Detecting and treating older adults who fall in an environment without others is essential. Millimeter-wave radar sensors do not have the disadvantage of invading user privacy like cameras, nor do they require users to wear them in real-time like wearable devices. Actual samples of older adults fall are difficult to collect, and it is unethical to require older adults to fall repeatedly to collect data. In addition, different body types and action patterns will inevitably reduce the model’s performance when new users use the model. In this paper, we constructed a fall detection model based on anomaly detection. The model is trained only using non-fall samples and detects falls as abnormal actions. The proposed model uses a domain generalization architecture based on domain feature alignment to extract domain-invariant features of the model, thereby improving the model’s generalization ability. In addition, we introduced the idea of denoising learning into the feature extractor and feature predictor to improve the model’s anti-interference ability. We conducted sufficient experiments to explore the effectiveness of the proposed method. When tested with new domain data, the proposed model has a true positive rate of 96.12%, a false positive rate of 0.97%, and an area under the receiver operating characteristic of 0.9979.

Multi-label arrhythmias classification is of great significance to the diagnosis of cardiovascular disease, and it is a challenging task as it requires identifying the label subset most related to each instance. In this paper, by integrating a deep residual neural network and auto-encoder, we propose an advanced deep neural network (DNN) framework with unified feature-aware and label embedding to perform multi-label arrhythmias classification involving 30 types of arrhythmias. Firstly, a deep residual neural network is built to extract the complex pathological features from varying-dimensional electrocardiograms (ECGs). Secondly, the mean square error loss is adopted to learn a latent space associating the deep pathological features and the corresponding label data, and then to achieve unified feature-label embedding. Thirdly, the label-correlation aware loss is introduced to optimize the auto-encoder architecture, which enables our model to exploit label-correlation for improved multi-label prediction. Our proposed DNN model can allow end-to-end training and prediction, which can perform feature-aware, label embedding, and label-correlation aware prediction in a unified framework. Finally, our proposed model is evaluated on the currently largest public dataset worldwide, and achieves the challenge metric scores of 0.492, 0.495, and 0.490 on the 12-lead, 3-lead, and all-lead version ECGs, respectively. The performance of our approach outperforms other current state-of-the-art methods in the leave-one-dataset-out cross-validation setting, which demonstrates that our approach has great competitiveness in identifying a wider range of multi-label arrhythmias.