Abstract
Introduction:
Early prediction of preeclampsia (PE) is of universal importance in controlling the disease process. Our study aimed to assess the feasibility of using retinal fundus images to predict preeclampsia via deep learning in singleton pregnancies.
Methods:
This prospective cohort study was conducted at Shanghai First Maternity and Infant Hospital, Tongji University School of Medicine. Eligible participants included singleton pregnancies who presented for prenatal visits before 14 weeks of gestation from September 1, 2020, to February 1, 2022. Retinal fundus images were obtained using a nonmydriatic digital retinal camera during their initial prenatal visit upon admission before 20 weeks of gestation. In addition, we generated fundus scores, which indicated the predictive value of hypertension, using a hypertension detection model. To evaluate the predictive value of the retinal fundus image-based deep learning algorithm for preeclampsia, we conducted stratified analyses and measured the area under the curve (AUC), sensitivity, and specificity. We then conducted sensitivity analyses for validation.
Results:
Our study analyzed a total of 1138 women, 92 pregnancies developed into hypertension disorders of pregnancy (HDP), including 26 cases of gestational hypertension and 66 cases of preeclampsia. The adjusted odds ratio (aOR) of the fundus scores was 2.582 (95% CI, 1.883–3.616; P < 0.001). Otherwise, in the categories of prepregnancy BMI less than 28.0 and at least 28.0, the aORs were 3.073 (95%CI, 2.265–4.244; P < 0.001) and 5.866 (95% CI, 3.292–11.531; P < 0.001). In the categories of maternal age less than 35.0 and at least 35.0, the aORs were 2.845 (95% CI, 1.854–4.463; P < 0.001) and 2.884 (95% CI, 1.794–4.942; P < 0.001). The AUC of the fundus score combined with risk factors was 0.883 (sensitivity, 0.722; specificity, 0.934; 95% CI, 0.834–0.932) for predicting preeclampsia.
Conclusion:
Our study demonstrates that the use of deep learning algorithm-based retinal fundus images offers promising predictive value for the early detection of preeclampsia.
Keywords: deep learning, prediction model, preeclampsia, retinal vessels, receiver operating characteristic curve
INTRODUCTION
Preeclampsia (PE) is one of the most severe complications of pregnancy [1,2], affecting approximately 3–5% of pregnancies globally, and is defined as new-onset hypertension with proteinuria or end-organ dysfunction after 20 weeks of gestation [3,4]. Unfortunately, there is no cure for preeclampsia except for placental delivery, and the disease can lead to severe adverse maternal or fetal outcomes if left untreated [5,6]. Therefore, the early prediction of preeclampsia is critical [7].
Currently, clinical risk assessment for preeclampsia relies mainly on maternal history and definitive guidelines or recommendations. The American College of Obstetricians and Gynecologists (ACOG) Practice Bulletin, Number 222 [8], recommends low-dose aspirin for pregnancies with any high-risk factors for preeclampsia or more than one moderate-risk factor for the primary prevention of preeclampsia [9,10]. However, this approach has limited detection rates of 5 and 2% for preterm and term preeclampsia [11], respectively. Although research has shown that there is no increase in maternal and neonatal adverse outcomes with prophylactic use of aspirin [12–15], such as bleeding, placental abruption, premature closure of the arterial canal, and neonatal bleeding, the broadening of the indications for aspirin has led to potential exposure to these side effects. To improve detection rates, recent screening methods have combined clinical risk factors with convincing predictors, such as ultrasound and maternal biochemical markers [16–19], which have been proven to have positive effects [16,17]. However, these predictors are expensive and highly dependent on technical expertise, making them less accessible in underdeveloped countries. Therefore, there is a need for additional noninvasive, inexpensive, and objectively monitored predictive indicators for routine prenatal visits [18,19].
Preeclampsia is associated with an elevated risk of damage to various organs, including the kidneys, liver, brain, eyes, and blood system [20–24]. As the retina is an extension of the brain and provides a convenient means of visualizing vascular changes in vivo[25], fundus images may reveal microvascular changes in the retina [26] that reflect changes in the cardiovascular system. Previous research has demonstrated that microvascular changes in the retinal vasculature are independent predictors of hypertension [27]. Therefore, we sought to predict preeclampsia by monitoring changes in the retinal vasculature using fundus images. Nonmydriatic fundus photography, which is widely used in hospitals and primary care settings, can easily capture retinal fundus images [28].
Given the advantages of deep learning in developing disease prediction models [29], we utilized this technique to automatically analyze fundus images because of its speed and accuracy in image processing. The details of the medical images can be automatically assessed quantitatively [30]. Deep learning techniques, such as Convolutional Neural Networks (CNNs), which are optimized for images [31], allow for a clear representation of raw data input, producing highly accurate algorithms for recognizing disease characteristics [32]. Previous studies have shown that a CNN-based retinal artificial intelligence diagnosis system has diagnostic accuracy equal to or better than that of ophthalmologists [33].
In our prospective cohort study of singleton pregnancies, we aimed to evaluate and synthesize the predictive value of retinal fundus images in preeclampsia using deep learning techniques.
MATERIALS AND METHODS
Study design and participants
This prospective cohort study was conducted at the Shanghai First Maternity and Infant Hospital (SFMIH), affiliated with Tongji University School of Medicine. Singleton pregnancies attending prenatal visits before 14 weeks of gestation at SFMIH were recruited between September 1, 2020, and February 1, 2022. Participants were offered several antenatal visits during which their weight and blood pressure were recorded. They underwent retinal fundus examinations during their initial prenatal visit upon admission before 20 weeks of gestation.
