Application of Support Vector Machine on Liver Diseases Diagnosis

[[abstract]]預防勝於治療的觀念盛行，讓健康檢查在國人的心目中的重要性大幅提升。由於人口老化、環境污染、生活壓力，癌症年輕化如30、40歲罹癌時有所聞，大家對健康的需求自然也就比過去高。慢性肝病及肝硬化一直位居於國人十大死因重要病因，肝炎的可怕在大多數肝炎感染者都無症狀，必須經由驗血才能確定感染，導致大多數人對肝炎掉以輕心，而輕忽慢性肝炎的結果，造成一、二十年後發展成肝硬化或肝癌的憾事。 本研究以肝病作為研究標的，以國內某醫學中心健檢資料庫之歷史資料為樣本。主要使用支援向量機的RBF 核心函數來建立分類系統，採用RBF(Radial Basis Function , RBF)核心函數時，有兩個參數(C, )需要設定，使用Grid search 演算法，可以選擇較佳的SVM 模型參數，以增進SVM 分類器之準確率；亦使用決策樹模型建構分類器，並與支援向量機的模型進行比較。目的為建立出一套可推測是否罹患肝病之診斷分類器，扮演臨床決策支援系統的角色，進而達到快速輔助醫生對肝病預測診斷之功效。本實驗方法將其研究標的分成兩部份，實驗一主要針對診斷出肝功能異常患者進行預測實驗；肝功能異常患者中脂肪肝患者佔了五成以上，實驗二則進一步針對診斷出肝功能異常和脂肪肝患者分類預測實驗。 研究結果顯示，實驗一兩者所建立的分類系統皆有非常高的正確率，平均正確率皆有九成以上且相關重要屬性如SGPT、SGOT為肝功能異常屬性皆被篩選出，但支援向量機搭配屬性篩選出其他有關肝病之危險因子如：體重、BMI、血尿酸、高密度脂蛋白膽固醇均為脂肪肝的重要屬性；乳酸脫氫脢、血紅素為肝硬化的重要屬性；AFP為肝癌重要檢驗指標。實驗二中結果顯示SVM分類器整體平均正確率72.67%，脂肪肝患者分類上其平均正確率為69.61%，無肝病患者分類上其平均正確率為77.4%；C5.0決策樹分類器中亦可挑選出上述屬性，但整體平均正確率僅66.42%，脂肪肝患者分類上其平均正確率為63.52 %，無肝病患者分類上其平均正確率為68.2%。腹部超音波是診斷脂肪肝的利器，而運用SVM建立的模型其診斷預測績效符合並稍高於其檢驗統計數據，脂肪肝患者平均正確率為67 %，無肝病患者平均正確率為74%，以SVM搭配屬性篩選確實可以得到輔助診斷的功效。[[abstract]]Due to the concept of “prevention is better than cure” is widely spread the importance of health examination is becoming more significant The aging population polluted environment and the stress in life bring about the age drop in cancer Many people even get cancer in their thirties or forties The chronic hepatitis and cirrhosis is always on the top ten of the leading causes of death The chronic hepatitis is dangerous because the patients feel nothing and the disease can only be confirmed by blood testing Therefore many people disregards the symptoms of the disease and it develop into cirrhosis or liver cancer after ten to twenty years The goal of this paper is to analyze whether the techniques such as constructs a hybrid SVM-based strategy with feature selection and Decision Trees can act as the diagnosis classification and clinical decision support system to help the doctors in diagnosing liver diseases efficiently According to the different targets the experiment is divided into two parts The first experiment aims at the patients of impairment of liver function Of all the patients over 50% of them also suffer from fatty liver The second experiments aims at both the patients of impairment of liver function and fatty liver According to the result of the experiments the overall hit rate of the first experiment reaches 90% Important symptoms of liver disease such as SGPT and SGOT are successfully detected The SVM-based classifier detects that body weight BMI uric aid and HDL-C are the indicators of the fatty liver; LDH and Hb are the indictors of the cirrhosis; AFP is the indicator of liver cancer In the second experiment the overall hit rate using the SVM-based classifier is 72 67% The accuracy percentage of the classification of fatty liver patients is 69 61%; the percentage of patients without livers is 77 4% The C5 0 Selected above attributes the overall hit rate using the C5 0 Decision Trees is 66 42% The accuracy percentage of the classification of fatty liver patients is 63 52%; the percentage of patients without livers is 68 2% Fatty liver usually can be diagnosed by ultrasonography The forecast of the diagnosis result using SVM-based classifier is a little higher than the statistics of the current medical report: The accuracy percentage of fatty liver patients is 67%; the percentage of patients without livers is 74% In short conclusion using SVM-based classifier can help doctors to identify and diagnose the liver diseases accuratel