多模态机器学习增强早期圆锥角膜检测:整合断层扫描、生物力学与临床危险因素
这项前瞻性多中心横断面研究纳入5个中心1035只眼,构建整合Scheimpflug断层扫描、角膜生物力学与临床危险因素的多模态机器学习诊断系统。特征选择后的CatBoost模型在检测顿挫型圆锥角膜时验证集AUROC达0.975,内部测试与外部测试均优于断层扫描与生物力学指数;对亚临床和临床圆锥角膜外部测试AUROC分别为0.991和1.000。作者认为该方法可能改善早期圆锥角膜检测及屈光手术候选者的筛查。
要点
- 特征选择后CatBoost模型检测顿挫型圆锥角膜验证集AUROC=0.975,优于仅临床危险因素(0.856)和仅设备参数(0.885)。
- 该模型内部测试AUROC=0.976优于断层扫描与生物力学指数的0.813(P=.048),外部测试0.952优于0.847(P=.012)。
- 外部测试中亚临床与临床圆锥角膜检测AUROC分别为0.991和1.000。
核心结论
- 整合临床危险因素、断层扫描与生物力学的多模态机器学习系统改善了早期圆锥角膜检测,尤其是顿挫型圆锥角膜。
- 作者认为该方法可能增强临床决策和屈光手术候选者的筛查。
研究设计与样本
- 前瞻性、多中心、横断面研究,纳入5个中心2020年至2024年诊断为圆锥角膜的患者及在术前屈光手术门诊评估的个体。
- 共纳入1035名参与者的1035只眼:正常对照590只眼、顿挫型圆锥角膜157只眼、亚临床圆锥角膜143只眼、临床圆锥角膜145只眼。
- 收集人口学、生活方式和临床眼科数据,包括Pentacam和Corvis ST测量,数据集分为训练集、验证集、内部测试集和外部测试集。
- 采用最小绝对收缩和选择算子回归识别预测变量,使用4种特征集(临床危险因素、设备衍生参数、组合特征、选择后特征)训练6种机器学习模型。
- 主要结局指标为受试者工作特征曲线下面积(AUROC)。
主要发现
- 检测顿挫型圆锥角膜时,特征选择后CatBoost模型在验证集准确度最高(AUROC=0.975),优于组合特征模型(AUROC=0.963)、仅临床危险因素模型(AUROC=0.856)和仅设备参数模型(AUROC=0.885)。
- 该模型在内部测试中优于断层扫描与生物力学指数(AUROC=0.976 vs 0.813;P=.048),在外部测试中亦优于该指数(AUROC=0.952 vs 0.847;P=.012)。
- 外部测试中亚临床圆锥角膜检测AUROC=0.991,临床圆锥角膜检测AUROC=1.000。
关键数值与推荐
- 顿挫型圆锥角膜检测:特征选择后CatBoost模型验证集AUROC=0.975,组合特征模型AUROC=0.963,仅临床危险因素模型AUROC=0.856,仅设备参数模型AUROC=0.885。
- 内部测试:特征选择后CatBoost模型AUROC=0.976 vs 断层扫描与生物力学指数0.813(P=.048)。
- 外部测试:特征选择后CatBoost模型AUROC=0.952 vs 断层扫描与生物力学指数0.847(P=.012)。
- 外部测试:亚临床圆锥角膜AUROC=0.991,临床圆锥角膜AUROC=1.000。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To develop and validate a machine learning (ML) diagnostic system that integrates Scheimpflug tomography, corneal biomechanics, and clinical risk factors (CRF) to enhance the early detection of keratoconus (KC). DESIGN: Prospective, multicenter, cross-sectional study. PARTICIPANTS: Patients diagnosed with KC and individuals evaluated in preoperative refractive surgery clinics. METHODS: Demographic, lifestyle, and clinical ophthalmic data, including Pentacam and Corvis ST measurements, were collected from patients with KC and refractive surgery candidates across 5 centers between 2020 and 2024. The dataset was divided into training, validation, internal test, and external test sets. Least absolute shrinkage and selection operator regression was used to identify predictive variables. Six ML models were trained using 4 feature sets: CRF, device-derived parameters, combined features, and selected features. MAIN OUTCOME MEASURES: Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). RESULTS: The dataset included 1035 eyes from 1035 participants across 5 centers: 590 normal controls, 157 eyes with forme fruste keratoconus (FFKC), 143 with subclinical KC, and 145 with clinical KC. For FFKC detection, the post-feature selection CatBoost model achieved the highest accuracy (AUROC = 0.975), outperforming the combined-feature (AUROC = 0.963), CRF-only (AUROC = 0.856), and device-only models (AUROC = 0.885) in the validation set. This model also outperformed the tomographic and biomechanical index in internal (AUROC = 0.976 vs 0.813; P = .048) and external testing (AUROC = 0.952 vs 0.847; P = .012). For subclinical and clinical KC, external testing yielded near-perfect performance (AUROC = 0.991 and 1.000, respectively). CONCLUSIONS: A multimodal ML system integrating CRF, tomography, and biomechanics improved early KC detection, particularly for FFKC. This approach may enhance clinical decision-making and screening for refractive surgery candidates.
引用
- Du K, Peng R, Chen Y, et al. Enhancing Early Keratoconus Detection With Multimodal Machine Learning: Integrating Tomography, Biomechanics, and Clinical Risk Factors. Am J Ophthalmol. 2025. PMID: 40840580. doi:10.1016/j.ajo.2025.08.035
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。