Multi-label electrocardiogram (ECG) classification is of great significance for the diagnosis of cardiovascular diseases, and it is also a challenging task because of the various of pathological patterns across different cardiac abnormalities. This paper analyzes the complex spatial-temporal variations information contained in the 12-lead ECG from a new perspective, and proposes an advanced deep neural network framework to accurately identify various multi-label ECG abnormalities. Firstly, the 12-lead ECG was modeled as 3-dimensional (3-D) spatial-temporal variations involving the intra-period, inter-period temporal dimensions, and cross-lead spatial dimension. Then, a 3-D spatial-temporal learning network (3D-STLNet) was carefully designed to extract the complex pathological features from the 3-D spatial-temporal variations. Secondly, the multi-head attention (MHA) mechanism was introduced into the proposed network to extract global temporal dependencies in ECG segments, which further improved the performance of the proposed network. By integrating a residual backbone network, a 3D-STLNet, and the MHA module, our proposed network allows end-to-end learning and prediction in a unified framework. The influence of several classic threshold selection strategies on the performance of multi-label ECG classification algorithm was also analyzed, which provides a reference for subsequent research on optimal threshold selection for multi-label ECG classification. Finally, our proposed network was evaluated on multiple large multi-label ECG public datasets, and the performance of our network outperforms that of current state-of-the-art networks, which demonstrates that our network has great competitiveness in identifying a wider range of multi-label ECG abnormalities.

Pulse signals can serve as important indicators of one’s cardiovascular condition. However, capturing signals with stable morphology using radar under varying measurement periods remains a significant challenge. This paper reports a non-contact arterial pulse measurement method based on mmWave radar, with stable signals achieved through a range–angle focusing algorithm. A total of six subjects participated in the experiment, and the results showed that, under different measurement times, the pulse morphology of the same body part for each subject had good consistency, reaching a correlation of over 0.84, and four selected pulse signs remained stable. This is a quantitative assessment revealing a high correlation in pulse morphology measured by radar over different periods. In addition, the influence of array size and measurement distance was analyzed, providing a reference of array selection for research work with different requirements. This work offers an effective reference framework for long-term pulse measurement using radar technology.

It is essential to find and treat the elderly in time after they fall. Millimeter wave radar sensor not only solves the problem of invasion of privacy caused by the camera but also avoids letting patients wear wearable devices, which has received more and more attention in the field of fall detection and motion recognition. Due to the different modulation of radar signals by different users and environments, the performance of the fall detection model is inevitably reduced. In this paper, we propose a highly generalized radar-based fall detection framework to improve the generalization of the fall detection system in the new domain. We use a feature extractor with a ResNet structure constructed from three-dimensional convolution to extract features of the signal, and a multilayer perceptron as a classifier for fall detection. We measure the cross-domain cross entropy of the features by setting subclassifiers for each domain to measure the cross-domain invariance of the features, enabling the feature extractor to extract domain-invariant features. We propose a specific data augmentation technology for radar time series signals, which is combined with the proposed model to achieve more generalized fall detection. We have conducted sufficient experiments to explore the relationship between the number of users and domains and the model’s generalization. The proposed method achieves an F1 score of 0.9747 on the proposed dataset containing 10 scenarios, 21 subjects, 52 non-falling actions and 12 falling actions, and an F1 score of 0.9949 on the public dataset, which shows superior effectiveness.

Emotion recognition is vital for mental and physical health, but current methods often rely on active user participation (e.g., questionnaires) or wearable sensors, limiting their real-world adoption. This paper proposes a non-contact BCG-based cross-subject emotion recognition framework. Subject-specific data distribution variations challenge model generalizability across domains. For this reason, we propose a LSCS-BDA method, extending Balanced Distribution Adaptation (BDA) with a Label Similarity Clustering for Selection (LSCS) module to optimize source domain selection. After feature extraction and baseline normalization, only the source domain individuals that are more similar to the target are included in the final source domain set through our method’s LSCS module (reducing training data to 9.78%), and finally the transfer is completed by the BDA algorithm. The experimental results show that the classification accuracy of our proposed method for three emotions of positive, negative, and neutral is 70.38%, which achieves higher accuracy and faster model training than classical transfer algorithms. The LSCS module in our method can be extended to other transfer learning methods. Our approach offers a new idea for contactless emotion recognition based on BCG and contributes to advancing its applications.