Chinese Journal of Biomedical Engineering

Home

Editorial Board

Download

Quick Search

Office Online

Current Issue

Download Articles

Download

Links

2026 Vol. 45, No. 4 Published: 2026-08-20

Reviews Communications Regular Papers Special Section: Brain-Computer Interface

Special Section: Brain-Computer Interface

	385	Ultrasound Brain Computer Interface
		Zheng Hairong, Niu Lili, Qiu Weibao, Liu Chengbo, Li Zhicheng, ZhouYihang, Meng Long, Pan Yue
		DOI: 10.3969/j.issn.0258-8021.2026.04.001
		Brain-computer interface, as a core technology for achieving human-machine integration and tackling neurological diseases, has become a key area in global technological competition and national strategic planning, holding significant strategic potential in supporting intelligent healthcare, human-machine collaboration, and breakthroughs in brain science. Ultrasound brain computer interface, with its unique advantages of non-invasive penetration of the skull and high spatiotemporal resolution, has emerged as a critical technological direction in the field of brain computer interaction. This review summarized the key technologies of ultrasound brain computer interfaces, including ultrasonic brain functional imaging for real-time dynamic observation of cerebral blood flow and neural activity correlations, the construction of decoding models, and neural writing techniques combining sonogenetics for cell-specific modulation. It elaborated on the application progress of ultrasonic brain computer interfaces in revealing neurovascular coupling mechanisms, which can visually demonstrate the relationship between brain functional operations and disease pathology, providing new insights for the diagnosis and treatment of brain disorders such as Alzheimer’s disease. Furthermore, technical challenges faced by ultrasound brain computer interfaces and application prospects in fields such as healthcare, military, industry, autonomous driving, and brain-to-brain communication were discussed.
		2026 Vol. 45 (4): 385-404 [Abstract] ( 15 ) HTML (1 KB) PDF (9031 KB) ( 5 )

	405	Progress and Challenges in the Clinical Translation of Invasive Brain-Computer Interfacesfor Rehabilitation of Neurological Dysfunction
		Shen Zijian, Hu Feng, Shu Kai, Wu Dongrui
		DOI: 10.3969/j.issn.0258-8021.2026.04.002
		Brain-computer interfaces(BCIs), a frontier interdisciplinary technology that connects the brain to external devices, have achieved breakthrough advances in the diagnosis, treatment, and rehabilitation of neurological diseases. In particular, invasive BCIs (iBCIs), with their exceptionally high spatiotemporal resolution and superior signal-to-noise ratio, have become a pivotal technological pathway for tackling severe neurological dysfunctions such as stroke, severe spinal cord injury, and amyotrophic lateral sclerosis(ALS). During the 2024—2026 critical window for global clinical translation of the BCI industry, China has achieved multiple significant breakthroughs in core minimally invasive hardware development, adaptive neural decoding algorithms, the establishment of multicenter clinical trial platforms, and the development of ethical and regulatory standards. Building on the foundational scientific discoveries and prospective clinical trial data from recent leading research teams at home and abroad, this review systematically surveys the latest progress in iBCI core hardware fabrication and the evolution of signal decoding theory; examines in depth the landmark achievements across the three key areas of clinical translation-motor function restoration, speech-assisted communication, and neural function modulation; and comprehensively assesses the evolution of current multicenter clinical research designs, the reconstruction of rehabilitation efficacy evaluation systems, and the ongoing development of a collaborative governance framework integrating technology, clinical practice, ethics, and regulation. This work aims to provide a systematic academic reference for future basic research and large-scale clinical application in this field.
		2026 Vol. 45 (4): 405-415 [Abstract] ( 10 ) HTML (1 KB) PDF (1149 KB) ( 3 )

	416	Comparative Study of SSVEP Characteristics Evoked by Semantic Stimuli and Luminance Flicker
		Li Xiaoyang, Gao Xiaorong
		DOI: 10.3969/j.issn.0258-8021.2026.04.003
		Steady-state visual evoked potential (SSVEP) based brain-computer interfaces (BCI) provide an efficient communication pathway for individuals with severe motor impairments. Conventional SSVEP paradigms typically rely on luminance flicker stimuli, which primarily engage early visual pathways and may therefore be constrained by the response characteristics of the primary visual cortex. To expand the encoding capacity of SSVEP-based systems, it is necessary to explore alternative stimuli capable of engaging richer visual processing mechanisms. However, the response characteristics of semantic visual stimuli under SSVEP paradigms remain insufficiently investigated. In this study, we compared SSVEP responses evoked by faces, cars, Chinese characters, and luminance flicker at four stimulation frequencies (3, 6, 9, and 12 Hz). The results showed that, compared with luminance flicker, semantic stimuli exhibited distinct temporal, spatial, and spectral response patterns. Among the tested stimuli, Chinese characters showed better overall performance, with higher signal-to-noise ratio (SNR) than faces and cars, and better classification performance than luminance flicker in some time windows. Notably, at 6 Hz, character stimuli produced a broader spatial distribution and a peak fundamental SNR of 13.29 dB. These findings suggest that character-based stimulation may offer a promising alternative to luminance flicker and provide new directions for visual encoding in SSVEP-BCI.
		2026 Vol. 45 (4): 416-427 [Abstract] ( 9 ) HTML (1 KB) PDF (5587 KB) ( 3 )

