Pre-eclampsia prediction system based on retina parameters

Abstract

The invention relates to a machine learning and biomedical technology, in particular to a pre-eclampsia prediction system based on retina parameters, which comprises a data acquisition module used for acquiring clinical information data and fundus photo data; the fundus parameter extraction module is used for carrying out retina capillary segmentation on the fundus photo data and then carrying out quantification to obtain fundus parameters; the data set preparation module is used for matching and processing the fundus parameters and the clinical information data according to the identity of the pregnant woman to generate a data set for constructing and developing a prediction model; the prediction model construction and development module is used for constructing and training a prediction model by integrating clinical risk factors, MAP and fundus parameters based on a logistic regression (LR) algorithm according to the data set; and the prediction module is used for carrying out real-time acquisition and online risk prediction on the clinical information data of the pregnant woman and the fundus photo data. By noninvasively integrating the fundus photo, the MAP and the pregnant woman characteristic data, the early efficient prediction of the PE and the newborn bad pregnancy outcome is realized.

Description

Technical Field

[0001] The present invention relates to machine learning and biomedical technology, and in particular to a preeclampsia prediction system based on retinal parameters. Background Art

[0002] Preeclampsia (PE) is one of the most serious complications of pregnancy. It refers to hypertension that develops after 20 weeks of gestation, accompanied by proteinuria or abnormal changes in vital organs such as the heart, lungs, liver, and kidneys, as well as in the digestive, nervous, and hematological systems. PE can lead to a variety of adverse pregnancy outcomes (AO), including multi-organ damage, preterm birth (PTB), and small for gestational age (SGA). It also significantly impacts maternal cardiovascular health and the risk of long-term disease.

[0003] Currently, management strategies following the diagnosis of PE remain limited. Aside from symptomatic treatment and termination of pregnancy, no treatments have been found to affect disease progression. Current approaches to improving clinical outcomes in PE primarily focus on early prevention and intervention. Regarding preventive strategies, low-dose aspirin has been shown to be effective in reducing the incidence of PE, particularly when initiated before 16 weeks of gestation, which can reduce the risk of preeclampsia by 62%. This situation highlights the need for effective early detection methods to accurately identify high-risk individuals and initiate timely intervention. Traditional PE risk assessment relies primarily on clinical risk factors such as maternal characteristics and obstetric history, such as advanced maternal age (≥35 years), obesity (BMI ≥30 kg / m²), chronic hypertension, diabetes, a history of kidney disease, and a previous history of PE. However, this clinical feature-based assessment approach has significant limitations, including limited predictive efficacy, an inability to quantify individual-specific risk, and a lack of objective biological markers.

[0004] In recent years, with the advancement of molecular biology and imaging technology, research on the prediction of PE in early pregnancy has made significant progress. Among them, the "triple test" proposed by the Fetal Medicine Foundation (FMF) is the most representative and influential. This method integrates maternal risk factors with new biomarkers, including gestational age 11-13. +6A multicenter prospective validation study conducted in China demonstrated that a model combining mean arterial pressure (MAP), uterine artery pulsatility index (UtA-PI), placental growth factor (PlGF), and pregnancy-associated plasma protein-A (PAPP-A) for the prediction of preterm PE (PE) achieved a predictive efficacy of up to 0.86 under the AUC. However, these novel biomarkers face numerous challenges in their clinical application. For example, UtA-PI measurement requires specialized ultrasound physicians and is not a routine ultrasound examination. PlGF and PAPP-A testing are invasive serological tests and are costly, particularly in remote areas with limited medical resources. Furthermore, while analytical techniques based on circulating free DNA methylation and free RNA offer new possibilities for the early prediction of PE, their high cost limits their widespread application in clinical practice. Therefore, there is an urgent need to develop more innovative, convenient and non-invasive prediction methods to adapt to universal screening scenarios, increase the early detection rate of PE and improve maternal and infant prognosis, which has important practical significance for improving the PE screening capabilities of primary medical institutions in low- and middle-income countries.