A flowchart of patient selection is shown in Fig. 1. The inclusion criteria were as follows: singleton pregnancies, attending prenatal visits before 14 weeks of gestation in SFMIH, taking fundus images before 20 weeks of gestation, and delivered in SFMIH between 28 and 42 weeks of gestation. The exclusion criteria were as follows: pregnancies not delivered in SFMIH; intrauterine fetal demise; multiple pregnancies; inability to perform fundus photography examination; and any ophthalmic diseases such as refractive media opacity or retinal diseases, which would hinder the acquisition of the fundus image or influence the image quality or the process of analysis.
FIGURE 1.
Flow chart for patient selection.
Data collection
Hospital electronic medical records were obtained based on the inclusion criteria. The collected information included sociodemographic characteristics, obstetric characteristics, and neonatal outcomes.
Primary outcome
Preeclampsia was diagnosed using the standardized criteria suggested by the ACOG [8]. The outcome was collected by certified medical record abstractors to determine the preeclampsia diagnosis.
Measurements
Gestational weeks were estimated using first trimester ultrasonographic measurements. Prepregnancy BMI (kg/m2) was calculated based on self-reported prepregnancy weight (kg) and height (m) at the first prenatal visit from the medical chart and categorized as normal weight (BMI<24.0), overweight (24.0 ≤ BMI<28.0), or obesity (BMI ≥ 28.0), according to the standards for Chinese participants proposed by the Chinese Working Group on Obesity [34]. Fundus images were obtained before 20 weeks of gestation using a nonmydriatic digital retinal camera (CA-CR-2/2 PLUS AF). Each image was independently evaluated by two experienced ophthalmologists, and images that combined other ophthalmic pathological situations were excluded.
Fundus scores generated via the deep learning algorithm
In our study, the predictive values of hypertension were generated by the hypertension detection model based on a deep learning algorithm [25], which is an artificial intelligence recognition tool for assessing hypertension. The algorithm was developed by Airdoc Ltd. in 2020 [35], in which a hypertension prediction branch was added to the Inception-ResNet-v2 backbone. The model accepts input from the fundus images of each participant, and the trained neural networks were used to export predictive values of hypertension to predict the probability of hypertension and were defined as ‘fundus scores’. Fundus scores indicated hypertension severity, with values of 0 to 1.0, representing increasing signs of typical hypertension. Because of the skewed distribution, we loge transformed fundus scores and modeled them as a continuous variable; the higher the values of the model outputs, the greater the influence of hypertension on fundus tissue, especially microvascular.
Before implementing the algorithm, an image quality control model was applied to filter out unqualified images. The quality assessment was evaluated using a deep learning model based on the architecture of the variational autoencoder and generative adversarial networks, which was developed by Airdoc Ltd. in 2021 [36].
Evaluating the predictive value of fundus score in preeclampsia
To evaluate the predictive value of the fundus score in preeclampsia, we used stratified analyses and multivariate logistic regression to assess the diagnostic value of the fundus score in predicting hypertension disorders of pregnancy (HDP), gestational hypertension, and preeclampsia. The area under the curve (AUC), sensitivity, and specificity were determined to assess the predictive values. Sensitivity analyses were used to further validate universality.
Statistical analysis
Continuous variables are expressed as medians (interquartile range [IQR]), and n (%) for categorical variables. One-way ANOVA, Pearson's chi-square test, and Fisher's exact test were used to compare the outcome groups for continuous and categorical variables.
All data were double-entered using the EpiData software (EpiData 3.1 Windows, EpiData Association Odense, Denmark). All analyses were performed using R Studio (version 2022.07.1 Build 554) with R (version 4.2.1).
Ethics statement and consent
This study was approved by the Ethics Committee of Shanghai First Maternity and Infant Hospital (reference number: KS20268). All procedures were performed in accordance with the tenets of the Declaration of Helsinki. All retinal fundus images were anonymized and de-identified prior to the analysis. Informed consent was obtained from all participants.
RESULTS
Characteristics of the study population
A total of 1209 singleton pregnancies attending prenatal visits before 14 weeks of gestation were recruited, and retinal fundus images were obtained before 20 weeks of gestation. After excluding 12 cases of intrauterine fetal demise, and 59 multiple pregnancies, 1138 women were included in the final analysis, with 92 pregnancies developing into HDP, including 26 cases of gestational hypertension and 66 cases of preeclampsia (Fig. 1). Representative examples of fundus images of unaffected pregnancies and gestational hypertension and preeclampsia pregnancies are shown in Fig. 2.
FIGURE 2.
Examples of the original fundus images. Original fundus images of (a) unaffected pregnancy; (b) GH; and (c) PE patients.
Table 1 summarizes the demographic, obstetric, and neonatal characteristics. Chronic hypertension, gestational weeks, preterm birth, fetal birth weight, and Apgar scores differed significantly among the groups of unaffected pregnancies, gestational hypertension, and PE pregnancies. However, there were no significant differences in primiparity, rate of assisted reproductive technology (ART), occurrence of hypothyroidism, gestational diabetes mellitus (GDM), polycystic ovary syndrome (PCOS), or postpartum hemorrhage (PPH) among the three groups. Maternal age at delivery, BMI at the first prenatal visit, previous pregnancy with preeclampsia, autoimmune disease, diabetes mellitus, delivery method, and small for gestational age (SGA) were significantly different between unaffected pregnancies and gestational hypertension pregnancies.
TABLE 1.
