利用3D血管成像、血流模擬數據及病歷資料建立中風預測分類模型

Abstract

中風是日本死亡的主要疾病之一，而因為其通常發生突然且難以預防，導致中風所造成每年之醫療費用相當高。同時，日本老年人罹患中風的機率逐年上升，且日本老齡化人口趨勢增長。因此，預測和治療中風成為需要正視之問題。隨著近年的科技發展，許多新技術被運用於相關醫學研究，也致使兼具客觀及正確率的中風預測結果有機會可能達成。為能進一步解決此問題，本研究與日本一醫療機構合作，取得三種不同類型之資料集，包括立體血管成像、血流模擬數據及病歷資料，搭配近年發展之機器學習技術加以研究。此研究為能將立體血管成像與其他資料集合併使用，故針對立體光刻格式擷取曲率特徵作為使用；為改善數據不平衡之問題，本研究選擇合成少數類別技術 (Synthetic Minority Oversampling Technique, SMOTE)以創建新的破裂動脈瘤實例；且用決策樹演算法篩選特徵與支持向量機演算法分類數據，以建立一基於破裂動脈瘤之中風預測分類模型。最後，透過本研究模型成功合併三類資料集並將之導入分類模型，並獲得具傑出效能之結果 (準確率: 98.91%)。鑒於此結果，期許此中風預測分類模型未來可作為醫生在接受多類資訊時進行診斷中風之參考。

Stroke is one of the major causes of death in Japan, and it caused annually medical cost that is quite high. At the same time, the prevalence of stroke in the elderly population increased year after year and trends in Japan’s ageing population raised. Therefore, predicting and treating stroke become the problem that needs to be faced. This paper cooperated with one of the Japanese medical institutions to get three different types of data, including Three-Dimensional (3D) vascular imaging, simulations of blood flow and medical history data. This paper extracted curvature features from STL (STereoLithography) format of 3D vascular imaging, and used SMOTE (Synthetic Minority Oversampling Technique) to create new instances, then build a predicting classification model with Decision Tree and SVM (Support Vector Machine) algorithms. Through this model of this research, it successfully constructed the classification models with great performance. The model is expected to assist doctors in diagnosis while they accept different types of information.