Electronic health records (EHRs) have been widely adopted among modern hospitals to collect and track clinical data. Secondary analysis of EHRs could complement the traditional randomized control trial (RCT) research model. However, most researchers in China lack either the technical expertise or the resources needed to utilize EHRs as a resource. In addition, a climate of cross-disciplinary collaboration to gain insights from EHRs, a crucial component of a learning healthcare system, is not prevalent. To address these issues, members from the Massachusetts Institute of Technology (MIT) and the People’s Liberation Army General Hospital (PLAGH) organized the first clinical data conference and health datathon in China, which provided a platform for clinicians, statisticians, and data scientists to team up and address information gaps in the intensive care unit (ICU).

Physiological signals can contain abundant personalized information and indicate health status and disease deterioration. However, in current medical practice, clinicians working in the general wards are usually lack of plentiful means and tools to continuously monitor the physiological signals of the inpatients. To address this problem, we here presented a medical-grade wireless monitoring system based on wearable and artificial intelligence technology. The system consists of a multi-sensor wearable device, database servers and user interfaces. It can monitor physiological signals such as electrocardiography and respiration and transmit data wirelessly. We highly integrated the system with the existing hospital information system and explored a set of processes of physiological signal acquisition, storage, analysis, and combination with electronic health records. Multi-scale information extracted from physiological signals and related to the deterioration or abnormality of patients could be shown on the user interfaces, while a variety of reports could be provided daily based on time-series signal processing technology and machine learning to make more information accessible to clinicians. Apart from an initial attempt to implement the system in a realistic clinical environment, we also conducted a preliminary validation of the core processes in the workflow. The heart rate veracity validation of 22 patient volunteers showed that the system had a great consistency with ECG Holter, and bias for heart rate was 0.04 (95% confidence interval: −7.34 to 7.42) beats per minute. The Bland-Altman analysis showed that 98.52% of the points were located between Mean ± 1.96SD. This system has been deployed in the general wards of the Hyperbaric Oxygen Department and Respiratory Medicine Department and has collected more than 1000 cases from the clinic. The whole system will continue to be updated based on clinical feedback. It has been demonstrated that this system can provide reliable physiological monitoring for patients in general wards and has the potential to generate more personalized pathophysiological information related to disease diagnosis and treatment from the continuously monitored physiological data.

Understanding the sleep quality and architecture is essential to human being's health, which is usually represented using multiple sleep stages. A standard sleep stage determination requires Electroencephalography (EEG) signals during the expensive and labor-intensive Polysomnography (PSG) test. To overcome this inconvenience, cardiorespiratory signals are proposed for the same purpose because of the easy and comfortable acquisition by simplified devices. In this paper, we leverage our low-cost wearable multi-sensor system to acquire the cardiorespiratory signals from subjects. Three novel features are designed during the feature extraction. We then apply a Bidirectional Recurrent Neural Network architecture with Long Short-term Memory (BLSTM) to predict the four-class sleep stages. Our prediction accuracy is 80.25% on a large public dataset (417 subjects), and 80.75% on our 32 enrolled subjects, respectively. Our results outperform the previous works which either used small data sets and had the potential over-fitting issues, or used the conventional machine learning methods on large data sets.

No abstract