Human vital signs are essential to the healthcare industry and applications of Internet of medical things (IoMT). As a significant vital sign signal, heart rate variability (HRV) provides into the functioning of the human stress levels and overall well-being. However, it is inconvenient for daily long-term HRV monitoring with conventional monitoring methods under the uncomfortable feelings. Moreover, it is hard to detect anomalies in HRV data to provide real-time previews about vital signs. To address these problems, a novel fiber interferometer-based optical fiber sensor is proposed to monitor human vital signs, and we propose a novel deep learning model (TWFN). First, a novel module (ADSN) is proposed to apply graph modeling for obtaining spatial and temporal characteristics of HRV. Subsequently, an unsupervised generative adversarial network (VS-GAN) is proposed for the effective overcoming mode collapse and generator failure to converge to the target distribution better. The outcomes of the experiment might encourage the use of HRV-based healthcare application in IoMT-enabled healthcare sectors.

Objectives: To investigate the association of sporadic and bouted steps with all-cause mortality. Bouted steps were defined as those accumulated during walking bouts lasting at least 10 minutes. Method: A cohort study of 3072 participants (mean age: 48.6 y, 51.7% female) from the United States National Health and Nutrition Examination Survey 2005–2006. Baseline step counts were recorded using a waist-worn accelerometer over a 7-day period. Mortality was ascertained through December 2019. A random forest classifier, trained on “ground truth” image data from the Capture-24 study, was used to differentiate between sporadic and bouted steps. Sporadic and bouted steps were mutually adjusted in the Cox model after controlling for important confounders. Results: Over a mean follow-up period of 13.1 years (2.7 y), 632 deaths were identified. Each 1000-step increase in sporadic steps was associated with a 10% (95% CI, 7%–13%) reduction in all-cause mortality. Each 1000-step increase in bouted steps was associated with a 27% (17%–35%) reduction. In the joint analysis, we observed an 80% reduction in all-cause mortality among individuals who took either 6000 steps (comprising 4000 sporadic and 2000 bouted steps) or 10,500 steps (comprising 10,000 sporadic and 500 bouted steps), compared with the reference group (2000 sporadic and 0 bouted steps). Conclusion: Both sporadic and bouted steps were inversely associated with all-cause mortality. More sporadic steps were better than less, but increasing bouted steps led to more rapid reductions in mortality. Health-benefiting daily step goals can be achieved through various combinations of these 2 step patterns.

A large group of people, nowadays, has been suffering from chronic sleep disorders and diseases, resulting in wide attention on sleep quality assessment. Conventional sleep-staging networks frequently consider multiple channel inputs, hindering the feasibility of the network to single-channel input or other sensor data input. In this paper, we proposed an Auto-SleepNet: a CPU-driven and end-to-end deep learning network for sleep stage classification using single-lead electroencephalogram (EEG) signals. The network is composed of a tailored Auto-Encoder for feature extraction and correction, and an LSTM network for temporal-signal classification. Compared with multi-lead connections, our design renders a higher accuracy in comparison to state-of-the-art, provides a meaningful reference for simplifying the hardware requirements of the EEG measurement device, and simultaneously lowers the computational loads significantly. We used the Per-Class precision (PR), Recall (RE), Per-Class F1 Score, overall accuracy, confusion matrix, and Cohen’s kappa coefficient (κ) to evaluate the performance. The overall accuracy, RE, and Cohen’s Kappa of our model are 95.7%, 95.19% and 0.91, respectively. Compared to state-of-the-art methods mentioned in the paper, Auto-SleepNet outperforms single-channel methods by 13.97%, and multiple-channel methods by 15.97% on average. Furthermore, it is not compulsory to use a GPU to train our Auto-SleepNet. Experiments show that our model can converge in 15.6 minutes using a CPU only. The results highlight the practicability of the network to sleep stage classification problems.

The increasing prevalence of chronic sleep disorders has intensified the need for accessible sleep monitoring. Existing sleep-staging models often rely on multi-channel inputs, thereby limiting their adaptability. Unlike comparable methods that frequently utilize multiple channels of information in a polysomnography, such as electrocardiogram, electrooculogram, and electromyogram, this paper proposes Auto-SleepNet, a central processing unit (CPU)-driven and end-to-end deep learning network for sleep stage classification using only a single-lead electroencephalogram. The architecture could make fully use of given information for the classification task. Compared to state-of-the-art methods, Auto-SleepNet achieves superior accuracy while significantly reducing computational cost. The proposed model was evaluated by Per-Class Precision, Recall, F 1-score, overall accuracy, confusion matrix, and Cohen’s kappa ( κ ). The model attained an accuracy of 92.67% and κ = 0.85, which outperforms single-channel benchmarks by 9.53% and multi-channel ones by 10.32% on average. Furthermore, the model can be trained on a CPU in just 15.6 min, demonstrating that it does not require a graphics processing unit (GPU) for acceleration and this highlights the efficiency of the learning strategy. The public dataset used in this paper could be downloaded from URL: https://researchdata.ntu.edu.sg/dataset.xhtml?persistentId=doi:10.21979/N9/MA1AVG .