论文题名(中文): | 息肉状脉络膜血管病变抗VEGF治疗反应的影像学及血液学生物标志物 |
姓名: | |
论文语种: | chi |
学位: | 博士 |
学位类型: | 专业学位 |
学校: | 北京协和医学院 |
院系: | |
专业: | |
指导教师姓名: | |
论文完成日期: | 2025-03-15 |
论文题名(外文): | Imaging and Hematological Biomarkers for Anti-VEGF Therapy Response in Polypoidal Choroidal Vasculopathy |
关键词(中文): | |
关键词(外文): | Polypoidal choroidal vasculopathy anti-VEGF classification imaging biomarkers hematological biomarkers |
论文文摘(中文): |
第一部分:息肉状脉络膜血管病变基于影像学标志物的抗VEGF疗效的列线图预测模型 研究目的:探索并识别预测PCV患者抗VEGF治疗反应的影像学生物标志物,建立列线图预测模型并进行多中心验证。 研究方法:这项前瞻性多中心研究纳入来自全国15家眼科中心的初治PCV患者。我院的患者被随机分为训练集和内部验证集。通过单因素回归分析、最小绝对收缩和选择算子算法(LASSO回归)和多因素回归分析建立列线图。其他14家中心的患者作为外部测试集。计算曲线下面积(AUC)、敏感性、特异性和准确性,并使用决策曲线分析(DCA)和临床影响曲线(CIC)评估模型在临床决策中的实际应用价值。 研究结果:训练集、内部验证集和外部测试集的患眼分别为66只、31只和71只。治疗反应良好组表现出较薄的黄斑中心凹下脉络膜厚度(SFCT)(230.67±61.96 μm vs. 314.42±88.00 μm,P<0.001)、较低的脉络膜血管指数(CVI)(0.31±0.08 vs. 0.36±0.05,P=0.006)、较少的脉络膜血管高通透性(CVH)(31.0% vs. 62.2%,P=0.012)以及更多的视网膜内液(IRF)(58.6% vs. 29.7%,P=0.018)。最终,SFCT(OR=0.990,95% CI 0.981-0.999,P=0.033)和CVI(OR=0.844, 95% CI 0.732-0.971,P=0.018)被选为最佳预测生物标志物,并以列线图形式呈现。该模型在训练集、内部验证集和外部测试集中预测“治疗反应良好”的AUC分别为0.837(95% CI 0.738-0.936)、0.891(95% CI 0.765-1.000)和0.901(95% CI 0.824-0.978),表现出较优的敏感性、特异性和实际应用价值。 研究结论:较薄的SFCT和较低的CVI可作为预测PCV患者对抗VEGF单药治疗反应良好的影像学生物标志物。基于这些标志物构建的列线图表现出令人满意的预测性能。 第二部分:息肉状脉络膜血管病变的机器学习模型:特征选择与可解释性研究 研究目的:构建并验证基于机器学习的预测模型,识别与PCV抗VEGF治疗疗效相关的关键特征,并进行模型可解释性分析。 研究方法:本研究采用多中心、前瞻性观察性研究设计,纳入716例PCV初治患者。通过Boruta算法和最小绝对收缩和选择算子算法(LASSO回归)筛选出与抗VEGF治疗反应相关的关键特征。构建多种机器学习模型,通过网格搜索优化超参数,通过多项指标评估模型性能,并采用沙普利加性解释(SHAP)分析方法进行模型解释。 研究结果:本研究共纳入716例PCV患者(716只眼),按照7:3的比例随机分为训练集和测试集。训练集中,反应良好和反应不佳的患者分别为193例和308例。反应不佳的患者具有以下显著特征:平均年龄较低(66.08±8.70岁,P=0.001)、脉络膜毛细血管层厚度较厚(25.06±10.15 μm,P=0.013)、中心凹下脉络膜厚度(SFCT)显著较厚(360.22±107.81 μm,P<0.001)、脉络膜肥厚血管发生率较高(80.2%,P<0.001)、大出血发生率较高(32.5%,P=0.003),且AMD样特征较少(20.5%,P<0.001)。通过Boruta算法和LASSO回归筛选出SFCT、脉络膜肥厚血管、年龄、大出血和AMD样特征为抗VEGF治疗反应的关键预测因子。基于这些特征,构建了多种机器学习模型,其中LightGBM模型表现最优,其AUC值为0.9345(95% CI 0.898-0.971),敏感性为0.9146,特异性为0.8889,精准率为0.8427,召回率为0.9146,F1分数为0.8772,平衡准确率为0.9018。SHAP分析进一步揭示了存在脉络膜肥厚血管、较厚SFCT和存在大出血对预测治疗反应不佳的正向贡献,而有AMD样特征和年龄较大则具有负向贡献。 研究结论:本研究构建了预测PCV抗VEGF疗效的机器学习模型,识别了疗效的关键预测特征。LightGBM模型结合SHAP分析方法,显著提高了模型的预测性能和可解释性。 第三部分:肥厚型息肉状脉络膜血管病变的分型标准及血液生物标志物 研究目的:提出基于预后的肥厚型PCV(Pachy PCV)分型标准,揭示其特异性蛋白和代谢物表达谱,探索潜在的生物标志物及发病机制。 研究方法:本研究为一项前瞻性、多中心临床研究。基于Logistic回归分析,提出PCV分型标准,将PCV分为Pachy PCV和Nonpachy PCV。此外,收集40例Pachy PCV患者、31例Nonpachy PCV患者及35名健康对照者的血浆样本,进行蛋白质组学和代谢组学分析。采用液相色谱-串联质谱(LC-MS/MS)和非靶向代谢组学技术检测差异蛋白和代谢物,并通过GO和KEGG数据库进行功能富集分析,探索Pachy PCV的分子机制。 研究结果:共纳入716 例初治PCV患者(716只眼),与治疗反应良好组相比,治疗反应不佳组表现出显著的肥厚脉络膜特征。多因素Logistic回归显示,中心凹下脉络膜厚度(SFCT)(OR=0.995,95% CI 0.992-0.999,P=0.005)、大出血(OR=0.455,95% CI 0.277-0.745,P=0.002)和脉络膜肥厚血管的存在(OR=0.125,95% CI 0.067-0.232,P<0.001)是预测抗VEGF治疗反应的独立危险因素。根据约登指数确定SFCT的最佳预测阈值为295.75 μm。