Baseline characteristics
| HDP (n = 92) | |||||
| Characteristics | Unaffected (n = 1046) | GH (n = 26) | PE (n = 66) | P valuea | P valueb |
| Demographic characteristics | |||||
| Primiparity, n (%) | 769 (73.5%) | 22 (84.6%) | 44 (66.7%) | 0.123 | 0.711 |
| Maternal age at delivery (year), Median (IQR) | 31.00 (29.00, 33.00) | 30.50 (27.25, 32.00) | 33.00 (30.00, 36.00) | 0.002 | <0.001 |
| Maternal age at delivery (year), n (%) | 0.015 | <0.001 | |||
| <35.0 | 876 (83.7%) | 23 (88.5%) | 38 (57.6%) | ||
| 35.0–40.0 | 153 (14.6%) | 3 (11.5%) | 20 (30.3%) | ||
| ≥40 | 17 (1.6%) | 0 (0%) | 8 (12.1%) | ||
| BMI at first prenatal visit (kg/m2) | 21.45 (19.84, 23.43) | 22.32 (19.65, 24.36) | 26.06 (22.78, 29.04) | <0.001 | <0.001 |
| BMI at first prenatal visit, n (%) | 0.003 | <0.001 | |||
| ≤24.0 | 840 (80.3%) | 19 (73.1%) | 22 (33.3%) | ||
| 24.0–28.0 | 170 (16.3%) | 4 (15.4%) | 24 (36.4%) | ||
| ≥28.0 | 36 (3.4%) | 3 (11.5%) | 20 (30.3%) | ||
| Obstetric characteristics | |||||
| Previous pregnancy with PE, n (%) | 6 (0.6%) | 0 (0%) | 3 (4.5%) | 0.556 | 0.030 |
| ART, n (%) | 51 (4.9%) | 1 (3.8%) | 8 (12.1%) | 0.437 | 0.052 |
| Chronic hypertension, n (%) | 5 (0.5%) | 0 (0%) | 17 (25.8%) | 0.002 | <0.001 |
| Autoimmune disease, n (%) | 23 (2.2%) | 2 (7.7%) | 4 (6.1%) | >0.999 | 0.025 |
| DM, n (%) | 7 (0.7%) | 0 (0%) | 3 (4.5%) | 0.556 | 0.040 |
| Hypothyroidism, n (%) | 81 (7.7%) | 4 (15.4%) | 6 (9.1%) | 0.460 | 0.289 |
| GDM, n (%) | 112 (10.7%) | 3 (11.5%) | 15 (22.7%) | 0.261 | 0.010 |
| PCOS, n (%) | 5 (0.5%) | 0 (0%) | 1 (1.5%) | >0.999 | 0.398 |
| Delivery method, n (%) | 0.003 | <0.001 | |||
| Spontaneous delivery | 606 (57.9%) | 12 (46.2%) | 10 (15.2%) | ||
| Caesarean section | 440 (42.1%) | 14 (53.8%) | 56 (84.8%) | ||
| Gestational weeks (week), Median (IQR) | 39.29 (38.43, 40.14) | 38.86 (38.14, 39.39) | 37.43 (36.04, 38.79) | 0.002 | <0.001 |
| Preterm birth, n (%) | 0.030 | <0.001 | |||
| Normal | 1008 (96.4%) | 24 (92.3%) | 43 (65.2%) | ||
| PTB (28–37 GW) | 36 (3.4%) | 2 (7.7%) | 21 (31.8%) | ||
| extremely PTB (<28 GW) | 2 (0.2%) | 0 (0%) | 2 (3.0%) | ||
| PPH, n (%) | 20 (1.9%) | 0 (0%) | 2 (3.0%) | >0.999 | 0.696 |
| Neonatal characteristics | |||||
| Fetal birth weight (g), Median (IQR) | 3330.00 (3070.00, 3590.00) | 3135.00 (2922.50, 3512.50) | 2955.00 (2341.25, 3388.75) | 0.103 | <0.001 |
| Low birth weight, n (%) | 25 (2.4%) | 2 (7.7%) | 22 (33.3%) | 0.016 | <0.001 |
| Fetal growth, n (%) | 0.750 | 0.111 | |||
| AGA | 900 (86.0%) | 23 (88.5%) | 55 (83.3%) | ||
| SGA | 33 (3.2%) | 0 (0%) | 0 (0%) | ||
| LGA | 113 (10.8%) | 3 (11.5%) | 11 (16.7%) | ||
| 1-minutes Apgar score, Median (IQR) | 9.00 (4.00, 10.0) | 9.00 (8.00, 10.0) | 9.00 (3.00, 10.0) | 0.017 | 0.003 |
| 1-minutes Apgar score ≤7, n (%) | 8 (0.8%) | 0 (0%) | 4 (6.1%) | 0.574 | 0.012 |
| 5-minutes Apgar score, Median (IQR) | 10.0 (6.00, 10.0) | 10.0 (9.00, 10.0) | 10.0 (7.00, 10.0) | 0.008 | <0.001 |
| 5-minutes Apgar score ≤7, n (%) | 1 (1%) | 0 (0%) | 1 (1.5%) | >0.999 | 0.155 |
Note: Variables are presented as Median (IQR) or n (%). HDP, hypertension disorders of pregnancy; GH, gestational hypertension; PE, preeclampsia; IQR, interquartile range; BMI, Body mass index; ART, assisted reproductive technology; DM, pregestational diabetes; GDM, gestational diabetes; PCOS, polycystic ovary syndrome; PTB, preterm birth; PPH, postpartum hemorrhage; AGA, average for gestational age; SGA, small for gestational age; LGA, large for gestational age.
Unaffected vs. HDP groups.
GH vs. PE groups.
Fundus scores significantly differed among unaffected, gestational hypertension, and preeclampsia groups
First, a correlation analysis was performed to elucidate the absence of a linear relationship between fundus scores and gestational weeks (Fig. 3). Subsequently, in light of the crucial need to discern the association between fundus score and the occurrence of HDP, logistic regression analyses were performed (Table 2), and the results demonstrated significant differences in fundus scores among HDP (adjusted odds ratio [aOR], 2.547; 95% CI, 1.944–3.385, P < 0.05), gestational hypertension (aOR, 2.144; 95% CI, 1.397–3.267, P < 0.05), and preeclampsia (aOR, 2.582; 95% CI, 1.883–3.616, P < 0.001).
FIGURE 3.
Linear correlation between fundus score and gestational weeks before 28 weeks of gestation. (a) Linear correlation for all participants. (b) Linear correlation for HDP pregnancies. (c) Linear correlation for GH pregnancies. (d) Linear correlation for PE pregnancies.