[0005] The eyes, known as "windows to the soul," are not only organs of vision but also unique "observation windows" that reflect overall health. As the only organ in the human body capable of direct, noninvasive observation of neurovascular processes, fundus microvascular changes can reflect the pathophysiological changes of various systemic diseases, offering the advantages of easy acquisition, non-invasiveness, and multimodality. Pregnancy is often accompanied by ocular changes, such as physiological changes caused by increased systemic blood volume and increased fundus perfusion due to hormonal changes. These changes can also reflect underlying pathological processes, including various pregnancy complications. For example, patients with PE may experience fundus vasospasm, hemorrhage, and decreased vascular density, all of which are associated with endothelial dysfunction and ischemic damage caused by hypertension. This ocular information provides important clues to the development and progression of pregnancy-related diseases, suggesting that ophthalmological examination has the potential to serve as a window for assessing and monitoring pregnancy complications. Current obstetric clinical guidelines recommend routine ophthalmological consultation for patients with PE, which can be assessed through a variety of ophthalmological examination methods. Currently, the most commonly used fundus examination method for pregnant women is non-dilated fundus color photography. However, traditional fundus evaluation methods cannot provide quantitative assessments and are subjective. They can only be evaluated when the lesions are macroscopically visible and have a certain lag. They can only provide a basis for the severity and progression of the disease and cannot provide a reasonable reference for disease prediction. Summary of the Invention

[0006] In order to solve the technical problems that existing PE screening methods rely on invasive testing, are costly and difficult to promote in primary medical care, the present invention provides a preeclampsia prediction system based on retinal parameters. By non-invasively integrating fundus photographs (fundus images), MAP and maternal characteristic data, it combines advanced AI technology with clinical medicine to achieve early and efficient prediction of PE and adverse neonatal pregnancy outcomes.

[0007] The technical solution adopted in the embodiment of the present invention is: a preeclampsia prediction system based on retinal parameters, comprising:

[0008] Data collection module, used to collect data of single pregnant women from 11 weeks to 13 weeks of pregnancy +6 Clinical information data and fundus color photography data between weeks;

[0009] The fundus parameter extraction module is used to segment retinal microvessels from fundus photograph data and then quantify fundus parameters after segmentation. Fundus parameters include arteriovenous ratio, vascular bone density, tortuosity, angle-based tortuosity, and angle-based tortuosity standard deviation.

[0010] The dataset preparation module is used to match fundus parameters with clinical information data according to maternal status, perform data cleaning and standardization, and generate a dataset for predictive model construction and development. It is also used to shuffle the samples in the dataset and randomly split the dataset into multiple sub-datasets, with the distribution of preeclampsia cases in each sub-dataset being the same.

[0011] A prediction model construction and development module is used to construct and train a prediction model based on the data set, using a logistic regression (LR) algorithm, and integrating clinical risk factors, MAP, and fundus parameters;

[0012] The prediction module deploys the fundus parameter extraction module and prediction model on a portable terminal device and / or a cloud platform to collect clinical information data and fundus photograph data of pregnant women in real time and perform online risk prediction.

[0013] Preferably, before segmenting the retinal microvessels in the fundus photograph data, the fundus parameter extraction module first evaluates the overall image quality of the fundus photograph data, eliminates unqualified images, and selects qualified left and / or right eye fundus images.

[0014] Further preferably, the fundus parameter extraction module includes an image quality control module, a segmentation module and a measurement module connected to each other;

[0015] During the fundus parameter extraction process, the image quality control module first evaluates the overall image quality of the fundus photograph data. Subsequently, the segmentation module divides the fundus image after quality assessment to generate a segmentation map of arteries, veins, and optic discs. Next, the measurement module calculates the measurement values of the corresponding areas in the segmentation map to extract the fundus parameters.

