常规检验标志物联合机器学习的青光眼风险预测模型的构建与评估

Development and Evaluation of a Glaucoma Risk Prediction Model Using Routine Laboratory Biomarkers and Machine Learning

摘要: 目的：基于常规检验标志物联合机器学习，构建并在独立验证集中评价一个可解释、可部署的青光眼个体化风险预测模型，综合考察其判别力、校准度与决策净获益，旨在为基层医疗提供低成本、易推广的早筛分诊工具。方法：纳入自贡市第一人民医院2022年3月~2025年3月就诊人群共1200例，按7:3分为训练集(n = 840)与验证集(n = 360)，两集青光眼/对照比例近1:1。以单因素分析与LASSO正则化联合筛选特征，确定6个关键变量；构建多种模型并以Logistic作为主模型。主要评价指标包括AUC/ROC、Brier分数与校准曲线、决策曲线分析；同时报告阈值相关指标(准确率、敏感度、特异度、F1、Kappa、Youden指数、PPV、NPV)。结果：验证集中，主模型(Logistic)取得较优的综合性能(AUC约0.78~0.86；Brier约0.17)，校准曲线贴近理想线；在DCA的0.1~0.6阈值区间获得稳定且高于Treat-all/none的净获益。SVM与XGBoost的区分度接近，但校准与净获益整体略逊。结论：仅依托常规检验数据即可获得具有良好区分度、可靠概率刻度与实用净获益的青光眼风险模型。该方案可及、低成本、标准化，适合在基层与体检场景作为“先分层、后影像/专科”的补充路径，值得在多中心前瞻性研究中进一步验证与再校准。

Abstract: Objective: To develop an interpretable and deployable individualized glaucoma risk prediction model based on routine laboratory biomarkers combined with machine learning, and to evaluate its discrimination, calibration, and clinical net benefit in an independent validation cohort, aiming to provide a low-cost and scalable early triage tool for primary care. Methods: A total of 1200 individuals presenting to Zigong First People’s Hospital from March 2022 to March 2025 were included and randomly split 7:3 into a training set (n = 840) and a validation set (n = 360), with an approximately 1:1 glaucoma/control ratio in both sets. Features were selected by combining univariable analysis with LASSO regularization to determine six key variables. Multiple algorithms were trained and evaluated, with logistic regression prespecified as the primary model. Primary endpoints were AUC/ROC, Brier score with calibration curves, and decision curve analysis (DCA); threshold-based metrics (accuracy, sensitivity, specificity, F1 score, Cohen’s κ, Youden’s index, PPV, NPV) were also reported. Results: In the validation set, the primary model (logistic regression) showed favorable overall performance (AUC ~0.78~0.86; Brier ~0.17) with calibration close to the ideal line; across the 0.1~0.6 threshold range, DCA indicated consistent net benefit exceeding treat-all and treat-none strategies. SVM and XGBoost achieved comparable discrimination but exhibited slightly inferior calibration and net benefit. Conclusions: A glaucoma risk model can be derived solely from routine laboratory data, achieving good discrimination, reliable probability scaling, and practical clinical net benefit. This standardized, low-cost approach is well suited for primary-care and health-check settings as a “pre-stratification before imaging/specialist referral” pathway and merits further multicenter prospective validation and recalibration.