Construction of Early Warning Model for Pulmonary Complications Risk of Surgical Patients Based on Multimodal Data Fusion

Overview

The goal of this observational study is to establish an intelligent early warning system for acute and critical complications of the respiratory system such as pulmonary embolism and respiratory failure. Based on the electronic case database of the biomedical big data research center and the clinical real-world vital signs big data collected by wearable devices, the hybrid model architecture with multi-channel gated circulation unit neural network and deep neural network as the core is adopted, Mining the time series trends of multiple vital signs and their linkage change characteristics, integrating the structural nursing observation, laboratory examination and other multimodal clinical information to establish a prediction model, so as to improve patient safety, and lay the foundation for the later establishment of a higher-level and more comprehensive artificial intelligence clinical nursing decision support system. Issues addressed in this study 1. The big data of vital signs of patients collected in real-time by wearable devices were used to explore the internal relationship between the change trend of vital signs and postoperative complications (mainly including infection complications, respiratory failure, pulmonary embolism, cardiac arrest). Supplemented with necessary nursing observation, laboratory examination and other information, and use machine learning technology to build a prediction model of postoperative complications. 2. Develop the prediction model into software to provide auxiliary decision support for clinical medical staff, and lay the foundation for the later establishment of a higher-level and more comprehensive AI clinical decision support system.

The project prospectively collected the clinical information of patients in general surgery (gastrointestinal surgery, biliary surgery, pancreatic surgery), noninvasively monitored the patient's temperature, heart rate, ECG, respiratory rate through the body surface with wireless vital signs sensor, guided the patient to wear correctly from the first day of admission, and continuously monitored the patient's preoperative, intraoperative and postoperative vital signs throughout the whole process until leave hospital，The traditional "point" vital signs monitoring will be updated to continuous "line" monitoring, returning to the real world of patients' vital signs. The vital signs data of patients are continuously collected, transmitted in real-time and stored in the local central workstation of the ward, and the researchers of the project are specially assigned to be responsible for data export, storage and analysis. In view of the lack of early warning means for acute and critical complications in surgical ward, wearable devices with verified accuracy were used to collect continuous vital signs big data, fully mining the internal relationship between the change trend of patients' multiple vital signs parameters and complications, and establishing an intelligent risk warning model for perioperative complications of surgical patients, It lays the foundation for the establishment of a higher-level real-time patient risk warning and clinical decision support system, so as to improve the perioperative safety of patients and promote the penetration of artificial intelligence in clinical medicine and nursing. Participants will use the Clavien Dindo grading criteria to determine the severity of complications according to the Cohort study method. According to whether the patient has experienced complications and the type of complications, the patient is divided into different subgroups and the patient outcome is manually calibrated. Mining the change trend of patient's vital signs over time, Convolutional neural network is used to extract morphological features from the continuous vital sign curves of the past 24 hours, and two-way short-term memory neural network is used to extract temporal features to obtain two sets of feature vectors. All data are feature fused, and sparse features are specially processed, and then sent to another weighted recurrent neural network to establish a complication prediction model, and predict the patient's Respiratory failure risk level, ICU risk level, and death risk level in the next 24 hours. Continuously revise the algorithm and set alarm threshold logic based on personal baseline data. Compare the predictive power of this model with the predictive power obtained from the national warning score NEWS, and determine the sensitivity, specificity, false positive rate, false negative rate, positive predictive value, and negative predictive value of the two in predicting patient outcomes.