[0016] Preferably, in the prediction model construction and development module, the prediction model construction process includes:

[0017] The LASSO logistic regression algorithm combined with 10-fold cross-validation was used to select feature variables; the selected feature variables included MAP, overweight, history of kidney disease, history of PE in the clinical information data, and the extracted fundus parameters;

[0018] Based on the selected feature variables, the dataset preparation module randomly splits multiple sub-datasets, builds models based on the logistic regression LR algorithm and 10-fold cross validation, and evaluates model performance;

[0019] The model is trained on all the datasets generated by the dataset preparation module to obtain the final prediction model.

[0020] Further preferably, the selection of feature variables is achieved by tuning the hyperparameters of LASSO logistic regression; wherein, in the process of hyperparameter tuning, the hyperparameter when the average area under the curve AUC on the validation set is the largest is selected as the optimal hyperparameter; when the optimal hyperparameter is taken, the feature whose model parameter is not equal to 0 is selected as the feature variable.

[0021] Preferably, the LASSO logistic regression algorithm adds the L1 norm on the basis of the logistic regression algorithm, and the loss function Loss of the LASSO logistic regression algorithm is:

[0022] Loss = Logistic Regression Loss Function + λ*L1 Norm

[0023] Where λ is a hyperparameter used to characterize the penalty term of the L1 norm.

[0024] The present invention, through the above technical solution, achieves non-invasive, convenient and economical early prediction of PE, providing a scientific and reliable decision-making tool for improving maternal and infant health outcomes. Compared with the existing technology, the technical effects achieved by the present invention include:

[0025] 1. Non-invasive

[0026] This method, based on fundus photographs and maternal clinical data, eliminates the need for invasive serum testing and, compared with biomarker-based prediction methods, avoids blood sampling. Furthermore, the method uses non-mydriatic fundus color photography to collect retinal parameters, reducing the physical burden of the testing process on pregnant women.

[0027] 2. Economical

[0028] By collecting fundus photographs and clinical data of pregnant women, the present invention makes fundus examination cheaper and easier to obtain than serum marker detection and ultrasound technology based on biomarker requirements, effectively reducing detection costs, saving more medical resources, and being more suitable for primary medical environments.

[0029] 3. Accessibility

[0030] This invention combines fundus photographs and clinical data of pregnant women, avoiding the need for serological tests with complex laboratory support or ultrasound examinations performed by highly qualified technicians in biomarker-based prediction technologies. It can complete data collection and risk assessment in primary healthcare scenarios; it reduces the requirements for testing technology and resources, and is suitable for promotion and application in primary healthcare institutions and low- and middle-income countries.

[0031] 4. Early prediction

[0032] This method can effectively predict PE and adverse neonatal outcomes (such as fetal growth restriction and premature birth) before 14 weeks of gestation. The prediction model used combines retinal parameters, mean arterial pressure, and maternal clinical data to provide risk assessment before clinical symptoms appear, providing data support for medical staff to formulate intervention strategies in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of a modeling of a prediction model used by the prediction system in an embodiment of the present invention;

[0034] Figure 2 Schematic diagram of variable importance in the PROMPT model provided by an embodiment of the present invention;

[0035] Figure 3 is a nomogram of the PROMPT model provided by an embodiment of the present invention;

[0036] Figure 4 2 is a schematic diagram of the performance of the PE prediction model adopted in the embodiment of the present invention and the results of subgroup analysis; wherein, (a) is a ROC curve diagram of the PE prediction performance based on the baseline model, eye model, MAP model and PROMPT model, (b) is a lift diagram of the PROMPT model and the MAP model, (c) is a prediction performance curve diagram of the PROMPT model in predicting preterm PE and term PE, and (d) is a subgroup analysis comparison diagram of the PROMPT model in the primipara and multipara groups. DETAILED DESCRIPTION

[0037] The present invention proposes an innovative algorithm based on deep learning (DL) and machine learning (ML). By integrating maternal risk factors in early pregnancy, clinical examination data and fundus characteristics, a PROMPT model (Preeclampsia Risk factors-Ophthalmic data-Mean arterial pressure Prediction Test, PROMPT, preeclampsia risk factors-ophthalmic data-mean arterial pressure prediction test) is constructed to predict PE. Compared with traditional methods, the PROMPT model has the advantages of being non-invasive, convenient and low-cost, and is particularly suitable for promotion and application in primary medical institutions. The present invention aims to develop an objective and reliable early prediction tool for PE, provide a new solution for improving the early identification and management of PE and related adverse pregnancy outcomes, and has important clinical significance for reducing maternal and perinatal morbidity and mortality.

