Feasibility study on quantifying retinal vascular features for predicting preeclampsia based on artificial intelligence models

Abstract

Objective·To explore the predictive capability of retinal vascular features in preeclampsia (PE) based on artificial intelligence (AI) models.Methods·This retrospective study enrolled 789 pregnant women who registered from June 2020 to January 2021 at Shanghai First Maternity and Infant Hospital of Tongji University in the first 16 weeks of gestation. These women underwent regular prenatal examinations, had retinal fundus images captured, and delivered singleton live births at the hospital. According to whether they developed hypertensive disorders of pregnancy (HDP), they were divided into unaffected group (n=685) and HDP group (n=104). Within the HDP group, pregnancies were further categorized into gestational hypertension (GH) group (n=36) and PE group (n=68) based on the occurrence of PE. Based on the gestational age at onset, the PE group was further divided into early-onset PE group (gestational age<34 weeks) and late-onset PE group (gestational age≥34 weeks). Fundus images of the pregnant women were obtained, and an AI algorithm was utilized to diagnose retinal lesions and quantify retinal vascular features. Comparative analyses were conducted on fundus features and retinal vascular features. Univariate Logistic regression model was employed to analyze the influencing factors of PE occurrence, and multivariate Logistic regression model was further utilized to assess the correlation between retinal vascular features and the occurrence of PE. The predictive capability of retinal vascular features for PE (both early- and late-onset PE) was analyzed by using area under the curve (AUC) of receiver operator characteristic curve (ROC curve).Results·The comparative analysis of fundus features and retinal vascular features demonstrated statistically significant differences between the unaffected group and PE group in terms of central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arteriole-to-venular ratio (AVR), retinal artery tortuosity and retinal artery fractal dimension (all P<0.05). Univariate Logistic regression analysis indicated that second-trimester mean arterial pressure (MAP), second-trimester estimated fetal weight (EFW), CRAE, CRVE, AVR, retinal artery tortuosity and retinal artery fractal dimension were the influencing factors for PE occurrence (all P<0.05). Multivariate Logistic regression analysis revealed that second-trimester EFW, CRAE, CRVE, AVR, retinal artery tortuosity and retinal artery fractal dimension were the protective factors for the occurrence of PE, while second-trimester MAP was the risk factor for PE (all P<0.05). The analysis of ROC curves revealed that maternal risk factors along with second-trimester prenatal examination data (including MAP and EFW) and retinal vascular features model had good predictive ability for PE [AUC (95% CI)=0.784 (0.725－0.843), and this model exhibited better predictive capability for early-onset PE, with an AUC (95% CI) of 0.840 (0.756－0.924).Conclusion·The integration of quantified retinal vascular features based on AI models with maternal risk factors and second-trimester prenatal examination data (including MAP and EFW) enables a more effective prediction of PE occurrence, particularly early-onset PE.

Keywords

• preeclampsia (pe)
• retinal vessel
• artificial intelligence (ai)