Current studies still lack the use of electronic health data to establish predictive models of death and respirator disengagement in critically ill patients. Electronic health records have been successfully used to predict the occurrence of various diseases, however, due to insufficient training samples and a high degree of imbalance in the distribution of categories, prediction performance is often insufficient in the case of rare occurrences.
 

In the team's recent research, our team successfully applied a time-series data Transformer deep model for stroke prediction in young populations, a deep multi-mode fusion algorithm is applied to predict bleeding events. These methods enable higher disease prediction performance compared to traditional methods. In the past two years, we have also established the TCVGH’s 2015-2020 critical care database based on the international critical care database (e-ICU) architecture and applied the above deep learning methods to establish respirator detachment and long-term prognosis prediction modules. 

In this study, our team will combine a transformer deep model using a deep multi-mode fusion algorithm and timely sequence data to analyze a clinical information database. The team's goal is to use new methods to achieve intelligent predictive models that use cross-database information to explore death and respirator disengagement in critically ill patients. The team plan to build models with deep learning and make cross-validation of machine learning modules of SubProgram 3 to optimize modules.