Biomedical Engineering Letters最新文献

Recognition of human emotions holds great potential for various daily-life applications. With the increasing interest in virtual reality (VR) technologies, numerous studies have proposed new approaches to integrating emotion recognition into VR environments. However, despite recent advancements, camera-based emotion-recognition technology faces critical limitations due to the physical obstruction caused by head-mounted displays (HMDs). Facial electromyography (fEMG) offers a promising alternative for human emotion-recognition in VR environments, as electrodes can be readily embedded in the padding of commercial HMD devices. However, conventional fEMG-based emotion recognition approaches, although not yet developed for VR applications, require lengthy and tedious calibration sessions. These sessions typically involve collecting fEMG data during the presentation of audio-visual stimuli for eliciting specific emotions. We trained a machine learning classifier using fEMG data acquired while users intentionally made posed facial expressions. This approach simplifies the traditionally time-consuming calibration process, making it less burdensome for users. The proposed method was validated using 20 participants who made posed facial expressions for calibration and then watched emotion-evoking video clips for validation. The results demonstrated the effectiveness of our method in classifying high- and low-valence states, achieving a macro F1-score of 88.20%. This underscores the practicality and efficiency of the proposed method. To the best of our knowledge, this is the first study to successfully build an fEMG-based emotion-recognition model using posed facial expressions. This approach paves the way for developing user-friendly interface technologies in VR-immersive environments.

Accurate segmentation of pulmonary nodule based on computed tomography (CT) images is of great significance for the diagnosis and treatment of lung cancer. However, the current popular segmentation algorithms usually do not involve expert knowledge from radiologists, thereby carrying the risk of failing to produce generalizable and trustworthy models. In this study, we develop an expert-informed active pulmonary nodule segmentation method that iteratively optimize a deep segmentation model using an active learning scheme. The uncertainties from both intermediate segmentation results and correction inputs from radiologists are combined effectively. Interactive graph interfaces are developed to enable online corrections, significantly facilitating the integration of expert knowledge from radiologists. Evaluation results on the Luna16 dataset demonstrate that the proposed approach significantly promotes the segmentation performance of pulmonary nodules. The proposed method can effectively incorporate expert knowledge of multiple radiologists into deep segmentation algorithms, which not only promote the segmentation performance, but also enhance the validity, reliability, and generalizability of computer-aided diagnosis methods.

This systematic review examines the transformative impact of artificial intelligence (AI), including machine learning (ML) and deep learning (DL), on cardiovascular signal analysis, focusing on electrocardiograms (ECG), phonocardiograms (PCG), and photoplethysmograms (PPG). It evaluates state-of-the-art methodologies that enhance diagnostic accuracy and predictive analytics by leveraging AI-driven systems. A wide range of public and private datasets are assessed, with attention to their strengths and limitations in supporting cardiovascular diagnostics. Key preprocessing techniques, such as noise reduction, signal normalization, and artifact mitigation, are explored for their role in improving signal quality. The review also highlights feature extraction methods, from time-domain and frequency-domain analyses to advanced morphological and spectral approaches, which contribute to robust classifier performance. Traditional ML models, such as K-Nearest Neighbor (KNN), Support Vector Machines (SVM) and Random Forests (RF), are compared with advanced DL architectures, including Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), and transfer learning models for detecting cardiovascular diseases. Despite these advancements, challenges such as dataset heterogeneity, preprocessing variability, and computational complexity persist, hindering clinical adoption. The review underscores the importance of large-scale, diverse datasets, multi-modal signal integration, and explainable AI to foster clinical trust and facilitate ethical deployment. By addressing these challenges, this review highlights the potential of AI to revolutionize cardiovascular healthcare through early diagnosis, wearable technology, and real-time decision support, paving the way for precision medicine and improved patient outcomes.

Electroencephalogram (EEG) signal is receiving much attention from recent studies since it is highly associated with intrinsic emotion. However, EEG signals contain underlying factors of variations across different sessions of the same subject, which make it difficult to learn temporal relationships between successive time steps. To disentangle invariant features, we propose a feature re-weighting mechanism on the extracted EEG features for temporal sequence modeling. Based on this method, our proposed model, called Convolutional Channel Modulator for Transformer and LSTM networks (CCMTL), extracts emotion-related inter-channel correlations using convolution operations and emphasizes important features by generating a channel attention map. This attention map is then used to perform matrix multiplication on the extracted features, which helps the subsequent Transformer to focus on important affective features. Furthermore, the sequential temporal modeling enhances the overall model's capability to understand temporal relationships both in global and sequential contexts. Experimental settings on public EEG emotion datasets demonstrate the superiority of the proposed CCMTL, surpassing six state-of-the-art models. Our code is publicly available at https://github.com/affctivai/CCMTL.