TABLE 2.
The stratified analyses of the association between hypertensive disorders of pregnancy and fundus scores
| Stratification factors | Unaffected | HDP | OR (95% CI) | GH | OR (95% CI) | PE | OR (95% CI) |
| Overall | 1046 | 92 | 26 | 66 | |||
| Crude | 3.040 (2.456–3.814)∗∗∗ | 1.487 (1.055–2.043)∗ | 3.714 (2.884–4.877)∗∗∗ | ||||
| Adjusted | 2.547 (1.944–3.385)∗∗∗ | 2.144 (1.397–3.267)∗∗∗ | 2.582 (1.883–3.616)∗∗∗ | ||||
| Maternal age (years) | |||||||
| <35.0 | 876 | 61 | 23 | 38 | |||
| Crude | 3.128 (2.381–4.188)∗∗∗ | 1.694 (1.130–2.477)∗∗ | 3.919 (2.809–5.644)∗∗∗ | ||||
| Adjusted | 2.525 (1.816–3.551)∗∗∗ | 1.896 (1.181–3.024)∗∗ | 2.845 (1.854–4.463)∗∗∗ | ||||
| ≥35.0 | 170 | 31 | 3 | 28 | |||
| Crude | 3.235 (2.145–5.165)∗∗∗ | 1.855 (0.680–4.629) | 3.132 (2.071–5.009)∗∗∗ | ||||
| Adjusted | 3.109 (1.937–5.358)∗∗∗ | 5.115 (1.267–38.151)∗ | 2.884 (1.794–4.942)∗∗∗ | ||||
| BMI at first prenatal visit (kg/m2) | |||||||
| <28.0 | 1010 | 69 | 23 | 46 | |||
| Crude | 2.212 (1.693–2.919)∗∗∗ | 1.580 (1.060–2.276)∗ | 2.645 (1.886–3.788)∗∗∗ | ||||
| Adjusted | 2.954 (2.274–3.890)∗∗∗ | 2.244 (1.461–3.441)∗∗∗ | 3.073 (2.265–4.244)∗∗∗ | ||||
| ≥28.0 | 36 | 23 | 3 | 20 | |||
| Crude | 5.281 (3.385–8.921)∗∗∗ | 1.215 (0.571–2.381) | 6.118 (3.775–10.887)∗∗∗ | ||||
| Adjusted | 5.914 (3.388–11.289)∗∗∗ | 1.955 (0.799–5.150) | 5.866 (3.292–11.531)∗∗∗ |
CI, confidence interval; GH, gestational hypertension; HDP, hypertension disorders of pregnancy; OR, odds ratio; PE, preeclampsia.
P < 0.05.
P < 0.01.
P < 0.001.
We next established the association between the fundus score and HDP when participants were stratified based on a maternal age of at least 35.0 and BMI of at least 28.0 (Table 2). Individuals in the high-risk population subgroup (maternal age ≥ 35.0) had increased aORs of HDP, gestational hypertension, and preeclampsia prediction of 3.109 (95% CI, 1.937–5.358), 5.115 (95% CI, 1.267–38.151), and 2.884 (95% CI, 1.794–4.942), respectively. Meanwhile, the subgroup of overweight pregnancies (BMI ≥28.0) had increased aOR of HDP, GH, and PE prediction of 5.914 (95% CI, 3.388–11.289), 1.955 (95% CI, 0.799–5.150), and 5.866 (95% CI, 3.292–11.531), respectively.
Predictive performance of preeclampsia
Subsequently, we conducted a receiver operating characteristic curve (ROC) analysis to evaluate the predictive performance of the fundus score, and the AUCs, sensitivity, and specificity are presented in Table 3. Figure 4 shows the ROC curves for the predictive performance of HDP, gestational hypertension, and preeclampsia.
TABLE 3.
Screening performance of different marker combinations for hypertensive disorders of pregnancy
| Groups screening indicators | AUC | 95% CI | Sensitivity | Specificity |
| HDP | ||||
| Fundus score | 0.800 | 0.746–0.853 | 0.663 | 0.847 |
| Risk factors | 0.685 | 0.625–0.745 | 0.551 | 0.804 |
| Fundus score + risk factors | 0.837 | 0.790–0.885 | 0.786 | 0.745 |
| GH | ||||
| Fundus score | 0.641 | 0.523–0.759 | 0.500 | 0.811 |
| Risk factors | 0.509 | 0.412–0.606 | 0.038 | 0.984 |
| Fundus score + risk factors | 0.775 | 0.696–0.854 | 0.962 | 0.519 |
| PE | ||||
| Fundus score | 0.845 | 0.791–0.899 | 0.681 | 0.904 |
| Risk factors | 0.742 | 0.675–0.810 | 0.667 | 0.803 |
| Fundus score + risk factors | 0.883 | 0.834–0.932 | 0.722 | 0.934 |
| EPE | ||||
| Fundus score | 0.844 | 0.760–0.928 | 1.000 | 0.551 |
| Risk factors | 0.694 | 0.516–0.872 | 0.692 | 0.779 |
| Fundus score + risk factors | 0.872 | 0.762–0.983 | 0.769 | 0.897 |
| LPE | ||||
| Fundus score | 0.838 | 0.771–0.906 | 0.698 | 0.902 |
| Risk factors | 0.744 | 0.669–0.819 | 0.642 | 0.794 |
| Fundus score + risk factors | 0.874 | 0.814–0.933 | 0.736 | 0.909 |
AUC, area under the curve; CI, confidence interval; HDP, hypertension disorders of pregnancy; GH, gestational hypertension; PE, preeclampsia; EPE, early-onset preeclampsia; LPE, late-onset preeclampsia.
FIGURE 4.
