利用邏輯斯迴歸模型進行特徵選擇延伸諾模圖以獲得預測計分

Abstract

近年來資訊系統的發展，各種資料也都儲存於電腦資料庫系統當中，而各種預測模型也因此蓬勃發展，其應用的範圍相當廣泛舉凡社會學、經濟學到臨床醫學皆廣為使用，以臨床醫學來說，能夠預先估計病患的死亡機率或者是特定疾病的復發機率對研究人員或者醫生是相當重要的。這些模型的建構過程中亦能識別出一些重要的資訊，將一些其他因素放入或許能從中發現一些意料之外被忽略卻重要的因素，因而對臨床的方向或是目標有所改變。而對於這些海量的資料若想從中提取一些潛在的有用信息，便需要一些特定的方法計算分析，因而發展了很多種不同的資料挖掘演算法，而在資料挖掘的過程中，為了提升機器學習的效能與資料挖掘結果的精準度，特徵變數選擇也越來越受到重視，進而產生了許多特徵變數選擇演算法以提升資料挖掘效能。對於大量資料且高維度的資料庫來說，若欲進行資料挖掘，特徵變數選擇是不可或缺的。有了好的特徵變數選擇能夠幫助研究人員去除不必要的資料，更針對有意義的資料作分析。 綜合以上之論點，本研究主要目的為利用諾模圖推導出病人結果的預測計分，而諾模圖為邏輯斯迴歸模型所延伸出的圖像視覺呈現，並對於迴歸模型進行特徵變數選擇以篩選出最佳變數子集合。研究材料以台大醫院神經重症電腦斷層及手術相關資料庫為資料來源依據，先應用概似比(likelihood ratio)分析各個特徵變數與目標變數的相關程度，以達到第一階段去除冗餘特徵變數的目的。接著在利用一致性指數特徵選擇(c-index feature selection)進一步的逐一去除相對不重要的特徵變數以提取最佳的一組特徵變數，在以邏輯斯迴歸所產生的諾模圖製作成預測模型，並證實其演算法篩選出的該組特徵變數所產生的預測模型的預測能力為一定程度上的精準，可提供臨床研究人員對於特徵變數更加注意其重要性。

Most information system had been developed for many years, which various kinds of data are stored in the database system of computer. As various kinds of predictive model are booming. The applications of predictive model in sociology, economics and clinical medicine are commonly used. In clinical medicine, to estimate the patient's risk of death or the probability of recurrence for particular disease is very important. The process of model construction may explore some unexpected factors which had been ignored. If the researchers want to retrieve some important information which had been hidden in mass data, particular methods for computing and analysis are need. For this purpose, several data mining algorithms are created. In the process of data mining, in order to enhancing generalization capability, speeding up learning process and improving model interpretability, feature selection becomes more and more in attention. For high dimensions database, feature selection is important way to reduce the redundant variables before processing the data mining. Taking above arguments, the main purpose of present research is to derive the predictive scoring function for patients’ outcome from nomogram which is a visualization tool extended from logistic regression model. We used the database of “CT finding and cranial surgery report” from NTUH as material, and apply several methods to obtain the feature subset : first, we use likelihood ratio to provide a summary of how many times more (or less) likely patients with the good outcome are to have that particular result than with poor outcome for each feature variables. After dropping the irrelevant features. Second, we apply the c-index selection to repeatedly calculate the c-index of the logistic regression model which drop a different feature variable a time, and select the subset with maximum c-index of the model. In this thesis, we also compare it with other common variable selection criterion of logistic regression model, and validate the predictive scoring by clinicians with expert knowledge.