	428	Adaptive Mixture of Experts for EEG-Based Emotion Recognition
		Duan Mengtong, Cui Heng, Chen Xun, Li Yu, Liu Aiping
		DOI: 10.3969/j.issn.0258-8021.2026.04.004
		Human emotions are closely linked to specific brain regions. Recent studies have increasingly focused on extracting regional representations for EEG-based emotion recognition (EER). However, the EEG signals associated with these regions exhibit high individual specificity in both feature distributions and brain connectivity patterns, making effective generalization for cross-subject EER a key challenge. To address this challenge, we proposed an adaptive mixture of experts (AdaMoE), leveraging MoE's expert specialization mechanism for specialized regional processing while introducing a dual test-time adaptation strategy over feature representations and graph topology to enhance cross-subject generalizability. Specifically, AdaMoE utilized multiple graph convolution experts that adaptively process different combinations of EEG channels and their corresponding inter-regional interaction patterns to capture diverse regional representations. On this basis, feature adaptation dynamically calibrated target sample representations to improve the domain-invariance of emotion-discriminative features, while topology adaptation adaptively adjusts graph connections to more accurately characterize subject-specific brain-region interaction patterns. Extensive experiments on two public datasets demonstrated that AdaMoE achieved accuracies of 80.5% and 57.7%, with F1 scores of 79.6% and 56.5% on the SEED and SEED-IV datasets, respectively, thereby effectively improving cross-subject EEG emotion recognition performance. These results demonstrate that regional graph-expert modeling and dual feature-topology test-time adaptation can jointly mitigate feature and structural shifts in unseen subjects while maintaining stable online updates.
		2026 Vol. 45 (4): 428-443 [Abstract] ( 10 ) HTML (1 KB) PDF (6409 KB) ( 3 )

	444	Transformational Spiking Neural Network and its Application in Brain-Computer Interface
		Zhao Zechen, Yu Tianyou, Yu Zhuliang, Gu Zhenghui
		DOI: 10.3969/j.issn.0258-8021.2026.04.005
		The monitoring of electroencephalogram (EEG) signals imposes stringent demands on the real-time performance and energy efficiency of computing systems, requiring the preservation of high temporal resolution while minimizing resource consumption. Spiking neural networks (SNNs), through their spike-based information transmission mechanism, replace the multiplication operations in traditional neural networks with addition operations. Their asynchronous communication characteristics enhance the processing efficiency of temporal data, thereby meeting real-time requirements. As a result, SNNs exhibit significant application potential in continuous EEG monitoring tasks. However, the direct design and optimization of SNNs are highly complex and challenging to deploy efficiently. To address this issue, this study proposed a quantize-clamp-quantize (QCQ) method for converting convolutional neural networks (CNNs) to SNNs and applied it to the SEED dataset for EEG emotion classification and the BCI Competition IV Dataset 2a for motor imagery. Specifically, the QCQ method introduced clamping and quantization operations in the activation and output layers of computational layers, adopted a single quantization strategy for input and pooling layer computations, and migrated batch normalization parameters to achieve efficient CNN-to-SNN conversion. The proposed QCQ method achieved accuracies of 93.79% on the SEED dataset and 78.64% on the MI dataset, yielding average performance gains of 2.77% and 4.31% over advanced existing methods, respectively. These improvements were accomplished with substantial energy efficiency enhancements of nearly 75% and 80% on the two datasets. These results validated that the QCQ method effectively inherited the classification performance of CNNs, brought SNN performance closer to CNNs on both datasets, and provided a practical implementation pathway for SNN applications.
		2026 Vol. 45 (4): 444-456 [Abstract] ( 10 ) HTML (1 KB) PDF (7200 KB) ( 3 )

