| 摘要: | 摘要
背景與目的：依據Cirrus Optical Coherence Tomography (OCT)儀器所測出的17項視神經纖維層厚度參數個別診斷能力，以及，比較邏吉斯迴歸及支撐向量機法並且改善區別台灣地區正常人與青光眼患者的準確度，輔助建立一套適合台灣地區的青光眼臨床決策支援系統，以達輔助醫師臨床診斷的目的。
研究方法：2009年11月至2010年5月前往中部某醫學中心眼科部執行完整的眼科檢查為研究對象，隨機挑選50名患有青光眼患者的其中一隻眼睛以及性別、年齡、眼睛位置、中央角膜厚度和屈光度匹配的50名正常人。使用SPSS12.0執行描述性統計、卡方檢定、t檢定、皮爾森相關、邏吉斯迴歸統計分析向前wald法、繪製ROC曲線並計算曲線下面積；使用Libsvm3.0執行基因演算法和支撐向量機法。
研究結果：Cirrus OCT的單一參數，最高AUC為平均厚度和Hr6(AUC=0.836)。邏吉斯迴歸分析，最好的每小時時區模式，挑選出Hr6(P <0.001)、Hr9( P=0.002)及屈光度(P=0.006)之三項參數，分類準確度為82%，AUC為0.898。支撐向量機法，最好模式為四象限，挑選出年齡、上側、下側及屈光度之四項參數，得到分類準確度為87%，AUC達0.914。
結論與建議：Cirrus OCT的單一參數有很好的性能之區別台灣民眾青光眼和正常人，而邏吉斯迴歸分析與支撐向量機法皆能改善Cirrus OCT的分類性能。未來能執行青光眼決策支援系統以達輔助醫師判斷青光眼的準確度。
Abstract
Purpose: The present study sought to assist in establishing a Clinical Decision Support System (CDSS) for glaucoma diagnosis, by comparing the classification performance of Logistic Regression Analysis (LRA) and Support Vector Machine (SVM) based on 17 parameters of Retinal Nerve Fiber Layer Thickness (RNFLT) measured by Cirrus Optical Coherence Tomography (Cirrus OCT).
Methods: 50 glaucoma patients receiving ophthalmological outpatient services from November 2009 to May 2010 in a medical center of central Taiwan were randomly selected as the first half of the study sample. Additionally, 50 individuals free of glaucoma whose age, gender, eye location, central corneal thickness, and refraction matched those of the 50 glaucoma patients were chosen as the comparison group. Statistic analyses including descriptive statistics, Chi-square, t-test, Pearson correlation, Wald method of LRA, and calculating Receiver Operating Characteristic (ROC) curve and the area under ROC curve (AUC) were carried out in SPSS 12. Genetic Algorithm (GA) and SVM were conducted using Libsvm 3.0.
Results: The average RNFLT and Hr6 were the best individual parameters of Cirrus OCT for differentiating between normal and glaucomatous eyes (AUC 0.836). The best parameters in the model of O’clock hour segment thickness generated by the forward Wald method of LRA were Hr6, Hr9, and refraction (P<0.001, P=0.002, and P=0.006, respectively), reaching 0.898 of AUC and 82% of accuracy. The best model of quadrants that included age, superior, inferior, and refaction generated by SVM achieved 0.914 of AUC and 87% of accuracy.
Conclusion: The performance of individual parameters obtained from Cirrus OCT is substantially reliable for differentiating glaucomatous eyes from normal eyes. Both LRA and SVM can effectively improve the classification performance in Cirrus OCT, advancing the accuracy in the implementation of CDSS for glaucoma diagnosis among ophthalmologists in the near future. |
