This study proposes a three-branch hybrid transformer model, termed EEG Branch Transformer (EBT), to improve cross-subject classification of cognitive fatigue and mental workload from electroencephalography (EEG) signals. Materials and Methods: EBT separately encodes temporal dynamics with a lightweight EEG-specific transformer, time-frequency representations with shared convolution and multi-scale Swin transformers, and spatial topology using 2D convolution on electrode grids, followed by learnable weighted feature fusion. The model was evaluated on three public datasets (Driving EEG dataset for fatigue, EEGMAT and STEW for mental workload) under a leave-one-subject-out cross-validation protocol. Performance was assessed using accuracy and macro F1-score. Compared with the EEG-Deformer baseline, EBT achieved higher accuracy and macro F1 on EEGMAT and STEW, indicating better cross-subject generalization for mental workload estimation. On the Driving EEG dataset, EEG-Deformer obtained slightly higher accuracy, whereas EBT showed comparable macro F1 and reduced inter-subject variability. These results suggest that balancing temporal, time-frequency, and spatial features through branch-wise transformers and weighted fusion is effective for robust EEG-based cognitive state classification across heterogeneous datasets and subjects.

Alzheimer's disease (AD) lacks effective disease-modifying therapies. This study evaluated whether noncontact capacitive alternating electric field therapy (Electric Field Therapy, EFT) could induce amyloid plaque degradation via magnetite-mediated reactive oxygen species generation. Synthetic amyloid-β fibril models were exposed to alternating electric fields under magnetite-containing and magnetite-free conditions. Amyloid changes were assessed using Thioflavin-T assays and transmission electron microscopy, while hydroxyl radical generation was analyzed using aminophenyl fluorescein. Frequency- and time-dependent effects were examined at 200 kHz and 1 MHz. In vivo efficacy was evaluated in 6-month-old 5xFAD mice using Congo-red staining, and neuronal safety was assessed in primary neuronal cultures. Electric field exposure selectively increased hydroxyl radical generation and reduced amyloid fibrils only in magnetite-containing conditions. Magnetite embedded within the protein structure exhibited high responsiveness to the alternating electric field, facilitating localized hydroxyl radical generation and promoting oxidative degradation of nearby amyloid plaques. EFT also decreased amyloid plaque burden in 5xFAD mice without inducing neuronal damage. These findings suggest that non-contact EFT is a promising non-invasive strategy for early-stage AD.

In this study, we applied an unsupervised learning framework to identify potential molecular subtypes exhibiting AD-related molecular patterns from large-scale blood protein expression data with limited clinical labels. A total of 2,933 blood-derived proteins obtained from 12,675 participants were initially analyzed, and after missingness-based quality control, 2,926 proteins and 11,953 samples were retained for downstream analyses. Dimensionality reduction was performed using principal component analysis (PCA) followed by uniform manifold approximation and projection (UMAP), and hierarchical density-based spatial clustering (HDBSCAN) was subsequently applied, resulting in the identification of three structurally distinct latent groups. Among these, Cluster -1 exhibited a distinct protein expression pattern and was observed in a separated region within the UMAP embedding space. Differential expression analysis and functional pathway enrichment analyses based on KEGG (Kyoto Encyclopedia of Genes and Genomes) and Reactome databases revealed that Cluster -1 was characterized by the enrichment of pathways showing biological relevance to previously reported AD-related molecular mechanisms, including DNA damage response, disruption of protein homeostasis, NF-κB signaling, and apoptosis. In contrast, Cluster 0 showed relatively mild dysregulation related to immune and metabolic processes and was therefore interpreted as an intermediate molecular state. Collectively, our findings demonstrate that an unsupervised learning-based approach can identify AD-related molecular subgroups solely from blood protein expression data, even in the absence of prior clinical label information. This framework provides a valuable foundation for the development of future biomarker-driven risk stratification and molecular subtype characterization for Alzheimer's disease.