	457	Development of a Highly Integrated Brain-Computer Interface Protective Helmet Based onDetachable Flexible Comb-Shaped Sponge Electrodes
		Yuan Xuezhi, Li Hui, Tang Qi, Chen Jie, Wang Kun, Xiao Xiaolin, Xu Minpeng, Fan Yi, Bai Ru, Wang Junlei
		DOI: 10.3969/j.issn.0258-8021.2026.04.006
		Aiming to solve core problems faced in the integration of brain-computer interface (BCI) and protective helmets, such as portable electrode replacement, small hardware anti-interference, stable signal transmission and wearing comfort, this paper proposed an innovative solution of "integration of detachable flexible comb-shaped sponge electrodes and highly integrated amplifiers". The electrode used an elastic sponge-flexible silicone substrate, flexible metal wires, and a comb-shaped distributed layout, improving comfort while ensuring signal quality. The amplifier, with high-integration components and a simplified circuit, achieved a volume of 6.6 cm×5.2 cm×1.2 cm, a common-mode rejection ratio exceeding 110 dB, a system battery life of 29.3 hours, and a 0.1%~0.2% packet loss rate. Wearing comfort evaluation and preparation time tests on 10 subjects showed an average preparation time of (2.28±0.48)minutes, demonstrating good usability and applicability. The unmanned aerial vehicle (UAV) control experiment based on steady-state visual evoked potential (SSVEP) verified that the average control accuracy for the 10 subjects reached 90.88%±2.61%. The system was able to operate stably and reliably, providing an effective technical path for the deep integration of BCI and protective equipment.
		2026 Vol. 45 (4): 457-468 [Abstract] ( 11 ) HTML (1 KB) PDF (7608 KB) ( 3 )

Regular Papers

	469	A Fast Deep Convolutional Neural Network for Mammograms Classification
		Cai Lvjun, Fu Yixuan, Ruan Yaoping, Li Yane
		DOI: 10.3969/j.issn.0258-8021.2026.04.007
		Breast cancer is one of the most common cancers among women worldwide. Early detection, diagnosis and treatment are necessary to elevate survival rates and quality of life for breast cancer patients. Mammography has been recognized as the most popular modality for breast cancer screening. However, it is quite challenging to interpret mammograms accurately primarily due to the overlapping of tissues, under the universal screening model which yields a large number of mammographic images. Many methods have been developed for computer-aided diagnosis schemes to classify patients into benign and malignant with mammography. These models are useful for early breast cancer detection, while approaches are needed to improve accuracy and reduce the time and costs. In this paper, we proposed a new lightweight deep convolutional neural network to classify breast cancer rapidly with mammograms. A public digital database for screening mammography, the Digital Database for Screening Mammography (DDSM), was used in this study. The dataset contains 2 604 patients, of which 854 cases were pathologically confirmed benign, 141 cases were benign without callback, 914 patients were malignant confirmed by biopsy and pathology examinations and 695 were negative patients. We first enhanced the data with image processing methods including rotating, flipping, brightening, adding salt-and-pepper noise and Gaussian noise. The post-processing image dataset has 48 300 mammograms, including 5 208 original mammographic images and 43 092 enhanced images. We proposed a deep neural network model named classification of mammograms convolutional neural network (CMCNNet) by embedding VGGNet and inception modules to classify breast cancer with mammograms. Results showed that the new network architecture achieved an accuracy of 99.38% at a learning rate of 0.000 5. In addition, we figured out that when the number of convolution kernels was set to 32-32-32-32 corresponding to the convolutional layer group, and the size of the convolutional kernel was set to 3×3, the accuracy increased to 99.60%. Finally, the proposed model was validated on 48 300 mammographic images (2 604 patients). A three-fold cross-validation method was applied to different models using AlexNet, VGGNet, Inception-V3, ResNet, as well as the CMCNNet. Results showed that CMCNNet achieved an accuracy of 99.86% that outperformed all the other models adopted in this study with the same iteration numbers for training. In addition, the CMCNNet model needed the least training time. Results of the parameter optimization study and the new network model proposed in this paper have profound practical significance for the classification of breast cancer using mammography.
		2026 Vol. 45 (4): 469-481 [Abstract] ( 12 ) HTML (1 KB) PDF (10553 KB) ( 3 )