Remote photoplethysmography (rPPG) offers significant potential for health monitoring and emotional analysis through non-contact physiological measurement from facial videos. However, noise remains a crucial challenge, limiting the generalizability of current rPPG methods. This paper introduces Diffusion-Phys, a novel framework using diffusion models for robust heart rate (HR) estimation from facial videos. Diffusion-Phys employs Multi-scale Spatial-Temporal Maps (MSTmaps) to preprocess input data and introduces Gaussian noise to simulate real-world conditions. The model is trained using a denoising network for accurate HR estimation. Experimental evaluations on the VIPL-HR, UBFC-rPPG and PURE datasets demonstrate that Diffusion-Phys achieves comparable or superior performance to state-of-the-art methods, with lower computational complexity. These results highlight the effectiveness of explicitly addressing noise through diffusion modeling, improving the reliability and generalization of non-contact physiological measurement systems.

The development of conductive polymer-based dry electrodes with high conductivity is promising for practical applications in daily life due to their biocompatibility, flexibility, lightweight, and comfort. The objective of this study is to demonstrate the feasibility of using a novel silicone-based dry electrode for measuring various bioelectrical signals.The silicone-based electrode, manufactured using an optimized polymer matrix, combines high conductivity with flexibility, ensuring superior wearability and reliable bioelectrical signal monitoring. To evaluate its performance, we compared its impedance and flexibility with those of a commercial electrode. Additionally, its compatibility for measuring biological signals was assessed through performance comparisons across various bioelectrical signals. Fourteen healthy participants performed three experimental paradigms: (1) eyes closed and open to measure alpha electroencephalography (EEG) as well as resting-state electrocardiography (ECG), (2) eye blinking to measure electrooculography (EOG), and (3) wrist movement to measure electromyography (EMG). All bioelectrical signals were measured simultaneously using both the silicone-based dry electrode and a commercial dry electrode. The performance of the silicone-based dry electrode was evaluated by comparing the signal quality of both electrodes. The silicone-based dry electrode exhibited lower electrical impedance (39.43 kΩ on average, p = 0.0058) and greater flexibility (Young's modulus: silicone 1.51 ± 0.10 MPa vs. commercial 2.46 ± 0.38 MPa) compared to the commercial dry electrode. Overall, there were minimal differences in signal quality between the two electrodes: i) EEG (α power SNR: silicone 1.39 ± 0.34 vs. commercial 1.36 ± 0.29), ii) ECG (R-peak recall: 99.20 ± 2.50%, correlation coefficient: 0.96 ± 0.08), iii) EOG (eye blink recall: 100.00%, correlation coefficient: 0.98 ± 0.03), and iv) EMG (no significant difference in SNR values). These findings indicate that the developed electrode not only ensures superior flexibility but also maintains compatible electrical properties for measuring various bioelectrical signals.

Three-dimensional (3D) tumor spheroid models closely mimic in vivo tumor environment and play a vital role in studying oncological research. Despite their significance, the existing methods for analyzing 3D tumor spheroids often suffer from limitations, including low throughput, high cost, and insufficient resolution. To address these challenges, we developed a portable imaging system for the real-time sensing and quantitative analysis of the 3D tumor spheroids. The system integrated the seamless workflow of spheroid generation, cell morphology tracking, and drug screening. The spheroid generation was successfully characterized using MCF-7 breast cancer cells by optimizing cell concentration (5-20 × 10⁶ cells/mL), incubation time (24-96 h) and microwell diameter (400-600 μm). A custom-written algorithm was developed for automated analysis of spheroids, exhibiting high sensitivity (98.99%) and specificity (98.21%). Confusion matrices and receiver operating characteristic curve analysis further confirmed the robustness of the algorithm with an area under the curve value of 93.75% and an equal error rate of 0.79%. Following the characterization, the real-time sensing of spheroid generation and the response of spheroids to drug treatment were successfully demonstrated. Furthermore, the live/dead assays with chemotherapy provided a detailed insight into the efficacy and cytotoxic effects of the drug, demonstrating a significant dose-dependent decrease in a spheroid viability. Therefore, our system offers considerable potential for enhancing drug development processes and personalized treatment strategies, thereby contributing to more effective cancer therapies.

