利用資料探勘預測病人自控麻醉之劑量使用

Abstract

病人在手術後疼痛會受很多因素影響，儘管多項統計研究已經對術後疼痛與術後止痛麻醉劑量使用有相關探討，可以看出預測模型的相關係數R2並不高(術後疼痛的R2在0.17-0.59之間，而術後止痛劑量的R2則在0.27-0.46之間)。此研究呈現一組現實的應用裝置，能夠參考各式預測因素與病人在PCA上的使用行為，將資料探勘的技術致力於麻醉醫學。我們透過延伸先前在PCA上的研究，結合分群、分類與迴歸等監督與非監督學習演算法，研發出多迴歸模型的樹狀結構。而利用交叉驗證以及與麻醉醫學專家比較的驗證實驗中，兩者的結果都顯示此應用裝置的準確度與可行性。

Many factors affect individual variability in postoperative pain. Although several statistical studies have evaluated postoperative pain and analgesic consumption, precious research shows that the coefficient of determination of existing predictive models was small (e.g., R2 = 0.17-0.59 for postoperative pain, and 0.27-0.46 for postoperative analgesic consumption). This thesis presents the real-world application of data mining to anesthesiology, and considers a wider variety of predictive factors, including PCA demands over time. It extends previous works by proposing a multi-model regression strategy that combines clustering, classification, and regression to predict analgesic consumption. The results of the cross validation, and the comparison with human experts have demonstrated the feasibility of the proposed computational methods.