To investigate machine-learning (ML) algorithms to differentiate corneal biomechanical properties between different topographical stages of keratoconus (KC) by dynamic Scheimpflug tonometry (CST, Corvis ST, Oculus, Wetzlar, Germany). In the following, ML models were used to predict the severity in a training and validation dataset. Three hundred and eighteen keratoconic and one hundred sixteen healthy eyes were included in this monocentric and cross-sectional pilot study. Dynamic corneal response (DCR) and corneal thickness related (pachymetric) parameters from CST were chosen by appropriated selection techniques to develop a ML algorithm. The stage of KC was determined by the topographical keratoconus classification system (TKC, Pentacam, Oculus). Patients who were classified as TKC 1, TKC 2 and TKC 3 were assigned to subgroup mild, moderate, and advanced KC. If patients were classified as TKC 1–2, TKC 2–3 or TKC 3–4, they were assigned to subgroups according to the normative range of further corneal indices (index of surface variance, keratoconus index and minimum radius). Patients classified as TKC 4 were not included in this study due to the limited amount of cases. Linear discriminant analysis (LDA) and random forest (RF) algorithms were used to develop the classification models. Data were divided into training (70% of cases) and validation (30% of cases) datasets. LDA model predicted healthy, mild, moderate, and advanced KC eyes with a sensitivity (Sn)/specificity (Sp) of 82%/97%, 73%/81%, 62%/83% and 68%/95% from a validation dataset, respectively. For the RF model, a Sn/Sp of 91%/94%, 80%/90%, 63%/87%, 72%/95% could be reached for predicting healthy, mild, moderate, and advanced KC eyes, respectively. The overall accuracy of LDA and RF was 71% and 78%, respectively. The accuracy for KC detection including all subgroups of KC severity was 93% in both models. The RF model showed good accuracy in predicting healthy eyes and various stages of KC. The accuracy was superior with respect to the LDA model. The clinical importance of the models is that the standalone dynamic Scheimpflug tonometry is able to predict the severity of KC without having the keratometric data. NCT04251143 at Clinicaltrials.gov, registered at 12 March 2018 (Retrospectively registered).

中文翻译：

---

使用人工智能预测圆锥角膜严重程度的基于角膜生物力学特性的分类系统的开发

研究机器学习 (ML) 算法，通过动态 Scheimpflug 眼压测量法（CST、Corvis ST、Oculus、Wetzlar、德国）区分圆锥角膜 (KC) 不同地形阶段之间的角膜生物力学特性。在下文中，ML 模型用于预测训练和验证数据集中的严重性。318 只圆锥角膜和 116 只健康眼睛被纳入这项单中心和横断面初步研究。通过适当的选择技术选择来自 CST 的动态角膜反应 (DCR) 和角膜厚度相关（测厚）参数以开发 ML 算法。KC 的阶段由地形圆锥角膜分类系统（TKC、Pentacam、Oculus）确定。被归类为 TKC 1、TKC 2 和 TKC 3 的患者被分配到轻度、中度、和高级 KC。如果患者被分类为 TKC 1-2、TKC 2-3 或 TKC 3-4，则根据其他角膜指数（表面方差指数、圆锥角膜指数和最小半径）的规范范围将他们分配到亚组。由于病例数有限，归类为 TKC 4 的患者未包括在本研究中。使用线性判别分析 (LDA) 和随机森林 (RF) 算法来开发分类模型。数据分为训练（70% 的案例）和验证（30% 的案例）数据集。LDA 模型以 82%/97%、73%/81%、62%/83% 和 68%/95% 的灵敏度 (Sn)/特异性 (Sp) 预测健康、轻度、中度和晚期 KC 眼验证数据集，分别。对于 RF 模型，可以达到 91%/94%、80%/90%、63%/87%、72%/95% 的 Sn/Sp 来预测健康、轻度、中度、和高级 KC 眼睛，分别。LDA 和 RF 的总体准确度分别为 71% 和 78%。在两种模型中，包括 KC 严重程度的所有亚组在内的 KC 检测准确度均为 93%。RF 模型在预测健康眼睛和 KC 的各个阶段方面表现出良好的准确性。相对于 LDA 模型，精度更高。这些模型的临床重要性在于，独立的动态 Scheimpflug 眼压计能够在没有角膜曲率数据的情况下预测 KC 的严重程度。Clinicaltrials.gov 上的 NCT04251143，于 2018 年 3 月 12 日注册（追溯注册）。RF 模型在预测健康眼睛和 KC 的各个阶段方面表现出良好的准确性。相对于 LDA 模型，精度更高。这些模型的临床重要性在于，独立的动态 Scheimpflug 眼压计能够在没有角膜曲率数据的情况下预测 KC 的严重程度。Clinicaltrials.gov 上的 NCT04251143，于 2018 年 3 月 12 日注册（追溯注册）。RF 模型在预测健康眼睛和 KC 的各个阶段方面表现出良好的准确性。相对于 LDA 模型，精度更高。这些模型的临床重要性在于，独立的动态 Scheimpflug 眼压计能够在没有角膜曲率数据的情况下预测 KC 的严重程度。Clinicaltrials.gov 上的 NCT04251143，于 2018 年 3 月 12 日注册（追溯注册）。