Stair descent is frequently performed in daily life but is a common setting for falls, so clarifying how modifiable initial foot-contact strategies influence neuromuscular control around initial contact is relevant. However, despite extensive biomechanical descriptions, the neuromuscular mechanisms that link initial contact strategy to short-window pre- and post-contact activation and co-contraction patterns remain insufficiently verified. To address this, eleven healthy adults performed stair descent with forefoot, midfoot, and rearfoot contact while sEMG was recorded from the tibialis anterior (TA), soleus (SOL), peroneus longus (PL), vastus medialis (VM), and semitendinosus (ST) and 500-Hz IMUs on the thigh and shank were recorded; initial contact was identified by shank-gyroscope zero-crossing and indices from -150 to 0 ms and 0 to 200 ms were compared using one-way repeated-measures ANOVA. Pre-contact TA differed significantly across conditions (F=7.25, p=0.004), with rearfoot contact higher than forefoot contact. Pre-contact TA-PL co-contraction index also differed significantly (F=6.87, p=0.022), with midfoot contact higher than both forefoot and rearfoot contact. Post-contact TA differed significantly across conditions (F=11.90, p<0.001), with both midfoot and rearfoot contact higher than forefoot contact. Other muscle activation and co-contraction variables did not show significant between-condition differences.

Purpose: This study examined how technology, industry, and policy keywords have been discussed and structurally co-occurred in Korean medical device research over the past three decades, providing an empirical basis for health technology assessment (HTA) policy and industrial strategy. Methods: A cross-sectional bibliometric analysis was performed on 447 papers from the Korea Research Information Service (RISS), with 1994-2024 as the main analytical period (final search: July 15, 2025; 24 partial-year 2025 papers included with caution). Four complementary text mining techniques were applied-TF-IDF, network and CONCOR cluster analysis, LDA topic modeling, and Sentence-BERT (SBERT) semantic similarity analysis. Results: TF-IDF revealed a differentiated discourse: 'surgery' (174.61) and 'robot' (151.45) formed specialized domains, whereas 'insurance' showed high frequency but low TF-IDF, functioning as a broad cross-cutting element. CONCOR yielded seven clusters, the largest (G.7, 26 keywords) bringing technology, industry, and policy/regulatory terms into a single cluster. LDA identified four optimal topics with K=4 selected based on the maximum Coherence Score (0.3356); one topic was insurance-centered (19.2%), while the remaining three (combined 80.7%) addressed digital technology, convergence-device industrialization, and tele-medicine/AI services. SBERT showed that device-related terms were semantically proximate to both innovation/ therapeutic terms (medical_device ↔ innovative_medical_device: 0.87; medical_device ↔ therapeutic_medical_device: 0.87) and legal/regulatory terms (medical_industry ↔ medical_law: 0.88), indicating semantic adjacency among device, innovation, and regulatory dimensions. Conclusion: All four methods consistently showed structural co-occurrence of technology, industry, and policy in Korean medical device discourse. These results offer an empirical basis for HTA policy design and industrial strategy.

This study investigated the performance of an artificial intelligence (AI)-based object detection model for spinal region segmentation in whole-spine X-ray images of patients with scoliosis. A total of 8,999 anonymized wholespine anteroposterior (AP) X-ray images were retrospectively collected and categorized into three groups (normal, mild, and severe) based on Cobb's angle. The dataset was divided into training, validation, and test sets at a ratio of 6:2:2. Several YOLO-based models, including YOLOv8, YOLOv9, YOLOv11, YOLOv12, and modified structures with attention modules and segmentation heads, were trained under the same experimental conditions. Performance evaluation was conducted using box detection mAP@0.5, mask segmentation mAP@0.5, precision, recall, F1-score, and false negative (FN) analysis. Among the tested models, YOLOv12 combined with CBAM showed the highest test box mAP@0.5 value of 0.761 and maintained relatively stable performance between the training and test datasets. Increasing the number of labeling points to 480 led to more frequent background misclassification and a higher number of FN cases. In contrast, the deepnet_v3_head structure reduced FN cases compared with the baseline segmentation head. Furthermore, combining the severe groups into a 3-class configuration showed more consistent detection results than the 4-class configuration. These results suggest that object detection-based segmentation can be effectively applied to automated spinal region segmentation and may contribute to the development of quantitative scoliosis assessment systems.

