動態區間切割演算法使用於不規則縱向資料探勘

Abstract

射頻燒灼術是一種治療肝癌的方法。雖然在手術後可以根除腫瘤，但是肝癌復發是一項非常重要的議題。在這項研究裡，我們收集了因為肝癌而接受過射頻燒灼術治療病人們的就診臨床資料，並用來建立預測射頻燒灼術後肝癌復發與否的模型。這些臨床資料是以縱向資料的形式存在，包含兩種特徵，靜態特徵與時序特徵。我們提出了一種叫做動態區間分割的資料前處理方法，並結合時間摘要法，從原始的時序特徵萃取摘要化的元特徵。我們使用動態區間分割法把給定的時間區間切成若干個分區，然後使用時間摘要法計算每個分區的統計量值。完成上述兩個方法以後，可以得到新的元特徵。結合新的元特徵和原本的靜態特徵得到新的資料。利用支持向量機作為分類器，並結合模擬退火和隨機森林兩種特徵選取的方法建立預測模型。

Radiofrequency ablation (RFA) is a common treatment for the hepatocellular carcinoma (HCC). Recurrence of HCC is an important issue despite effective treatments with tumor eradication. In this study, for those who had HCC and were treated by RFA, their clinical data are collected to build predictive models which can be used to predict the recurrence of HCC patients after RFA treatment. These clinical data are in the form of longitudinal data, which consists of static features and temporal features. We develop a data preprocessing method called Dynamic Period Slicing (DPS) combined with temporal abstraction (TA) to extract the high-level features from the original temporal features. We use DPS to divide a given time period into several partitions, then use quantitative TA to calculate the statistical measurements from the values in a given partition. After implementing the above two methods, we can obtain new meta-features, then combine the meta-features with original static features to form new processed data. Support Vector machine (SVM) was selected as the classifier to build predictive models with simulated annealing and random forest feature selection methods.