Supplementary information: The online version contains supplementary material available at 10.1007/s13534-025-00470-y.

Human-computer interaction (HCI) focuses on designing efficient and intuitive interactions between humans and computer systems. Recent advancements have utilized multimodal approaches, such as electroencephalography (EEG)-based systems combined with other biosignals, along with deep learning to enhance performance and reliability. However, no systematic review has consolidated findings on EEG-based multimodal HCI systems. This review examined 124 studies published from 2016 to 2024, retrieved from the Web of Science database, focusing on hybrid EEG-based multimodal HCI systems employing deep learning. The keywords used for evaluation were as follows: 'Deep Learning' AND 'EEG' AND ('fNIRS' OR 'NIRS' OR 'MEG' OR 'fMRI' OR 'EOG' OR 'EMG' OR 'ECG' OR 'PPG' OR 'GSR'). Main topics explored are: (1) types of biosignals used with EEG, (2) neural network architectures, (3) fusion strategies, (4) system performance, and (5) target applications. Frequently paired signals, such as EOG, EMG, and fNIRS, effectively complement EEG by addressing its limitations. Convolutional neural networks are extensively employed for spatio-temporal-spectral feature extraction, with early and intermediate fusion strategies being the most commonly used. Applications, such as sleep stage classification, emotion recognition, and mental state decoding, have shown notable performance improvements. Despite these advancements, challenges remain, including the lack of real-time online systems, difficulties in signal synchronization, limited data availability, and insufficient explainable AI (XAI) methods to interpret signal interactions. Emerging solutions, such as portable systems, lightweight deep learning models, and data augmentation techniques, offer promising pathways to address these issues. This review highlights the potential of EEG-based multimodal HCI systems and emphasizes the need for advancements in real-time interaction, fusion algorithms, and XAI to enhance their adaptability, interpretability, and reliability.

Cardiovascular diseases (CVD) are the leading cause of death worldwide, with heart failure (HF) being one of the most fatal conditions within CVD, greatly impacting patients' quality of life and imposing a heavy socioeconomic burden. Early intervention can significantly reduce HF mortality and hospitalization rates. However, current diagnostic methods are often expensive and complex, leading to delayed detection. To address this issue, this paper proposes a multimodal model, ECGEL, which combines electrocardiogram (ECG) and clinical text data for heart failure prediction. The model first denoises 12-lead ECG signals using LUNet, then converts the ECG signals into spectrograms via fast Fourier transform, extracting ECG features using EfficientNetv2. Simultaneously, clinical text is preprocessed with Bert, and textual features are extracted using BiLSTM. Finally, the ECG and text features are fused for heart failure prediction. Experimental results show that the ECGEL model achieved outstanding performance on a private dataset, with accuracy of 97.9%, recall of 98.3%, and F1 score of 97.6%. This model offers an efficient and accurate solution for the early diagnosis of heart failure, showing significant potential for clinical application.

Recent research has made significant progress with definitively identifying individuals with Parkinson's disease (PD) using speech analysis techniques. However, these studies have often treated the early and advanced stages of PD as equivalent, overlooking the distinct speech impairments and symptoms that can vary significantly across the various stages. This research aims to enhance diagnostic accuracy by utilizing advanced optimization strategies to combine speech recognition results (character error rates) with the acoustic features of vowels for more rigorous diagnostic precision. The dysphonia features of three sustained Korean vowels /아/ (a), /이/ (i), and /우/ (u) were examined for their diversity and strong correlations. Four recognized machine-learning classifiers: Random Forest, Support Vector Machine, k-Nearest Neighbors, and Multi-Layer Perceptron, were employed for consistent and reliable analysis. By fine-tuning the Whisper model specifically for PD speech recognition and optimizing it for each severity level of PD, we significantly improved the discernibility between PD severity levels. This enhancement, when combined with vowel data, allowed for a more precise classification, achieving an improved detection accuracy of 5.87% for a 3-level severity classification over the PD "ON"-state dataset, and an improved detection accuracy of 7.8% for a 3-level severity classification over the PD "OFF"-state dataset. This comprehensive approach not only evaluates the effectiveness of different feature extraction methods but also minimizes the variance across final classification models, thus detecting varying severity levels of PD more effectively.