부산대학교 도서관

Sepsis is a common disease with prohibitive medical costs and life-threatening consequences. Early prediction of sepsis and initiation of antibiotics is widely recognized as essential determinant of patient survival. In this work, this study proposes a novel sepsis accuracy prediction method called Cloud NAS-based Sepsis prediction system, which utilizes cloud technology (Cloud NAS Framework) - Container technique and neural architecture search/optimization technology. Furthermore, this study fully integrates the hyperparameter tuning mechanism and the matching neural network architecture search. The end user can use the model recommended by this study to land and make clinical sepsis and other related predictions to assist the first-line judgment. To verify the usability of this method, we also applied several public Datasets (e.g., MIMIC-IV) to conduct practical tests. The experimental data shows that this study not only searches for the model but also optimizes the hyperparameters in a low-cost way and mainly uses 2 different datasets for training and validation. Use the 2019 Pysionet dataset for model training and testing and MIMIC-IV for final verification. And the last model has an AUROC score of 0.95 within 12 hours of prediction and can predict the onset of sepsis eight hours earlier.