ROC curves of the predictive performance of different methods for HDP, GH, and PE. (a) ROC curves of the predictive performance for HDP. (b) ROC curves of the predictive performance for GH. (c) ROC curves of the predictive performance for PE. (d) ROC curves of the predictive performance for early-onset PE. (e) ROC curves of the predictive performance for late-onset PE. AUC, area under the curve; CI, confidence interval.
Regarding the prediction of HDP, The AUC of the fundus score and the combination of the fundus score and risk factors were 0.800 (sensitivity, 0.663; specificity, 0.847; 95% CI, 0.746–0.853) and 0.837 (sensitivity, 0.786; specificity, 0.745; 95% CI, 0.790–0.885). To further evaluate the predictive ability of the fundus score, we calculated its AUC of the fundus score in predicting preeclampsia. The AUCs of the fundus score and the combination of the fundus score and risk factors were 0.845 (sensitivity, 0.681; specificity, 0.904; 95% CI, 0.791–0.899) and 0.883 (sensitivity, 0.722; specificity, 0.934; 95% CI, 0.834–0.932), respectively (Table 4).
TABLE 4.
The sensitivity analyses of the association between hypertensive disorders of pregnancy and fundus scores
| Unaffected | HDP | OR (95% CI) | GH | OR (95% CI) | PE | OR (95% CI) | |
| Sensitive analysis 1a | 1041 | 75 | 26 | 49 | |||
| Crude | 2.871 (2.270–3.680)∗∗∗ | 1.620 (1.132–2.272)∗∗ | 3.412 (2.577–4.606)∗∗∗ | ||||
| Adjusted | 2.951 (2.231–3.967)∗∗∗ | 2.102 (1.374–3.186)∗∗∗ | 3.160 (2.264–4.523)∗∗∗ | ||||
| Sensitive analysis 2b | 1023 | 86 | 24 | 62 | |||
| Crude | 3.102 (2.489–3.919)∗∗∗ | 1.599 (1.126–2.218)∗∗ | 3.607 (2.797–4.744)∗∗∗ | ||||
| Adjusted | 2.698 (2.051–3.600)∗∗∗ | 2.429 (1.583–3.737)∗∗∗ | 2.641 (1.924–3.697)∗∗∗ | ||||
| Sensitive analysis 3c | 1039 | 89 | 26 | 63 | |||
| Crude | 2.975 (2.405–3.728)∗∗∗ | 1.504 (1.066–2.069)∗ | 3.545 (2.765–4.631)∗∗∗ | ||||
| Adjusted | 2.461 (1.893–3.24)∗∗∗ | 2.112 (1.383–3.197)∗∗∗ | 2.480 (1.829–3.419)∗∗∗ | ||||
| Sensitive analysis 4d | 1040 | 89 | 26 | 63 | |||
| Crude | 3.123 (2.521–3.921)∗∗∗ | 1.484 (1.053–2.039)∗ | 3.813 (2.962–5.008)∗∗∗ | ||||
| Adjusted | 2.632 (2.016–3.483)∗∗∗ | 2.102 (1.374–3.186)∗∗∗ | 2.728 (1.998–3.804)∗∗∗ | ||||
| Sensitive analysis 5e | 934 | 74 | 23 | 51 | |||
| Crude | 2.997 (2.368–3.852)∗∗∗ | 1.527 (1.055–2.149)∗ | 3.625 (2.742–4.905)∗∗∗ | ||||
| Adjusted | 2.489 (1.856–3.388)∗∗∗ | 2.263 (1.429–3.558)∗∗∗ | 2.443 (1.729–3.527)∗∗∗ | ||||
| Sensitive analysis 6f | 842 | 42 | 19 | 23 | |||
| Crude | 2.130 (1.581–2.906)∗∗∗ | 1.663 (1.081–2.480)∗ | 2.336 (1.612–3.454)∗∗∗ | ||||
| Adjusted | 2.296 (1.630–3.283)∗∗∗ | 2.356 (1.392–3.928)∗∗∗ | 2.147 (1.407–3.332)∗∗∗ |
Multivariate logistic regression analysis performed among pregnancies without chronic hypertension.
Multivariate logistic regression analysis performed among pregnancies without autoimmune diseases.
Multivariate logistic regression analysis performed among pregnancies without DM.
Multivariate logistic regression analysis performed among pregnancies without previous pregnancy with PE.
Multivariate logistic regression analysis performed among pregnancies without GDM.
Multivariate logistic regression analysis performed among pregnancies without high-risk factors for PE.
P < 0.05.
P < 0.01.
P < 0.001.
Sensitivity analysis
Subsequently, we performed sensitivity analysis by varying the risk factors for preeclampsia. Chronic hypertension, autoimmune diseases, previous pregnancy with preeclampsia, and diabetes mellitus were selected as crucial parameters for the multivariable logistic regression analysis. The results revealed that the fundus score exhibited favorable predictive performance for preeclampsia, excluding pregnancies with chronic hypertension (aOR, 3.160; 95% CI, 2.264–4.523), autoimmune diseases (aOR, 2.641; 95% CI, 1.924–3.697), diabetes mellitus (aOR, 2.480; 95% CI, 1.829–3.419), previous pregnancy with preeclampsia (aOR, 2.728; 95% CI, 1.998–3.804), GDM (aOR, 2.443; 95% CI, 1.729–3.527), and high-risk factors (aOR, 2.147; 95% CI, 1.407–3.332). These findings highlight the robustness of our results when uncertainty is considered, thereby bolstering the reliability and credibility of our study outcomes.
DISCUSSION
Main findings
In this study, we evaluated the use of deep learning on retinal fundus images for predicting preeclampsia in singleton pregnancies. Our findings demonstrated strong individual predictive values, with AUCs of 0.800, 0.641, and 0.845 for HDP, gestational hypertension, and preeclampsia, respectively. The excellent image processing capabilities of CNNs, combined with the emergence of nonmydriatic fundus photography, offer a noninvasive and promising approach for closely monitoring the health status and development of preeclampsia.
