The Prediction of Ventilator Weaning-A Comparison between Data Mining Analysis and Weaning Parameters

• Main Authors: Chia-Ming Chen, 陳嘉銘
• Other Authors: Chwei-Shyong Tsai
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/52794678309859230426

碩士 === 國立中興大學 === 資訊管理學系所 === 104 === Ventilator dependent patients accounted for a lot medical costs of NHI. In recent years, hospice medical care keeps costs of ventilator patients under control. Early predicting outcome of ventilator patients is necessary for physician-patient communication. Widely used weaning parameters are not suitable for the early prediction. This study employs data mining methods to construct early prediction models, to identify important predictors and retrieve and rules. Between January 2014 and December 2014, 300 patients with respiratory failure using mechanical ventilator in the intensive care unit of a regional hospital were included. We recorded demographic and clinical data and analyzed with decision tree (DT), K-nearest neighbor (KNN) and support vector machine (SVM) methods and weaning parameters. We found that admission department, admission source, and the reason for using ventilator were significant predictors of ventilator weaning. The prediction models constructed by DT, KNN and SVM have good performance on both weaning success and failure groups and the area under the ROC curve (AUC) are up to 0.848, 0.878 and 0.814 respectively. We employs weaning parameters such as rapid shallow breathing index (RSBI), respiratory rate (RR), and the maximum inspiratory pressure (PiMax) to predict weaning outcome of patients who has been treated and achieved stable status. The AUC of RSBI, RR, and PiMax only are 0.570, 0.5082 and 0.560 respectively. Our methods have better performance than weaning parameters in early prediction. The rules retrieved from decision tree model were reasonable and supported by other studies. It demonstrates the feasibility of data mining applications in the medical field, but also highlights the value of this study. After being validated by a greater amount of data, our study could be valuable in clinical practice.