使用資料探勘來分析與預測麻醉術後噁心嘔吐

Analysis and Prediction of Postoperative Nausea and Vomiting by Data Mining Techniques

王芷葳(Jhih-Wei Wang) ； 程毅君(Yih-Giun Cherng) ； 溫信財(Hsyien-Chia Wen) ； 張偉斌(Wei-Pin Chang)

《北市醫學雜誌》 16卷4期 (2019/12) Pp. 329-343

https://doi.org/10.6200/TCMJ.201912_16(4).0006

摘要

目的：本研究主要以台灣北部某準醫學中心醫院為實證基礎，利用資料探勘技術，建構預測術後噁心嘔吐高危險族群的特性模型，提供醫療人員臨床診斷、照護之參考。方法：先利用SPSS 18.0之卡方檢定與羅吉斯迴歸找出術後噁心嘔吐的相關因子與影響因子，再採用R語言3.5.3版將資料分為訓練集與驗證集，進行羅吉斯迴歸（GLM）、決策樹（rpart）、支援向量機（SVM）及類神經網絡（neuralnet）模組預測，以準確率、特異性、敏感性、陽性預測值、陰性預測值、F1值，作為演算法預測能力評估。結果：與術後噁心嘔吐有顯著相關的有性別、吸菸狀況、酗酒狀況、術後症狀（包含：術後喉嚨痛、術後暈眩）、手術類型。在羅吉斯迴歸分析，女性有術後噁心嘔吐的風險是男性的2.978倍。術後有喉嚨痛者有術後噁心嘔吐的風險是無喉嚨痛者的2.305倍。術後有暈眩者有術後噁心嘔吐的風險是術後無暈眩者的2.943倍。手術類別之減肥手術有術後噁心嘔吐的風險是骨折復位手術的4.528倍。在預測方面的準確率，羅吉斯迴歸為64.52％、決策樹為71.83％、支持向量器為75.65％、類神經網路為75.30％。結論：藉由本次研究可找出影響手術後噁心嘔吐的因子以及鑑別不同預測模型之優劣。

關鍵字

術後噁心嘔吐 ； 資料探勘 ； 預測模型

並列摘要

Objective: This study was based on the data of a Would-be Academic Medical Center hospital in northern Taiwan. It used data mining technology to construct a characteristic model for predicting the high-risk group of postoperative nausea and vomiting (PONV), and provided a reference for clinical diagnosis and care. Methods: Firstly, using the chi-square test of SPSS 18.0 and logistic regression, the related factors and influencing factors of postoperative nausea and vomiting were identified, and then the R software version 3.5.3 was used to divide the data into training set and verification set for further logistic regression (GLM), decision tree (rpart), support vector machine (SVM), and neural network (neuralnet) module prediction, with accuracy, specificity, sensitivity, PPV, NPV, and F1 value to index the performance of the algorithm. Result: Factors significantly associated with PONV were gender, smoking, alcohol abuse, and postoperative symptoms (including postoperative sore throat, and postoperative dizziness). In the logistic regression analysis, the risk of women having PONV was 2.978 times higher than men. The risk of PONV in patients with sore throat after surgery was 2.305 times higher than those without. The risk of PONV in patients with dizziness after surgery was 2.943 times higher than those without. The risk of PONV in the surgical category of weight loss surgery was 4.528 times that of fracture reduction surgery. The prediction accuracy rate was 64.35% in logistic regression, 71.83% in decision Tree, 75.65% in SVM, and 75.30% in neural network. Conclusion: We identified factors that affect PONV and the pros and cons of different predictive models.

並列關鍵字

Postoperative nausea and vomiting ； Data mining techniques ； Predictive model