Preeclampsia is a global pregnancy complication that can result in short-term and long-term adverse outcomes for both mothers and newborns [3]. However, there is no cure for preeclampsia except for placental delivery. A combination of maternal factors, biomarkers, and ultrasound markers can improve the predictive performance of preeclampsia. However, serum biomarker tests are invasive and costly, and the accuracy of the uterine artery pulsatility index (UtA-PI) is highly dependent on expensive equipment and specialized expertise that may not be available in remote areas. Given the notable alterations in retinal vascular patterns observed in pregnancies affected by preeclampsia and the independent predictive value of microvascular changes in the retinal vasculature for hypertension [27], researches have explored the utilization of arteriovenous ratio assessments of retinal arterioles and venules as predictive indicators for preeclampsia [37]. Consequently, our study aims to predict the occurrence of preeclampsia by monitoring changes in the retinal vasculature through the analysis of fundus images and quantification of retinal alterations. Nonmydriatic fundus photography, which is widely used in hospitals and primary care settings, can easily capture retinal fundus images [28]. Therefore, our study assessed the application value of nonmydriatic fundus photography in preeclampsia prediction.
Screening for preeclampsia typically involves the combination of maternal history with various convincing predictors [38]. The current best-performing screening algorithm is the FMF algorithm, which includes maternal factors, UtA-PI, MAP, placental growth factor (PlGF), and soluble fms-like tyrosine kinase-1 (sFlt-1) [17]. However, expensive and invasive examinations were also included. In contrast, deep learning algorithms based on retinal fundus images are noninvasive, low-cost, and do not require expert technicians to simplify the prediction process and save resources. Our study found that the fundus score alone showed an effective predictive value for preeclampsia, and fundus photography elevated the predictive performance of maternal risk factors. Specifically, a sensitivity of 0.883 for predicting preeclampsia showed superiority in screening and warning about risks. This approach may be particularly useful in underdeveloped regions with large populations where expensive and invasive examinations cannot be performed. As retinal microvascular parameters are negatively correlated with pregnancy weeks, including vessel density ratios of the macula and optic disc [39], we recommend retinal fundus examination before gestational week 20.
Deep-learning techniques, particularly CNNs, have revolutionized image processing and feature extraction, enabling the identification of abstract and complex imaging characteristics with high accuracy and efficiency [40–42]. In ophthalmology, CNN-based disease prediction models have demonstrated excellent performance in detecting fundus lesions associated with various diseases [43–46]. Given that retinal vascular anomalies, such as arteriolar narrowing, tortuosity, and segmental retinal artery vasospasm, have been observed in 70% of fundus examinations in pregnancies complicated by preeclampsia [47], the integration of fundus photography with CNNs represents a promising approach for predicting this disorder. By leveraging the power of CNNs, this approach offers exceptional clinical potential for enhancing diagnostic precision, reducing healthcare costs, and expanding the screening reach in regions with high population density. Furthermore, the development of handheld fundus cameras has simplified image acquisition and demonstrated promising feasibility, with imaging quality comparable to that of traditional tabletop fundus imaging devices [48].
Strengths and limitations
This study has several limitations that need to be acknowledged. First, the deep learning algorithm was not specifically developed for pregnant women but rather based on the general population, which may limit its predictive performance. Future studies should consider incorporating pregnancy-related features to improve the accuracy of the algorithm. Second, external validation is necessary to determine the generalizability of the algorithm to other healthcare settings. Third, the changes in fundus scores throughout pregnancy require further investigation in both high-risk and low-risk pregnancies to fully understand their predictive value in different gestational periods.
CONCLUSION
In this study, we demonstrated the practical usefulness of a deep learning algorithm that utilizes retinal fundus images for predicting preeclampsia. The employment of CNNs has facilitated the emergence of fundus photography, providing a noninvasive and cost-effective means of predicting preeclampsia, even in underdeveloped regions with limited access to ophthalmologists.
ACKNOWLEDGEMENTS
The authors thank our patients and all participants for the data collection.
This study was supported by the Shanghai Municipal Health Commission (202040128) and the Natural Science Foundation of Shanghai (21142201800).
X.L.H. and P.G. conceived of and designed the study. T.F.Z., S.Y.G., F.X.S., and P.L. performed data analysis and drafted the manuscript. Y.L.W., J.H.X., B.W., and C.C.Z. revised the manuscript. All authors have read and approved the final manuscript.
Conflicts of interest
The authors report no conflict of interest.
Tianfan Zhou and Shengyi Gu contributed equally to this work.
Peng Gao and Xiaolin Hua were both in charge of corresponding.
