摘要

目的

孕早期抑郁症状对孕妇身心健康及胎儿生长发育构成威胁，可能导致流产、早产等多种妊娠不良结局，早期识别和干预对保障母婴健康至关重要，但传统方法如爱丁堡量抑郁量表问卷依赖主观报告，尚无明确的生理生化指标。基于常规临床风险因素的模型评估性能有限，本研究旨在探索视网膜图像自动分析技术在孕早期抑郁症状风险评估中的潜力，通过构建客观、非侵入性的风险模型，比较孕早期抑郁症状患者与非抑郁孕妇视网膜特征差异，揭示微血管损伤特征，评估视网膜特征的诊断价值并识别关键指标与抑郁风险相关性，进而提出视网膜图像技术评估方案，优化个体化风险评估，探索视网膜图像自动分析技术在孕早期抑郁症状风险评估中的应用潜力，通过构建客观风险评估模型，为临床实践提供新型辅助工具，为中国孕产妇心理健康管理提供科学依据。

方法

本研究采用病例对照研究，收集2023年5月至2024年5月期间在通州区妇幼保健院孕早期检查的1052名孕妇数据。通过爱丁堡抑郁量表测量孕早期抑郁症状情况，并同时通过问卷调查收集其他基线信息，在医院信息系统收集生化检验结果，使用佳能眼底照相机CR-2AF拍摄双眼视网膜图像。采用“自动视网膜图像分析方法（Automatic Retinal Image Analysis, ARIA）”技术计算常见的视网膜图像参数并采用ResNet50机器学习算法方法提取与孕早期抑郁症状相关的视网膜特征高维参数，并使用主成分分析对其进行降维。采用年龄±2岁进行匹配分为抑郁组症状与对照组，以量表诊断结果为结局变量，临床特征和视网膜特征为自变量，通过lasso回归筛选特征变量，基于Logistic回归模型进行建模，采用十折交叉验证法评估模型性能。构建仅临床特征、仅视网膜特征及综合模型。模型内部评采用灵敏度、特异度、AUC衡量区分度，通过Hosmer-Lemeshow（HL）检验评估校准度，并使用净重新分类改善指数（NRI）和综合判别改善指数（IDI）分析特征联合的增益效应，通过DeLong检验比较模型间性能差异，为临床应用提供依据。

结果

本研究最终共纳入449名孕早期孕妇，其中抑郁症状组92例，非抑郁症状组357例。仅临床特征模型的AUC为0.81（95% CI: 0.76-0.86），灵敏度为72.42%（95% CI: 61.87-79.63%），特异度82.73%（95% CI: 69.54-88.37%），HL检验P=0.327；仅视网膜特征模型的AUC为0.94（95% CI: 0.92-0.96），灵敏度为86.14%（95% CI: 77.63-91.82%），特异度90.47%（95% CI: 86.92-93.26%），HL检验P=0.659；综合模型的AUC为0.96（95% CI: 0.94-0.98），灵敏度为92.47%（95% CI: 84.62-96.83%），特异度为95.81%（95% CI: 92.93-97.64%），HL检验P=0.802，三模型均显示良好的校准度。与临床模型相比，仅视网膜特征模型的NRI为0.27（95% CI: 0.18-0.36，P<0.001），IDI为0.15（95% CI: 0.11-0.20，P<0.001）；综合模型的NRI为0.39（95% CI: 0.30-0.48，P<0.001），IDI为0.22（95% CI: 0.17-0.27，P<0.001），表明视网膜特征和综合模型均显著提升了分类能力。然而，与视网膜特征模型相比，综合模型的NRI为0.08（95% CI: -0.02-0.18，P=0.078），IDI为0.04（95% CI: -0.01-0.09，P=0.095），未达统计学显著性。DeLong检验进一步显示，综合模型与视网膜特征模型的AUC差异无统计学意义（P=0.132）。

结论

结果表明，视网膜图像比传统的临床参数包含更多有价值的信息，并且将临床特征添加到视网膜特征的综合风险评估模型中并没有统计学上显著提高模型的性能。基于其优异表现支持视网膜特征风险评估模型作为孕早期抑郁症状筛查和诊断的有效工具。未来需进一步开展多中心研究以验证模型的普适性。鉴于视网膜特征与多种妊娠不良结局相关，推动其在孕产妇围产健康管理中的联合应用。