	482	Classification of Mild Cognitive Impairment, Alzheimer’s Disease, and Healthy ControlsBased on Weighted Random Forest-KNN
		Yuan Yuyang, Li Fangyun, Li Yingjie, Li Yunxia
		DOI: 10.3969/j.issn.0258-8021.2026.04.008
		Mild cognitive impairment (MCI), a critical prodromal stage of Alzheimer's disease (AD), exhibits distinct neural oscillatory features from healthy controls (HC). These differences are pivotal for early AD diagnosis and research on cognitive aging mechanisms. Brain functional connectivity in the α-band (8~13 Hz) has been proven to be closely associated with resting-state network activity, yet its classification efficacy and key biomarkers across MCI, AD, and HC populations remain incompletely elucidated. Meanwhile, traditional machine learning algorithms inadequately assign weights to high-dimensional EEG features, making it difficult to capture subtle differences among multiple groups. This study employed a feature importance weighting strategy via random forest to enhance the discriminability of critical brain connections in traditional machine learning models, addressing the inefficient feature utilization of single models. Resting-state EEG data from 43 participants each in MCI, AD, and HC groups were analyzed. A full feature matrix (AFM) for the α-band was constructed using Pearson correlation coefficients (PCC), and brain connections with significant inter-group differences were filtered via analysis of variance (ANOVA) to form an effective feature matrix (EFM). An integrated weighted random forest-KNN (WRF-KNN) model was designed, and its performance in three-class (MCI/AD/HC) and two-class (MCI/AD) tasks was evaluated using 5-fold stratified cross-validation. Additionally, a network of critical brain connections was constructed based on the feature importance of random forest. The WRF-KNN model achieved an average accuracy of 93% ±2.1% in the three-class task, significantly outperforming single KNN (88% ± 3.5%) and random forest (90% ± 2.8%, P < 0.05); its accuracy in MCI-AD two-class classification reached 92%±1.8%. Concurrently, the model identified 20 critical α-band connections. Their importance scores revealed key differences in α-band neural oscillations and functional connectivity between MCI/AD patients and HC, MCI and AD groups exhibited reduced prefrontal-parietal connectivity and enhanced intra-occipital connectivity, whereas HC displayed a more uniform connection distribution. These findings open new perspectives for the early diagnosis of AD and research on cognitive aging.
		2026 Vol. 45 (4): 482-494 [Abstract] ( 15 ) HTML (1 KB) PDF (14159 KB) ( 3 )

Reviews

	495	Research Progress of Novel Dressing Materials in Clinical Wound Care
		Wei Kun, Li Kai, Wang Peixin, Zhu Huan, Zhang Lijuan
		DOI: 10.3969/j.issn.0258-8021.2026.04.009
		Wound repair is one of the core research directions in the fields of traumatology and biomaterials. For complex refractory wounds such as chronic diabetic foot ulcers and deep burns, conventional dressings suffer from inherent limitations including limited moist environment maintenance capacity, suboptimal infection prevention and control efficacy, and prolonged healing cycles, failing to meet clinical treatment demands. With the interdisciplinary integration of materials science, biomedicine and clinical medicine, novel medical dressings can effectively improve wound repair quality by mimicking the physiological barrier function of skin and actively regulatingwound microenvironments. This review systematically summarized the research progress of novel medical dressings,including hydrogels, hydrocolloids, foams, alginates, silver-ion, and nanofibers dressings based on material properties. Thekey components, material characteristics, wound healing-promoting mechanisms, clinical application outcomes and existing challengeswere elaborated.At the same time, future development directions were prospected from three dimensions including multi-component synergistic design, precise targeted regulation and construction of clinical translation systems.
		2026 Vol. 45 (4): 495-502 [Abstract] ( 12 ) HTML (1 KB) PDF (825 KB) ( 4 )

Communications

	503	An EEG Artifact Suppression Method Combining Naive Bayes Noise Segments Identificationwith Wavelet Denoising
		Xie Wencan, Wang Guoqing, Xu Ce, Wang Chengmian, Zhao Mingwei, Li Guangye
		DOI: 10.3969/j.issn.0258-8021.2026.04.010

		2026 Vol. 45 (4): 503-507 [Abstract] ( 11 ) HTML (1 KB) PDF (5983 KB) ( 4 )

	508	Effects of Biofeedback Combined with Tongue Pressure Resistance Feedback Training on SwallowingDisorders, Nutritional Status and Quality of Life in Children with Craniocerebral Injury
		Yue Lei, Liu Yanjie, Wang Xu, Liu Juan, Zhao Huizhong, Luo Xu
		DOI: 10.3969/j.issn.0258-8021.2026.04.011

		2026 Vol. 45 (4): 508-512 [Abstract] ( 13 ) HTML (1 KB) PDF (753 KB) ( 3 )

Copyright © Editorial Board of Chinese Journal of Biomedical Engineering
Supported by:Beijing Magtech