Prathikanti Vinayak, Casentini Renee, Hwang Jonathan, Abdou Amza, Beidou Nassirou, Kadri Boubacar, Srivathsan Ariktha, Prieto Isabelle, Huang Winnie, Eyassu Daniel G, Gebreegziabher Elisabeth, Pierce Corinne, Burroughs Hadley, Keenan Jeremy D, Lietman Thomas M
F. I. Proctor Foundation, Department of Ophthalmology, University of California, San Francisco, California.
Programme National de Lutte Contre la Cecité, Niamey, Niger.
Am J Trop Med Hyg. 2025 Feb 25;112(5):1087-1090. doi: 10.4269/ajtmh.24-0321. Print 2025 May 7.
The WHO has a simplified grading system for assessing trachoma. However, even for experts, it can be difficult to classify certain cases as strictly positive or negative for a given grade. Given the absence of a true gold standard, we performed a Latent Class Analysis (LCA) on a set of 200 graded photos of the superior tarsal conjunctiva. Ten trained graders assessed the presence of two trachoma grades: trachomatous inflammation-follicular (TF) and trachomatous inflammation-intense (TI). The LCA was modeled in two different ways: first with two classes (presence/absence), and then with three classes, with the extra class presumed to represent a more discrepant "borderline" case. Cohen's κ-statistics measuring agreement between graders were calculated for TF and TI grades (separately) before and after removing the third latent class. The κ-statistic increased by 0.10 (95% CI = 0.72-0.85; P <0.001) for TF and 0.13 (95% CI = 0.81-0.90; P <0.001) for TI, indicating that the third latent class represented a discrepant-case borderline class. The identification of borderline grading cases using a three-class LCA may be useful in creating balanced grader certification examinations that represent the full spectrum of disease. Additionally, a multiclass LCA could act as a probabilistic gold standard used to train and analyze future convolutional neural network models.
世界卫生组织（WHO）有一个用于评估沙眼的简化分级系统。然而，即使对于专家来说，将某些病例严格分类为给定等级的阳性或阴性也可能很困难。鉴于缺乏真正的金标准，我们对一组200张上睑结膜的分级照片进行了潜在类别分析（LCA）。十位经过培训的分级人员评估了两种沙眼等级的存在情况：沙眼性炎症-滤泡型（TF）和沙眼性炎症-重度（TI）。LCA以两种不同方式建模：首先是两类（存在/不存在），然后是三类，额外的一类被假定代表更具差异的“临界”病例。在去除第三个潜在类别之前和之后，分别计算了衡量分级人员之间一致性的科恩κ统计量用于TF和TI等级。TF的κ统计量增加了0.10（95%置信区间 = 0.72 - 0.85；P <0.001），TI的κ统计量增加了0.13（95%置信区间 = 0.81 - 0.90；P <0.001），这表明第三个潜在类别代表了一个差异病例的临界类别。使用三类LCA识别临界分级病例可能有助于创建代表疾病全谱的平衡分级人员认证考试。此外，多类LCA可以作为一种概率金标准，用于训练和分析未来的卷积神经网络模型。