Abstract

Objectives

Depression symptoms in early pregnancy pose a threat to maternal physical and mental health as well as fetal growth and development, potentially leading to adverse pregnancy outcomes such as miscarriage and preterm birth. Early identification and intervention are critical for ensuring maternal and infant health. However, traditional methods, such as the Edinburgh Postnatal Depression Scale (EPDS), rely on subjective reports and lack clear physiological or biochemical markers. Models based on conventional clinical risk factors have limited predictive performance. This study aims to explore the potential of automated retinal image analysis in assessing the risk of depressive symptoms in early pregnancy. By developing an objective, non-invasive risk model, we compared retinal characteristics between pregnant women with and without depressive symptoms, investigated microvascular damage features, evaluated the diagnostic value of retinal features, and identified key indicators associated with depression risk. The study proposes a retinal image-based assessment framework to optimize individualized risk evaluation, providing a novel clinical tool for mental health management in Chinese pregnant women.

Methods

This study adopted a case-control design and collected data from 1,052 pregnant women who underwent first-trimester examinations at Tongzhou District Maternal and Child Health Hospital between May 2023 and May 2024. Depressive symptoms were assessed using the EPDS, and baseline information was gathered through questionnaires. Biochemical test results were retrieved from the hospital information system, and retinal images of both eyes were captured using a Canon CR-2AF fundus camera. The Automatic Retinal Image Analysis (ARIA) technique was employed to calculate common retinal parameters, and the ResNet50 machine learning algorithm extracted high-dimensional retinal features associated with early pregnancy depressive symptoms. Principal component analysis was used for dimensionality reduction. Participants were matched by age (±2 years) and divided into a depression group and a control group. With EPDS diagnosis as the outcome variable and clinical and retinal features as independent variables, lasso regression was applied for feature selection. Logistic regression models were constructed, and ten-fold cross-validation was used to evaluate model performance. Three models were developed: clinical features only, retinal features only, and a combined model. Model discrimination was assessed using sensitivity, specificity, and area under the curve (AUC). Calibration was evaluated with the Hosmer-Lemeshow (HL) test. The net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices were used to analyze the added benefit of combined features, and the DeLong test compared performance differences between models to provide a basis for clinical application.

Results

A total of 449 pregnant women in the first trimester were finally included in this study, among which there were 92 cases in the depressive symptom group and 357 cases in the non-depressive symptom group. The AUC of the model with only clinical characteristics was 0.81 (95% CI: 0.76-0.86), the sensitivity was 72.42% (95% CI: 61.87-79.63%), the specificity was 82.73% (95% CI: 69.54-88.37%), and the P value of the HL test was 0.327. The AUC of the model with only retinal characteristics was 0.94 (95% CI: 0.92-0.96), the sensitivity was 86.14% (95% CI: 77.63-91.82%), the specificity was 90.47% (95% CI: 86.92-93.26%), and the P value of the HL test was 0.659. The AUC of the comprehensive model was 0.96 (95% CI: 0.94-0.98), the sensitivity was 92.47% (95% CI: 84.62-96.83%), the specificity was 95.81% (95% CI: 92.93-97.64%), and the P value of the HL test was 0.802. All three models showed good calibration. Compared with the clinical model, the NRI of the model with only retinal characteristics was 0.27 (95% CI: 0.18-0.36, P<0.001), and the IDI was 0.15 (95% CI: 0.11-0.20, P<0.001). The NRI of the comprehensive model was 0.39 (95% CI: 0.30-0.48, P<0.001), and the IDI was 0.22 (95% CI: 0.17-0.27, P<0.001), indicating that both the model with retinal characteristics and the comprehensive model significantly improved the classification ability. However, compared with the model with retinal characteristics, the NRI of the comprehensive model was 0.08 (95% CI: -0.02-0.18, P=0.078), and the IDI was 0.04 (95% CI: -0.01-0.09, P=0.095), which did not reach statistical significance. The DeLong test further showed that there was no significant difference in the AUC between the comprehensive model and the model with retinal characteristics (P=0.132).

Conclusions

The results indicate that retinal images contain more valuable information than traditional clinical parameters, and adding clinical characteristics to the comprehensive risk assessment model of retinal characteristics does not significantly improve the performance of the model statistically. Based on its excellent performance, the risk assessment model based on retinal characteristics is supported as an effective tool for screening and diagnosing depressive symptoms during the first trimester of pregnancy. Further multicenter studies are needed in the future to verify the universality of the model. Given the association between retinal characteristics and various adverse pregnancy outcomes, the joint application of retinal characteristics in the perinatal health management of pregnant and lying-in women should be promoted.