透過案例式推理方法進行不平衡多重測量肝癌病患資料分析及處理

Abstract

在病患的臨床資料裡，有許多不同的用途，如果使用某項條件把整套資料一分為二，時常會有不平衡情況發生，意即某一邊的資料數量會多於另一邊的資料數量，而這樣的情況對於之後要拿來做分類或者預測的系統有著很大的影響，對於資料數量較多的一方，系統的訓練效能會比資料數量較少的一方良好許多，如此就會產生出有偏差的判定情況，在本研究中我們嘗試了常見的平衡資料模組的方法: 依大採樣 (Over-Sampling), 依小採樣 (Under-Sampling)，來對不平衡的肝癌病患資料作處理，且使用基於案例式推理原理的系統來進行復發的預測判定，同時我們也保留了不平衡的資料模組來當作一個比較的基準，根據系統的預測結果再進行靈敏度 (Sensitivity)、特異度 (specificity) 等相關統計，來比較各種處理資料方法對於預測的影響。

In nowadays, the medicine clinical data are increasing very rapidly and most clinical data usually have imbalanced data problem. In this study, over-sampling and under-sampling are used for handling data imbalanced condition. Case based reasoning is used for developing classification models to predict recurrent statuses of patients with liver cancer. Classification results of these two methods are compared with those of an original imbalanced data set. Classification results are evaluated by sensitivity, specificity, balanced accuracy (BAC), positive predictive value (PPV), negative predictive value (NPV), and accuracy. Experiment results appear that balanced data sets can provide benefits for classification models and efficiently reduce biased classification. Furthermore, we also use some feature selection methods to give the feature weights and rank the feature weights from the highest to lowest. Then, these features are added stepwise to train and evaluate classification models. According to evaluation results, we could realize that using how many features could have better classification results.