Human Action Behavior (HAB) systems critically depend on the reliability of human candidate generation at the initial stage, as detection errors directly propagate to subsequent behavior analysis. Conventional object detection approaches, such as YOLO, often suffer from false detections and missed detections (False Negative, FN) under real-world conditions including occlusion, illumination changes, and background clutter. To address these limitations, this study investigates a pose estimation-based approach that leverages structural constraints of the human body and compares it with object detection under identical conditions. A large-scale HAB dataset consisting of 389,040 images was constructed from multiple open-source sources. From the same dataset, both bounding box and keypoint annotations were generated to ensure a fair comparison. Models were trained under identical settings and evaluated using AP@50 and FN metrics. Experimental results show that all models achieved high accuracy (mAP@50 ≥ 0.994), confirming dataset reliability. However, pose estimation demonstrated superior performance in reducing FN. In particular, the YOLOv8-based pose model with CBAM reduced FN by approximately 6% compared to the baseline. Qualitative analysis further revealed that pose estimation maintains robustness in challenging scenarios such as occlusion, body deformation, and crowded environments. These results indicate that pose-based human candidate generation, combined with attention mechanisms, can significantly enhance the operational reliability of HAB systems.

Artificial intelligence-based Software as a Medical Device (AI-SaMD) inherently requires continuous learning and frequent model updates; however, current regulatory frameworks designed for static medical devices reveal limitations in ensuring change traceability and data integrity. Korea enacted the Digital Medical Products Act (DMPA, January 2024, effective January 2025), the world's first standalone digital medical product legislation, introducing a Change Management Plan system (Article 8) and exemption from change approval within the approved plan scope (Article 11 proviso). Nevertheless, a technical mechanism for determining whether a change falls within the approved plan scope remains absent. This study designs and implements a blockchain-based automated verification and audit trail system using Hyperledger Besu and Solidity smart contracts that registers DMPA Change Management Plan parameters on-chain and automatically determines scope compliance of AI model changes. The prototype comprises three Solidity smart contracts (~740 lines) deployed on a four-node IBFT 2.0 testnet. Five change scenarios-covering exemption approval, mandatory approval, performance degradation detection, regulatory audit, and boundary case-were designed based on DMPA enforcement rules classifications. Functional validation using simulation data (n = 5) confirmed correct rule execution in all test cases, average processing time of 2.05 seconds, and an estimated 87% time reduction compared to manual processes. Comparative analysis of DMPA, FDA PCCP, and EU AI Act suggests that the proposed architecture can be extended to serve as cross-regulatory infrastructure. This study presents an engineering solution for ensuring AI medical device safety through immutability, transparency, and automated verification of change histories throughout the AI-SaMD lifecycle.

The purpose of this study was to improve deep learning-based classification of distractor presence in N-back tasks by introducing flickering stimulus presentation. Specifically, steady-state visual evoked potential (SSVEP) stimulation was employed to enhance the neural discriminability of distractor conditions in electroencephalography (EEG) signals. EEG data were acquired from seven healthy adults during a 2-back task performed under controlled experimental conditions, including the presence or absence of flickering (SSVEP) stimulation and auditory and visual distraction stimuli. The recorded EEG signals were preprocessed through notch filtering, bandpass filtering, artifact removal using independent component analysis, and epoch segmentation. The preprocessed data were then analyzed using an improved EEGNet architecture incorporating temporal attention mechanisms and long short-term memory (LSTM) networks to capture both spatial and temporal EEG features. Model performance was evaluated using ten-fold cross-validation. The results showed that EEG signals recorded under flickering stimulus conditions yielded higher classification performance for distractor presence than those without flickering stimulation. In particular, classification accuracy, F1-score, and area under the receiver operating characteristic curve were consistently higher in the flickering condition, with an accuracy improvement of approximately 14.42%. These findings suggest that flickering stimulus presentation enhances the neural separability of distractor states, thereby improving deep learning-based classification performance. In conclusion, the proposed approach demonstrates strong potential for real-time assessment of attentional disruption during cognitive rehabilitation and may support the development of adaptive rehabilitation systems that optimize training difficulty and improve therapeutic outcomes.