Abbreviations: ART, Assisted reproductive technology; AUC, The area under curve; CI, Confidence interval; CNNs, Convolutional Neural Networks; DBP, Diastolic pressure; DM, Pregestational diabetes; EFW, Estimated fetal weight; GDM, Gestational diabetes; HDP, Hypertension disorders of pregnancy; IQR, Interquartile range; MAP, Mean arterial pressure; MoM, Multiples of the normal median; NICU, Neonatal ICU; OR, odds ratio; PCOS, Polycystic ovary syndrome; PE, Preeclampsia; PlGF, Placental growth factor; PPBMI, Prepregnant body mass index; SD, Standard deviation; sFlt-1, Soluble fms-like tyrosine kinase-1; UtA-PI, Uterine artery pulsatility index
REFERENCES
- 1.Rana S, Lemoine E, Granger JP, Karumanchi SA. Preeclampsia: pathophysiology, challenges, and perspectives. Circul Res 2019; 124:1094–1112. [DOI] [PubMed] [Google Scholar]
- 2.Magee LA, Nicolaides KH, von Dadelszen P. Preeclampsia. N Engl J Med 2022; 386:1817–1832. [DOI] [PubMed] [Google Scholar]
- 3.Chappell LC, Cluver CA, Kingdom J, Tong S. Preeclampsia. Lancet 2021; 398:341–354. [DOI] [PubMed] [Google Scholar]
- 4.Say L, Chou D, Gemmill A, Tuncalp O, Moller AB, Daniels J, et al. Global causes of maternal death: a WHO systematic analysis. Lancet Glob Health 2014; 2:e323–e333. [DOI] [PubMed] [Google Scholar]
- 5.McCarthy FP, Ryan RM, Chappell LC. Prospective biomarkers in preterm preeclampsia: a review. Pregnancy Hypertens 2018; 14:72–78. [DOI] [PubMed] [Google Scholar]
- 6.Lawn JE, Blencowe H, Waiswa P, Amouzou A, Mathers C, Hogan D, et al. Stillbirths: rates, risk factors, and acceleration towards 2030. Lancet 2016; 387:587–603. [DOI] [PubMed] [Google Scholar]
- 7.Lisonkova S, Joseph KS. Incidence of preeclampsia: risk factors and outcomes associated with early- versus late-onset disease. Am J Obstet Gynecol 2013; 209:544 e1–544 e12. [DOI] [PubMed] [Google Scholar]
- 8.Gestational Hypertension and Preeclampsia: ACOG Practice Bulletin, Number 222. Obstet Gynecol 2020; 135:e237–e260. [DOI] [PubMed] [Google Scholar]
- 9.Rolnik DL, Wright D, Poon LC, O’Gorman N, Syngelaki A, de Paco Matallana C, et al. Aspirin versus placebo in pregnancies at high risk for preterm preeclampsia. N Engl J Med 2017; 377:613–622. [DOI] [PubMed] [Google Scholar]
- 10.Lin L, Huai J, Li B, Zhu Y, Juan J, Zhang M, et al. A randomized controlled trial of low-dose aspirin for the prevention of preeclampsia in women at high risk in China. Am J Obstet Gynecol 2022; 226:251 e1–251.e12. [DOI] [PubMed] [Google Scholar]
- 11.O’Gorman N, Wright D, Poon LC, Rolnik DL, Syngelaki A, de Alvarado M, et al. Multicenter screening for preeclampsia by maternal factors and biomarkers at 11-13 weeks’ gestation: comparison with NICE guidelines and ACOG recommendations. Ultrasound Obstet Gynecol 2017; 49:756–760. [DOI] [PubMed] [Google Scholar]
- 12.Atallah A, Lecarpentier E, Goffinet F, Doret-Dion M, Gaucherand P, Tsatsaris V. Aspirin for prevention of preeclampsia. Drugs 2017; 77:1819–1831. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Groom KM, David AL. The role of aspirin, heparin, and other interventions in the prevention and treatment of fetal growth restriction. Am J Obstet Gynecol 2018; 218 (2S):S829–S840. [DOI] [PubMed] [Google Scholar]
- 14.Rotchell YE, Cruickshank JK, Gay MP, Griffiths J, Stewart A, Farrell B, et al. Barbados Low Dose Aspirin Study in Pregnancy (BLASP): a randomised trial for the prevention of preeclampsia and its complications. Br J Obstet Gynaecol 1998; 105:286–292. [DOI] [PubMed] [Google Scholar]
- 15.Di Sessa TG, Moretti ML, Khoury A, Pulliam DA, Arheart KL, Sibai BM. Cardiac function in fetuses and newborns exposed to low-dose aspirin during pregnancy. Am J Obstet Gynecol 1994; 171:892–900. [DOI] [PubMed] [Google Scholar]
- 16.Litwinska M, Syngelaki A, Wright A, Wright D, Nicolaides KH. Management of pregnancies after combined screening for preeclampsia at 19-24 weeks’ gestation. Ultrasound Obstet Gynecol 2018; 52:365–372. [DOI] [PubMed] [Google Scholar]
- 17.Gallo DM, Wright D, Casanova C, Campanero M, Nicolaides KH. Competing risks model in screening for preeclampsia by maternal factors and biomarkers at 19-24 weeks’ gestation. Am J Obstet Gynecol 2016; 214:619.e1–619.e17. [DOI] [PubMed] [Google Scholar]
- 18.Cnossen JS, Morris RK, ter Riet G, Mol BW, van der Post JA, Coomarasamy A, et al. Use of uterine artery Doppler ultrasonography to predict preeclampsia and intrauterine growth restriction: a systematic review and bivariable meta-analysis. CMAJ 2008; 178:701–711. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Shen J, Chen J, Zheng Z, Zheng J, Liu Z, Song J, et al. An innovative artificial intelligence-based app for the diagnosis of gestational diabetes mellitus (GDM-AI): development study. J Med Internet Res 2020; 22:e21573. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Rzeszotarska A, Szczapa-Jagustyn J, Kociecki J. Ophthalmological problems in pregnancy - a review. Ginekol Pol 2020; 91:473–477. [DOI] [PubMed] [Google Scholar]
- 21.Zhao Q, Huang ZJ, Yang S, Pan J, Smith B, Xu B. The utilization of antenatal care among rural-to-urban migrant women in Shanghai: a hospital-based cross-sectional study. BMC Public Health 2012; 12:1012. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Arab M, Entezari M, Ghamary H, Ramezani A, Ashori A, Mowlazadeh A, Yaseri M. Peripapillary retinal nerve fiber layer thickness in preeclampsia and eclampsia. Int Ophthalmol 2018; 38:2289–2294. [DOI] [PubMed] [Google Scholar]
- 23.Ponticelli C, Moroni G. Is preeclampsia a risk for end-stage renal disease? Kidney Int 2019; 96:547–549. [DOI] [PubMed] [Google Scholar]
