資料探勘在藥物安全上的應用：抗生素初次用量的反應預測

The Application of Data Mining in Drag Safety: Prediction of Outcome for Primary Antibiotic Treatment

用藥疏失常造成醫療資源浪費問題，進而增加不少的社會成本。泛可黴素因藥物作用範圍狹窄，若無法充分掌握藥物的作用結果，常會引起藥物在人體的毒性反應或抗藥性的副作用。在臨床上，醫療機構雖以藥物血中濃度監測程序(TDM)協助醫藥人員監控藥物在病患身上的作用結果，藉以調整藥物的用量與使用方式。然而，對於初次使用泛可黴素的病患而言，TDM程序卻無法有效地協助醫藥人員事先評估泛可黴素對於病患的影響，因此造成TDM程序在確保全面用藥安全上的限制。資料探勘技術已經廣泛地被應用在各項醫療研究上，也印證資料探勘從資料中所萃取潛在知識可用於醫療決策輔助的可行性。本研究嘗試運用資料探勘技術中的C4.5決策樹分析法，從醫療單位對泛可黴素進行TDM監控的歷史案例中，建構出可用以預測泛可黴素在病患身上的作用結果之分類模式，用以協助醫藥人員掌握泛可黴素的療效，進而提升泛可黴素的用藥醫療品質、降低可能的醫療資源浪費。實驗結果顯示，本研究所收集的TDM監控資料中有著類別間資料數量不對稱的現象，而造成模式預測偏向的問題。經由增加少數法的運用與適當複製比例的選擇，本研究可以利用調整類別資料數量比例後的資料集合，建構出對各類別資料都有良好預測效能的理想模式。

Inadequate medication usually results in wasted medical resources and increased social costs. Vancomycin can induce potent adverse side effects such as toxic response or antibiotic resistance if drag concentration is not controlled within a narrow efficacy / safety range. Measurement of circulatory drug concentration (TDM) is usually employed to adjust dose and help monitor treatment effects. Because TDM procedures are not helpful in patients taking vancomycin for the first time, use of TDM to ensure universal medication safety is limited. Data mining has been widely used in various aspects of medical research. The feasibility of data mining application in the extraction of potential beneficial and accessory information needed to assess medical strategies has been proved The present study aimed at using C4.5 strategic tree analysis data-mining techniques to construct a categorical model for predicting vancomycin effects with use of TDM monitoring data. Such a model would be greatly helpful in controlling effects of vancomycin treatment to promote medication quality and reduce putative waste of medical resources associated with vancomycin treatment. Our results demonstrate that the asymmetric quantity among categories in the TDM monitoring data base collected in the present study would result in deviation of model prediction. An ideal model with good predictive efficacy toward each data category could be established in the present study by adjusting the quantity ratio among data categories through over-sampling of the minority data and choosing adequate replication ratio.