Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference最新文献

Punch biopsy is a skin biopsy method that is used to remove a small sample of the epidermis and dermis. The procedure requires a trained dermatologist to excise the sample with a punch biopsy tool and then immediately apply sutures for wound closure. Punch biopsy is a common procedure necessary for the diagnosis of many conditions. Thus, the present shortage of dermatologists-especially in the developing world-motivates the design of convenient methods for punch biopsy that do not require clinical training. Here, we present a medical device that streamlines punch biopsy and wound closure into simple, sequential operations. The handheld device engages and securely locks the skin, excises the biopsy sample, then applies a N-butyl cyanoacrylate tissue adhesive for wound closure. The device is validated ex vivo using porcine ear skin, which has comparable biomechanical properties to human skin. For engagement, the device can target the desired biopsy area with an accuracy of 2 mm. For sample collection, the device reliably excises samples 7 mm in diameter and 4 mm in depth. For wound closure, the device streamlines the application of tissue adhesive to seal the wound, albeit with less strength than surgical sutures. Altogether, these results validate the design of a device for punch biopsy and wound closure that can be used within sequential steps with minimal training.

Studying the interplay between respiration patterns and the heart rate variability (HRV) through mathematical models led to valuable insights into autonomic nervous system (ANS) functioning. Despite several models have been proposed in the literature, there is a lack of a general mathematical framework based on systems theory and formulated according to a rigorous control theory formalism. This work aims to reframe existing cardiopulmonary models into a general finite-dimensional nonlinear time-invariant (FDNTI) framework to capture respiration-cardiovascular interactions. A MATLAB Simulink-based implementation is presented and a simulation study is carried out. By exploiting control theory formalism, a system theoretic oriented model is obtained, which addresses roles of state variables, inputs, and linear/nonlinear contributions. Simulation tests confirmed the validity of the proposed modeling approach. This generalized formulation could enable in-depth analysis of physiological and pathological states by adopting advanced control theory techniques to investigate stability properties of the cardiorespiratory system.Clinical relevance- An in-silico model of respiration-cardiovascular interactions for assessing autonomic functioning.

Current research on steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) predominantly focuses on utilizing the frequency- and phase-locking characteristics of SSVEP for encoding purposes. In this study, we propose an innovative paradigm wherein SSVEP serves as a marker, integrated with different types of motion animations to identify distinct neural processing pathways associated with these animations. This approach enables the classification of SSVEP-based BCIs without relying on frequency features. We designed six distinct animations corresponding to six behaviors commonly observed in daily life. Each animation was tagged with a uniform 6 Hz stimulus frequency, forming a six-target classification task. Offline testing was conducted with 10 participants. Despite identical frequency components, significant differences in spatial distribution corresponding to the animations were observed, likely due to the behavioral variations in the animations. Classification analysis demonstrated an accuracy of 0.93 within a 6-second window, validating the practical feasibility of this approach. This paradigm offers a novel direction for the advancement of SSVEP-based BCIs, potentially enabling the integration of multi-sensory information.

Automatic sleep staging typically requires multi-channel EEG data, limiting its application in portable devices. To address this, we propose a hybrid deep learning model that utilizes multi-domain features from single-channel EEG data collected via polysomnography (PSG). Our model employs two feature extractors to capture time-domain and time-frequency-domain features, which are fused for final predictions. Validated on the Haaglanden Medisch Centrum Sleep Centre Database (HMC) with EEG data from 151 subjects, the model achieves an accuracy of 0.747 and an F1 score of 0.742. Compared to state-of-the-art methods, it shows improved multi-classification performance, particularly in N3 stage detection. This study highlights the potential of single-channel EEG for accurate sleep staging and the development of portable PSG-based monitoring systems.Clinical Relevance-This study develops a deep learning model for automatic sleep staging only using a single-channel EEG. Our research would be helpful to automatically classify stages during sleep for sleep physicians.

Sleep stage classification is a critical task in sleep research, with significant implications for diagnosing and treating sleep disorders. Traditional methods rely on manual scoring of polysomnography (PSG) data, which is time-consuming and prone to human error. While recent advances in deep learning have enabled automated sleep stage classification, challenges persist in handling the complex, non-linear patterns of physiological signals. Existing models are often computationally expensive, require sophisticated feature extraction methods, and are unsuitable for real-time implementation. To address these limitations, we propose a lightweight and efficient dual-branch deep-learning model that leverages the feature extraction capabilities of CNNs and the channel-wise attention mechanisms of Transformers. Unlike conventional transformers, it avoids excessive computational complexity while effectively capturing both local and global dependencies in physiological signals. The model is validated on four benchmark datasets-SleepEDF-20, SleepEDF-78, SleepEDFx, and SHHS-and demonstrates superior performance compared to several baseline algorithms. Our proposed algorithm achieves state-of-the-art results across all datasets, highlighting its robustness and scalability for real-world applications. The code for the proposed algorithm is publicly available at link, enabling reproducibility and further research. Combining the strengths of CNNs and Transformers, it offers a promising solution for accurate and efficient sleep stage classification, paving the way for improved diagnosis and treatment of sleep disorders. The code is available at https://github.com/ang-frozen/embc2025.