Circadian rhythm analysis via wearable devices has emerged as a promising digital biomarker for health monitoring. Although commercial wearables utilize diverse step-counting algorithms, varying data thresholds often lead to inconsistent results. The impact of different step-counting algorithms on the calculation of circadian markers was evaluated to identify robust indicators for cardiometabolic disease discrimination. Using accelerometer data from NHANES (2011-2014), six algorithms (ADEPT, Oak, Stepcount RF, Stepcount SSL, Verisense, and ActiLife) were compared. Circadian markers, including Acrophase, Relative Amplitude (RA), and Intradaily Variability (IV), were derived and evaluated using Intraclass Correlation Coefficients (ICCs) and logistic regression. While absolute daily step counts varied by up to 24-fold across algorithms, markers such as Acrophase, RA, and IV demonstrated high structural stability (ICCs 0.57-0.88). Notably, diseased groups consistently exhibited lower RA and higher IV across all algorithms, with statistically significant odds ratios for disease prediction. These findings demonstrate that circadian markers remain robust despite technical differences in step-counting. Consequently, such indicators can serve as reliable digital biomarkers for early detection and monitoring of cardiometabolic diseases, even within diverse wearable ecosystems.

Artificial intelligence(AI)-based medical devices exhibit technical characteristics such as data dependency, and continuous learning, which challenge the regulatory frameworks developed for conventional software-based medical devices. This study aims to identify structural and operational regulatory gaps in Korea's regulatory system for AI-based medical devices by applying ISO/IEC international standards and International Medical Device Regulators Forum (IMDRF) guidelines as normative reference models. Through an analysis of the technical characteristics of AI-based medical devices and their associated risk across the total product life cycle, five key regulatory elements were derived: predetermined change control plans, real-world performance monitoring, explainability, data quality and fairness management, and AI-specific cybersecurity requirements. These elements served as an analytical framework to compare regulatory approaches in the United States and the European Union, with Korea as the primary subject of evaluation. The results indicate that, although Korea has established a legal foundation for regulating AI medical devices through the Digital Medical Products Act, critical operational criteria-particularly for post-market change management, real-world performance monitoring, data bias evaluation, and AI-specific cybersecurity-remain insufficiently specified. This study highlights the need for actionable, life-cycle-oriented regulatory mechanisms aligned with international standards to enhance regulatory predictability, consistency, and patient safety in governance of AI-based medical devices.

Accurate characterization of tissue acoustic properties, particularly the speed of sound (SoS), is important for improving medical ultrasound diagnosis. Most clinical ultrasound systems assume a constant SoS of 1,540 m/s, which can cause phase aberration and image degradation in heterogeneous tissues. However, reconstructing accurate SoS maps from backscattered signals remains a challenging inverse problem for conventional signal processing methods. In this study, we present LDAS-Net, a deep learning architecture for estimating SoS distributions using multiple data modalities. The proposed network employs a dual-encoder framework to extract complementary features from raw radiofrequency (RF) data and beamformed B-mode images. Its key component is the channel-wise Latent DAS (LDAS) strategy, which aligns RF and B-mode features by incorporating beamforming principles into the latent space. Experiments on simulation datasets show that LDAS-Net achieves an MAE of 2.6799, a PSNR of 27.4740, and a correlation coefficient of 0.9767. The model also produces clearer tissue boundaries and less noise in reconstructed SoS maps than existing baseline methods. These results suggest that latent feature alignment between RF and B-mode representations can improve simulation-based SoS reconstruction. Further validation using phantom and in vivo datasets is required before clinical application.