据此提出Pachy PCV的诊断标准包括:(1)主要标准:同时满足SFCT≥300 μm和存在脉络膜肥厚血管;(2)若仅满足一项主要标准,则需同时满足年龄≤66岁和无AMD样特征两项次要标准。蛋白质组学分析鉴定出135种Pachy PCV患者与Nonpachy PCV之间的差异蛋白,其中97种在Pachy PCV中显著上调,代谢组学分析揭示了Pachy PCV患者中168种差异代谢物,其中23种显著上调。GO及KEGG通路富集显示Pachy PCV中VEGF、MAPK、Rap1和PI3K-Akt等信号通路显著富集,并揭示了两个独特的关键通路:剪切力和动脉粥样硬化通路及HIF-1信号通路。参与剪切力反应和缺氧反应的相关蛋白(如NCF1、PIK3CB、CAMK2G、EIF4E2)和代谢物(如丙酮酸、Fe2+)在Pachy PCV中显著上调,提示剪切力和缺氧微环境在疾病进展中的重要作用。 研究结论:Pachy PCV和Nonpachy PCV具备不同的影像学特征及对抗VEGF反应,血浆蛋白组学及代谢组学揭示了Pachy PCV独特的发病机制,本研究为PCV临床治疗策略的优化提供了重要依据。
|
论文文摘(外文): |
Part I: Nomogram Prediction Model for Anti-VEGF Efficacy in Polypoidal Choroidal Vasculopathy Based on Imaging Biomarkers Objective: To explore and identify imaging biomarkers that predict the response to anti-VEGF therapy in patients with polypoidal choroidal vasculopathy (PCV), and to develop and validate a nomogram prediction model through a multicenter study. Methods: This prospective multicenter study included treatment-naïve PCV patients from 15 ophthalmic centers nationwide. Patients from our hospital were randomly divided into a training set and an internal validation set. Univariate regression analysis, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and multivariate regression analysis were used to construct the nomogram. Patients from the other 14 centers served as the external test set. The area under the curve (AUC), sensitivity, specificity, and accuracy were calculated. Decision curve analysis (DCA) and clinical impact curves (CIC) were used to evaluate the model’s practical application value in clinical decision-making. Results: The training set, internal validation set, and external test set included 66, 31, and 71 eyes, respectively. The good responder exhibited thinner subfoveal choroidal thickness (SFCT) (230.67±61.96 μm vs. 314.42±88.00 μm, P<0.001), lower choroidal vascularity index (CVI) (0.31±0.08 vs. 0.36±0.05, P=0.006), less choroidal vascular hyperpermeability (CVH) (31.0% vs. 62.2%, P=0.012), and more intraretinal fluid (IRF) (58.6% vs. 29.7%, P=0.018). Ultimately, SFCT (OR 0.990, 95% CI 0.981-0.999, P=0.033) and CVI (OR 0.844, 95% CI 0.732-0.971, P=0.018) were selected as the optimal predictive biomarkers and presented in the form of a nomogram. The model demonstrated excellent sensitivity, specificity, and practical application value, with AUCs for predicting good responder of 0.837 (95% CI 0.738-0.936), 0.891 (95% CI 0.765-1.000), and 0.901 (95% CI 0.824-0.978) in the training set, internal validation set, and external test set, respectively. Conclusion: Thinner SFCT and lower CVI could serve as imaging biomarkers to predict a favorable response to anti-VEGF monotherapy in PCV patients. The nomogram constructed based on these biomarkers exhibited satisfactory predictive performance.
