Objective: To investigate the diagnosis of normal cornea, subclinical keratoconus and keratoconus by artifical intelligence. Methods: Diagnostic study. From January 2016 to January 2019, who admitted to Tianjin Eye Hospital from 18 to 48 years old, with an average of (28.4±8.2) years of myopia patients in 2 018 cases. Two experienced ophthalmologists labeled keratoconus, subclinical keratconus and nomal cornea based on the topography. The data of 80% (1 615 cases) patients were randomly selected as the training set by computer random sampling method, and the data of 20% (403 cases) patients were used as the verification set. Using the Gradient Boosting Decision Tree (GBDT) algorithm to extract 28 corneal parameters, and establish an algorithm model to diagnose the corneal condition of the patient, verify the diagnostic accuracy of the model by using the 10-fold cross-validation method, and evaluate the model using the receiver operating characteristic curve. Sensitivity and specificity with the original labeling and ophthalmic resident labeling. Results: The diagnostic accuracy of the model was 95.53%. The area under the receiver operating characteristic curve (AUC) of the validation set was 0.996 6. The accuracy of the model for diagnosis of subclinical keratoconus and normal cornea was 96.67%, the AUC of the validation set was 0.993 6; the accuracy of diagnosis of keratoconus and normal cornea was 98.91%, and the AUC of the validation set was 0.998 2. The diagnostic accuracy of the model is 95.53%, which is significantly better than the resident's 93.55%. Conclusion: The model established by artifical intelligence can diagnose the subclinical keratoconus with high accuracy, which can greatly improve the clinical diagnosis efficiency and accuracy of young and primary ophthalmologists. (Chin J Ophthalmol, 2019, 55: 911-915).
目的： 探讨基于机器学习的数据模型对于正常角膜、亚临床圆锥角膜和圆锥角膜的诊断情况。 方法： 诊断性研究。收集2016年1月至2019年1月就诊于天津市眼科医院年龄（28.4±8.2）岁的近视眼患者2 018例。由2名经验丰富的眼科专家根据角膜地形图诊断并标注为圆锥角膜、亚临床圆锥角膜和正常角膜。采用计算机随机采样方法随机选取其中80%（1 615例）患者的数据作为训练集，另20%（403例）患者的数据作验证集。借助梯度提升树（GBDT）算法提取28个角膜参数特征，建立数据模型对患者角膜情况进行诊断，采用十折交叉验证法验证模型的诊断准确率，并采用受试者工作特征曲线评价数据模型与标注情况及高年资住院医师标注情况的敏感度与特异度。 结果： 模型诊断准确率为95.53%。验证集受试者工作特征曲线下面积（AUC）为0.9966。模型诊断亚临床圆锥角膜与正常角膜的准确率为96.67%，验证集AUC为0.9936；诊断圆锥角膜与正常角膜的准确率为98.91%，验证集AUC为0.998 2。模型的诊断准确率为95.53%，明显优于高年资住院医师（93.55%）。 结论： 借助机器学习方法建立的数据模型诊断亚临床期圆锥角膜有较高的准确率，可极大提升年轻和基层医师的临床诊断效率和准确率。（中华眼科杂志，2019，55：911-915）.
Keywords: Artificial intelligence; Diagnosis, computer-assisted; Early diagnosis; Keratoconus.