This study aims to investigate the engineering evolution and structural characteristics of Brain-Computer Interface (BCI) technologies through a bibliometric analysis of 8,500 articles published in the Web of Science (WoS) database between 2016 and 2025. The analysis was conducted based on quantitative bibliometric indicators derived from WoS data and network visualization using VOSviewer, including annual scientific production analysis, keyword co-occurrence network analysis, and thematic evolution analysis. Through this approach, the structural relationships among signal acquisition, signal processing, and control components of BCI systems were interpreted from an engineering perspective. The results indicate that BCI research has steadily expanded with an average annual growth rate of 11.3% over the past decade, with electroencephalography (EEG) identified as the dominant signal acquisition modality. In addition, a clear transition from conventional feature-based signal processing methods to deep learning-based neural decoding approaches was observed. The research focus has also shifted from algorithm-level performance improvement toward system-level integration. This study provides an engineering-oriented interpretation of the evolution of BCI technologies and proposes a structural framework for neural signal processing systems. The findings are expected to offer fundamental insights for the design of clinically applicable BCI systems and future research directions in neuroengineering.

Tiger gait has been proposed as an efficient strategy for uphill walking, yet its biomechanical characteristics have not been quantitatively examined. This study investigated how tiger gait influences joint mechanical work compared with that during Normal gait. Eleven healthy adults performed two tiger gaits (straight and cross) and Normal gait on a 10° inclined ramp. Three-dimensional kinematic data were collected using a nine-camera motion capture system, and ground reaction forces were measured with two embedded force plates. Joint mechanical work at the hip, knee, and ankle were calculated using inverse dynamics during stance phase. The stance phase was divided into four functional phases-weight acceptance, pull-up, forward continuance, and push-up-based on five gait events. Mechanical work was obtained by integrating joint power within each phase and was normalized to body weight and vertical center-of-mass displacement. Significant differences in joint mechanical work between tiger and Normal gaits were observed in the sagittal plane, but not in the coronal plane; therefore, the subsequent analyses were confined to the sagittal plane. Total work throughout the stance period in tiger gaits increased at the hip joint but decreased at the ankle joint, compared with that of Normal gait (p<0.05). Specifically, hip joint work increased in the pull-up phase and ankle joint work decreased in the forward continuance phase. These findings suggest that tiger gaits redistribute lower limb mechanical work across joints and phases, such that the burden on the ankle joint in a late phase is transferred to the hip joint in an early phase.

Transcranial Low-Intensity Focused Ultrasound (LIFUS) is a non-invasive neuromodulation technique that delivers focused acoustic energy to specific brain regions. While the skull has been reported as a primary factor affecting ultrasound transmission, the systematic influence of cancellous bone ratio remains insufficiently characterized. In this study, a simplified spherical skull model was used to perform 3D LIFUS simulations, analyzing the effects of skull thickness, cancellous bone ratio, and transducer distance on ultrasound transmission characteristics. 3D LIFUS simulations were performed with systematically varied skull thickness (4-12 mm), cancellous bone ratio(0-60%), and transducer distance (4-12 mm), resulting in a total of 1,053 simulation conditions. Simulations were evaluated based on peak pressure and focal volume. Pearson correlation analysis was performed to assess the statistical relationships between simulation outcomes and three structural variables. Machine learning models, including Linear Regression, Random Forest, XGBoost, and Support Vector Regression (SVR), were trained on simulation datasets to predict peak pressure and focal volume. Skull thickness showed the strongest relationship with both peak pressure and focal volume. The cancellous bone ratio showed a statistically significant but negligible correlation with peak pressure (r=0.06, p<0.01) and a weak negative relationship with focal volume. Across all prediction targets, SVR consistently achieved the highest performance, followed by XGBoost and Random Forest, while Linear Regression showed the lowest accuracy. These findings indicate that skull thickness and cancellous bone ratio are key structural variables in LIFUS. The results of this study may serve as preliminary data for further investigation into prediction models incorporating individual skull properties.