Part II: Machine Learning Models for Polypoidal Choroidal Vasculopathy: Feature Selection and Interpretability Study Objective: To construct and validate machine learning (ML)-based predictive models for identifying key features associated with anti-VEGF treatment efficacy in polypoidal choroidal vasculopathy (PCV) and to perform interpretability analysis of the models. Methods: This study adopted a multicenter, prospective, observational design, enrolling 716 treatment-naïve PCV patients. Boruta algorithm and Least Absolute Shrinkage and Selection Operator (LASSO) regression were performed to identify key features related to anti-VEGF treatment response. Multiple ML models were constructed. Hyperparameters were optimized through grid search, and model performance was evaluated using multiple metrics. Model interpretability was analyzed using SHapley Additive exPlanations (SHAP). Results: A total of 716 PCV patients (716 eyes) were included and randomly divided into a training set and a test set in a 7:3 ratio. In the training set, 193 patients showed a good response, while 308 patients had a poor response. Patients with a poor response exhibited the following significant characteristics: younger average age (66.08±8.70 years, P=0.001), thicker choroidal capillary layer (25.06±10.15 μm, P=0.013), significantly thicker subfoveal choroidal thickness (SFCT) (360.22±107.81 μm, P<0.001), higher prevalence of choroidal pachyvessels (80.2%, P<0.001), higher incidence of extensive hemorrhage (32.5%, P=0.003), and fewer AMD-like features (20.5%, P<0.001). Through Boruta and LASSO regression, SFCT, choroidal pachyvessels, age, extensive hemorrhage, and AMD-like features were identified as key predictors of anti-VEGF treatment response. Based on these features, multiple ML models were constructed, with the LightGBM model performing the best, achieving an AUC of 0.9345 (95% CI 0.898-0.971), sensitivity of 0.9146, specificity of 0.8889, precision of 0.8427, recall of 0.9146, F1 score of 0.8772, and balanced accuracy of 0.9018. SHAP analysis further revealed that the presence of choroidal pachyvessels, thicker SFCT, and extensive hemorrhage positively contributed to predicting a poor treatment response, while AMD-like features and older age had negative contributions. Conclusion: This study constructed ML models to predict anti-VEGF efficacy in PCV, identifying key predictive features. The LightGBM combined with SHAP analysis significantly improved the predictive performance and interpretability of the model.
Part III: Classification Criteria and Blood Biomarkers for Pachychoroid Polypoidal Choroidal Vasculopathy Objective: To propose a prognosis-based classification for pachychoroid PCV (Pachy PCV), reveal its specific protein and metabolite expression profiles, and explore potential biomarkers and pathogenic mechanisms. Methods: This was a prospective, multicenter clinical study. A PCV classification standard was proposed using logistic regression analysis, categorizing PCV into Pachy PCV and Nonpachy PCV. Additionally, plasma samples from 40 Pachy PCV patients, 31 Nonpachy PCV patients, and 35 healthy controls were collected for proteomic and metabolomic analyses. Liquid chromatography-tandem mass spectrometry (LC-MS/MS) and untargeted metabolomics were used to detect differential proteins and metabolites. Functional enrichment analysis was performed using GO and KEGG databases to explore the molecular mechanisms of Pachy PCV. Results: A total of 716 treatment-naïve PCV patients (716 eyes) were included. Compared to good responder group, the poor responder group exhibited significant pachychoroid features. Multivariate logistic regression analysis revealed that subfoveal choroidal thickness (SFCT) (OR=0.995, 95% CI 0.992-0.999, P=0.005), extensive hemorrhage (OR=0.455, 95% CI 0.277-0.745, P=0.002) and the presence of choroidal pachyvessels (OR=0.125, 95% CI 0.067-0.232, P<0.001) were independent risk factors for predicting response to anti-VEGF. The optimal predictive threshold for SFCT, determined by the Youden index, was 295.75 μm. Based on these findings, the diagnostic criteria for Pachy PCV included: (1) Major criteria: simultaneous presence of SFCT ≥ 300 μm and choroidal pachyvessels; (2) If only one major criterion was met, two minor criteria should also be satisfied: age ≤ 66 years and absence of AMD-like features. Proteomic analysis identified 135 differentially expressed proteins between Pachy PCV and Nonpachy PCV, with 97 significantly upregulated in Pachy PCV. Metabolomic analysis revealed 168 differential metabolites in Pachy PCV, with 23 significantly upregulated. GO and KEGG pathway enrichment analysis showed significant enrichment of VEGF, MAPK, Rap1, and PI3K-Akt signaling pathways in Pachy PCV, along with two unique key pathways: fluid shear stress and atherosclerosis pathway, and HIF-1 signaling pathway. Proteins (e.g., NCF1, PIK3CB, CAMK2G, EIF4E2) and metabolites (e.g., pyruvate, Fe2+) involved in fluid shear stress response and hypoxia response were significantly upregulated in Pachy PCV, suggesting the critical role of fluid shear stress and hypoxic microenvironments in disease progression. Conclusion: Pachy PCV and Nonpachy PCV exhibited distinct imaging features and responses to anti-VEGF therapy. Plasma proteomics and metabolomics revealed unique pathogenic mechanisms in Pachy PCV, providing important insights for optimizing clinical treatment strategies for PCV.
|
开放日期: | 2025-06-05 |