- 24.Escudero C, Kupka E, Ibanez B, Sandoval H, Troncoso F, Wikstrom AK, et al. Brain vascular dysfunction in mothers and their children exposed to preeclampsia. Hypertension 2023; 80:242–256. [DOI] [PubMed] [Google Scholar]
- 25.Ma Y, Xiong J, Zhu Y, Ge Z, Hua R, Fu M, et al. Deep learning algorithm using fundus photographs for 10-year risk assessment of ischemic cardiovascular diseases in China. Sci Bull (Beijing) 2022; 67:17–20. [DOI] [PubMed] [Google Scholar]
- 26.Ciloglu E, Okcu NT, Dogan NC. Optical coherence tomography angiography findings in preeclampsia. Eye (Lond) 2019; 33:1946–1951. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Zhu Z, Chen Y, Wang W, Wang Y, Hu W, Shang X, et al. Association of retinal age gap with arterial stiffness and incident cardiovascular disease. Stroke 2022; 53:3320–3328. [DOI] [PubMed] [Google Scholar]
- 28.Wang S, Jin K, Lu H, Cheng C, Ye J, Qian D. Human visual system-based fundus image quality assessment of portable fundus camera photographs. IEEE Trans Med Imaging 2016; 35:1046–1055. [DOI] [PubMed] [Google Scholar]
- 29.Zhou LQ, Wu XL, Huang SY, Wu GG, Ye HR, Wei Q, et al. Lymph node metastasis prediction from primary breast cancer US images using deep learning. Radiology 2020; 294:19–28. [DOI] [PubMed] [Google Scholar]
- 30.Xiao W, Huang X, Wang JH, Lin DR, Zhu Y, Chen C, et al. Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective, multicentre study. Lancet Digit Health 2021; 3:e88–e97. [DOI] [PubMed] [Google Scholar]
- 31.Milletarì F, Navab N, Ahmadi S. V-Net: fully convolutional neural networks for volumetric medical image segmentation. 2016 Fourth International Conference on 3D Vision (3DV), 565–571. [Google Scholar]
- 32.Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2018; 2:158–164. [DOI] [PubMed] [Google Scholar]
- 33.Lin D, Xiong J, Liu C, Zhao L, Li Z, Yu S, et al. Application of Comprehensive Artificial intelligence Retinal Expert (CARE) system: a national real-world evidence study. Lancet Digit Health 2021; 3:e486–e495. [DOI] [PubMed] [Google Scholar]
- 34.Zhou BF. Cooperative Meta-Analysis Group of the Working Group on Obesity in C. Predictive values of body mass index and waist circumference for risk factors of certain related diseases in Chinese adults--study on optimal cut-off points of body mass index and waist circumference in Chinese adults. Biomed Environ Sci 2002; 15:83–96. [PubMed] [Google Scholar]
- 35.Xiong JH, Wang B, Zhao X, Chen YZ, He C, Zhang DL. Hypertension classification method and device based on fundus images. Chinese Patent No. CN111047590A, 2020-04-21. [Google Scholar]
- 36.Hu X, Zhao X, He C, Zhang D, inventor. Fundus Image Quality Control Method, Device, Electronic Equipment and Storage Medium. Chinese Patent No. CN113449774A, 2021-09-28. [Google Scholar]
- 37.Brueckmann A, Seeliger C, Schwefer M, Schlembach D, Schleussner E. PP049. The arterio-venous ratio of retinal vessels in the first trimester as a predictor for preeclampsia. Pregnancy Hypertens 2013; 3:84–85. [DOI] [PubMed] [Google Scholar]
- 38.Moufarrej MN, Vorperian SK, Wong RJ, Campos AA, Quaintance CC, Sit RV, et al. Early prediction of preeclampsia in pregnancy with cell-free RNA. Nature 2022; 602:689–694. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Kiziltunc PB, Varli B, Buyuktepe TC, Atilla H. Ocular vascular changes during pregnancy: an optical coherence tomography angiography study. Graefes Arch Clin Exp Ophthalmol 2020; 258:395–401. [DOI] [PubMed] [Google Scholar]
- 40.Ching T, Himmelstein DS, Beaulieu-Jones BK, Kalinin AA, Do BT, Way GP, et al. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface 2018; 15:20170387. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Emin EI, Emin E, Papalois A, Willmott F, Clarke S, Sideris M. Artificial intelligence in obstetrics and gynaecology: is this the way forward? In Vivo 2019; 33:1547–1551. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Davidson L, Boland MR. Towards deep phenotyping pregnancy: a systematic review on artificial intelligence and machine learning methods to improve pregnancy outcomes. Brief Bioinform 2021; 22:1–29. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Ben-Shlomo Y, Spears M, Boustred C, May M, Anderson SG, Benjamin EJ, et al. Aortic pulse wave velocity improves cardiovascular event prediction: an individual participant meta-analysis of prospective observational data from 17,635 subjects. J Am Coll Cardiol 2014; 63:636–646. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Zhang K, Liu X, Xu J, Yuan J, Cai W, Chen T, et al. Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. Nat Biomed Eng 2021; 5:533–545. [DOI] [PubMed] [Google Scholar]
- 45.Bai J, Wan Z, Li P, Chen L, Wang J, Fan Y, et al. Accuracy and feasibility with AI-assisted OCT in retinal disorder community screening. Front Cell Dev Biol 2022; 10:1053483. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Zhang L, Yuan M, An Z, Zhao X, Wu H, Li H, et al. Prediction of hypertension, hyperglycemia and dyslipidemia from retinal fundus photographs via deep learning: a cross-sectional study of chronic diseases in central China. PLoS One 2020; 15:e0233166. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Cunningham FG, Fernandez CO, Hernandez C. Blindness associated with preeclampsia and eclampsia. Am J Obstet Gynecol 1995; 172:1291–1298. [DOI] [PubMed] [Google Scholar]
- 48.Das S, Kuht HJ, De Silva I, Deol SS, Osman L, Burns J, et al. Feasibility and clinical utility of handheld fundus cameras for retinal imaging. Eye (Lond) 2023; 37:274–279. [DOI] [PMC free article] [PubMed] [Google Scholar]