- 會議論文
- OpenAccess
Cloud-based Sepsis Prediction System with Machine Learning Hyperparameter Tuning Service
摘要
敗血症是一種常見疾病，醫療費用非常昂貴，且對生命是會有致命的影響。敗血症的早期預測和抗生素的開始被廣泛認為是患者存活的重要決定因素。在這項工作中，本研究提出了一種新的敗血症精確度預測方法，稱為雲端敗血症預測系統，它利用雲端技術－容器計算資源，以及主流超參數調整方法，如Hyperopt和Optunity。最後，本研究將系統用於實際數據集-MIMIC-III臨床數據庫數據集。除了敗血症預測的準確性之外，所提出的系統還有其自動化參數調整和突破可調參數數量限制的能力。而要如何有效率的使用雲端計算建立敗血症預測系統？首先，用戶可以使用用戶友好且直接的界面來操作雲端敗血症預測系統，以進行模型訓練和超參數調整。其次，雲端敗血症系統內核接收來自用戶的訊號，然後將驅動超參數調整機制。每個訓練工作都是容器化中發送，由Kubernetes管理與分配。最後，所搭配提出的系統將找到最佳超參數建議。在這種情況下，期望預測結果可以達到98％的準確度，AUROC可以高於0.8，由此可以驗證整個系統的可用性。
並列摘要
Sepsis is a common disease with very costly, potentially deadly implications. Early prediction of Sepsis and initiation of antibiotic is widely considered as an important determinant of patient survival. In this work, we proposed a novel approach towards sepsis precision prediction, called Cloud-based Sepsis prediction system, which leverages cloud technology - container computing resources, and mainstream hyperparameter tuning methods such as Hyperopt and Optunity. Finally, we dedicated our system to an actual dataset - MIMIC-III Clinical Database dataset for performance testing. Besides accuracy of sepsis prediction, the proposed system is also justified for its ability in automating parameter tuning and breaking through the limitation of the number of adjustable parameters. The proposed Cloud-based Sepsis prediction system is designed by research team in National Center for High-performance Computing. First of all, users can operate the Cloud-based Sepsis prediction system with user-friendly and straightforward interface for doing model training and hyperparameter tuning. Second, the Cloud-based Sepsis kernel receives the signal from users, and then will active hyperparameter tuning mechanism. Every single training job will be dispatched in container, which is managed by Kubernetes. Finally, the proposed system will find best hyperparameter suggestion. In this case we expect to achieve 98% accuracy and AUROC can be above 0.8, therefore we can proof that the proposed system is high availability.
延伸閱讀
- Ullah, Q. Z., Hassan, S., & Khan, G. M. (2017). Adaptive Resource Utilization Prediction System for Infrastructure as a Service Cloud. Computational Intelligence and Neuroscience, 2017(), 1-12. https://doi.org/10.1155/2017/4873459
- 張維晏、陳羿霖、陳煥、蔡崇煒（2020）。An Effective Hyperparameter Selection for Deep Learning Algorithm in Intrusion Detection System。資訊安全通訊，26(4)，1-16。https://www.airitilibrary.com/Article/Detail?DocID=a0000270-202012-202101270011-202101270011-1-16
- ARSHAD, J., JOKHIO, I. A., & TOWNEND, P. (2012). A Machine Learning Based Intrusion Impact Analysis Scheme for Clouds. Mehran University Research Journal of Engineering and Technology, 31(1), 107-118. https://www.airitilibrary.com/Article/Detail?DocID=P20170126001-201201-201702060032-201702060032-107-118
- Wu, C. Y. (2022). Early Prediction of Sepsis Using Machine Learning Methods on MIMIC Database [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202202105
- Zhao, S., Chen, H., Zhao, R., Zhao, Y., & Chen, G. (2016). A Big Data Processing-oriented Prediction Method of Cloud Computing Service Request. 淡江理工學刊, 19(4), 497-504. https://doi.org/10.6180/jase.2016.19.4.13