Sepsis is a leading cause-of-death in hospitalized patients worldwide. As the aging population and increasing number of immunocompromised patients, patients admitted to intensive care units (ICUs) has steadily increased in the past few decades. In 2017, sepsis affects up to 48.9 million worldwide with 11 million deaths, and the estimated incidence of sepsis in Taiwan was 643/100,000, with approximately 15,000 patients per year. ICU is a specialized unit with a number of multidomain and continuous physiological and ventilatory parameters; however, most of data were not stored.

There is still lack of predictive model for sepsis by using electronic health records (EHR). EHR has been successfully used for predicting occurrences of a variety of diseases. However, inadequate predictive performances have been observed in cases of rare occurrences due to both insufficient training samples and highly imbalanced class distribution. In our recent work, we have successfully applied transformer model for temporal sequence data for young stroke prediction and deep multimodal fusion algorithms for bleeding event prediction. These methods can achieve higher performance for prediction of disease occurrence than traditional approaches. In this study project, clinical databases will be analyzed by using deep multimodal fusion algorithms with transformer model for temporal sequence data approach. We aim to use the novel method for leveraging cross database information to achieve an intelligent prediction model for sepsis.