The morphology of the arterial blood pressure (ABP) waveform has been demonstrated to serve as a significant indicator of the patient's condition and a marker of impending changes. The wave separation analysis (WSA) approach is based on invasive measures of concomitant arterial blood flow (ABF) and ABP. Other methods were developed as well, but the analyses were limited to waveforms with the physiological shape, namely the Type A. This study introduces a bidirectional long short-term memory (BiLSTM) deep learning model to classify ABP beats into Type A and Type B/C, this last group refers to a condition of altered vascular compliance and resistance. The models were developed by using central (aortic) and peripheral (femoral) waveforms. The best models have achieved an accuracy of 96% and 90% for aortic and femoral signals, respectively. The ultimate objective of this research is to enhance non-invasive cardiovascular monitoring and facilitate the early detection of arterial alterations.

The increasing elderly population has made falls due to frailty a critical issue, with physical and cognitive factors interacting to elevate the risks of fractures and solitary deaths. Falls are challenging to predict due to the involvement of multiple factors, necessitating the development of continuous gait monitoring and fall detection technologies. Although various fall detection methods have been proposed, many rely on batteries requiring maintenance with circuit complexity, higher costs, or both. This study aims to develop an unintentional-falling detection system by utilizing triboelectric nanogenerator (TENG) technology to create a battery-free insole device. The device was tested to analyze the feature of the voltage signals produced by unintentional falls. The results suggest that the signals generated by the device are sufficiently distinguishable from the frequency-amplitude and the ratio of maximum and minimum voltage values during one gait cycle, indicating the feasibility of constructing a battery-free system capable of detecting unintentional-falling.

The rising number of total knee arthroplasty (TKA) revisions combined with the inferior outcomes compared to the primary TKA highlight the critical need for early detection of primary TKA failure. The present work aims to propose a radiomics-based machine learning model to automatically detect TKA failure from radiographs. The dataset comprised radiographs from 44 failed and 51 non-failed TKA patients. Following preprocessing phases, 465 radiomic features were extracted. A cross-validation procedure, consisting in 100 repeated training-validation splits was implemented. The training phase encompassed feature selection, data balancing and machine learning classifier training. Four feature selection approaches were evaluated combined with several classifiers. Based on the average performance metrics on the validation set, the Least Absolute Shrinkage and Selector Operator (LASSO) feature selection, combined with Logistic Regression (LR) classifier achieved the best performance, with an F1-score of 0.701, a balanced accuracy of 0.710 and area under the curve (AUC) of 0.783. The results demonstrate the potentialities of the developed radiomics-based approach in automatically detecting TKA failure from plain radiographs.Clinical Relevance-The increasing number of revision procedures poses significant challenges for healthcare systems, highlighting the critical need for automated early detection of primary TKA failure. The developed model can support clinicians by reducing their workload and minimizing inter- and intra-observer variability.

Closed-loop brain-computer interfaces (BCIs) hold promise for restoring function after neurological damage by dynamically processing neural signals and delivering targeted brain stimulation. To achieve clinically meaningful outcomes, such systems must operate with high spatiotemporal precision. This work aims to demonstrate a proof-of-concept neuromorphic BCI that processes neural spike events in near-real time, without necessitating preprocessing besides signal filtering and spike detection. Methods - We developed a system that acquires neural signals and streams spike events into a spiking neural network (SNN) running on SpiNNaker neuromorphic hardware. We evaluated the system's performance using both in vivo recordings from mouse visual cortex and simulated neural waveforms. We measured the roundtrip latency, defined as the time from spike detection to an output spike generated by the SNN. Results - Under baseline conditions with no hidden SNN layers, mean roundtrip latency was 4.69 ms (±1.70 ms). Adding hidden layers increased latency by approximately 3.65 ms per layer, reflecting the computational overhead of deeper networks. The system successfully detected and processed spikes in near real-time, demonstrating that neuromorphic hardware can manage spike-based input at speeds suitable for closed-loop intervention. Discussion - These findings indicate that neuromorphic SNNs can rapidly process neural signals, providing a foundation for closed-loop BCIs capable of bypassing damaged neural pathways. Future efforts will involve implementing stimulation protocols and functional SNNs. Such developments may ultimately facilitate more effective, flexible, and power-efficient neuroprosthetic devices.

Motor intent (MI)-based brain-computer interfaces (BCIs) have been extensively studied to improve the performance and clinical realization of assistive robots for motor recovery in stroke patients. However, challenges arise in their low decoding performance. This can be attributed to the low spatial resolution and signal-to-noise ratio of electroencephalography (EEG), particularly in accurately deciphering hand movements, which reduces classification performance. Therefore, we have developed a novel feature extraction technique that exploits Levant's differentiators to extract distinct patterns in EEG signals and employs symmetric positive definite matrices (SPD) to effectively leverage the spatial-temporal properties of the EEG signal. Results from nine post-stroke patients and fifteen normal subjects showed an improved decoding accuracy of 99.16±0.64% and 99.30±0.69%, respectively in classifying twenty-four hand motor intents, significantly outperforming existing related methods. Thus, the proposed technique has the potential to greatly enhance the reliability and effectiveness of EEG-based control systems for post-stroke rehabilitation.Clinical Relevance- The outcome of this study can lead to better control of rehabilitation robots and improve the recovery speed of the stroke patients.