fourier Transform (FFT) magnitude coefficients, amplitude and variance, coherence, Hjorth parameters, and wavelet transform coefficients in distinguishing between these two tutoring modalities. A transformer-based binary classifier was employed to evaluate the effectiveness of these features in classifying EEG data collected during interactions with both types of tutors. The wavelet transform coefficients demonstrated the highest classification accuracy of 98.75%, precision of 99.17%, recall of 98.18%, F1 score of 98.61%, and Area Under the Curve (AUC) of 1.00, indicating exceptional performance across all metrics. This high accuracy represents a significant advancement in EEG-based educational classification systems. Hjorth parameters (90.13% accuracy, 89.91% precision, 95.45% recall) and preprocessed EEG (91.73% accuracy, 91.57% precision, 90.91% recall) also showed strong performance. Simpler features, such as amplitude and variance, exhibited limited discriminatory power with only 72.67% accuracy. These findings underscore the importance of feature engineering and robust preprocessing in EEG-based educational studies and demonstrate the feasibility of creating personalized learning environments that can adapt to individual students’ neural preferences. Limitations such as a small sample size and topic-specific focus are noted, paving the way for future research to generalize findings and optimize methodologies.
11 June 2025 in Research Square
This paper presents a comparative study of three deep learning architectures for one-dimensional magnetotelluric (MT) inversion: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Informer models. We developed a comprehensive framework for generating realistic synthetic MT data, training the models, and evaluating their performance through multiple quantitative metrics. Our synthetic data set consisted of 10,000 samples with 25 periods spanning 10-3 to 103 seconds, created using statistical parameters derived from real MT data. Each model was trained on apparent resistivity and phase responses to recover subsurface resistivity profiles. The results show that the recurrent neural network architectures (GRU and LSTM) slightly outperform the attention-based Informer model, with the GRU achieving the best overall performance (MSE of 97.67 Ohm.m2, R² of 0.44). Despite differences in their architectures, all models successfully captured the major resistivity contrasts in the subsurface. When applied to real MT data from the UAE, the models showed promising results in reconstructing subsurface structures. This study demonstrates the viability of deep learning approaches for MT inversion, with potential applications in subsurface imaging for efficient field interpretations.