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings in this embodiment. Obviously, this embodiment is only one of the preferred embodiments of the present invention, and any modification, replacement, combination or simplification made without departing from the spirit and principle of the present invention shall be regarded as an equivalent embodiment of the present invention and included in the scope of protection of the present invention.

[0039] Example

[0040] This embodiment provides a preeclampsia prediction system based on retinal parameters, including a data set acquisition module, a data preprocessing module, a fundus parameter extraction module, a data set preparation module, a prediction model construction and development module, and a prediction module. Figure 1 As shown in the figure, each module is described and introduced in detail below.

[0041] 1. Data Acquisition Module

[0042] In this embodiment, the data collection module collects clinical information data and fundus photograph data of the target pregnant women.

[0043] Specifically, during pregnancy 11-13 +6 At the gestational age, clinical information data of pregnant women with singleton pregnancy were collected, including age, gestational age, body mass index (BMI), MAP, history of chronic hypertension, history of kidney disease, and history of PE.

[0044] At the same time, non-mydriatic fundus photography equipment should be used to collect color fundus photographs of pregnant women. The collection time of fundus photographs should be synchronized with the collection of clinical information data.

[0045] It can also be further combined with pregnancy follow-up data to annotate pregnancy outcomes (such as small for gestational age, premature birth, and other serious complications) to provide real labels for subsequent model training.

[0046] 2. Data Preprocessing Module

[0047] The data preprocessing module preprocesses the collected clinical information data and fundus photograph data respectively.

[0048] Clinical information data is first formatted, converted, and encoded to meet the requirements of machine learning algorithms; then, missing and abnormal items are filled with the mode or mean according to preset standards.

[0049] Preprocess fundus photograph data, including rotation, flipping, scaling and resizing, to unify image specifications.

[0050] 3. Fundus Parameter Extraction Module

[0051] In this embodiment, the fundus parameter extraction module includes an image quality control module, a segmentation module, and a measurement module that are connected to each other.

[0052] During fundus parameter extraction, the image quality control module first assesses the overall image quality of the fundus photograph data, eliminating unqualified images due to blurring, underexposure, and other factors that could affect subsequent processing. Finally, qualified fundus images are selected. Generally, the right eye data is included in the analysis. If the right eye fundus image quality is unqualified or missing, qualified left eye fundus images are selected. Subsequently, the segmentation module divides the quality-assessed fundus image into segments, generating a segmentation map of arteries, veins, and the optic disc. Next, the measurement module calculates the measurements of the corresponding regions in the segmentation map and extracts fundus parameters, including the arteriovenous ratio (AVR), vascular skeleton density (VSD), and tortuosity density (tortD).

[0053] Among them, the segmentation module uses a multi-branch U-Net architecture based on deep learning to perform fully automatic segmentation of retinal microvessels in fundus images after quality assessment; after segmentation, the measurement module is used to quantify fundus parameters such as arteriovenous ratio, vascular bone density and tortuosity.

[0054] Please refer to Table 1 for the definitions of the extracted fundus parameters and their clinical significance.

[0055] Table 1. Definition and clinical significance of fundus parameters

[0056]

[0057] 4. Dataset Preparation Module

[0058] In the dataset preparation module, the extracted fundus parameters are first matched with the preprocessed clinical information data according to the identity of the pregnant woman, and the data is cleaned and standardized to generate a dataset for predictive model construction and development.

[0059] Then, the samples of the dataset were shuffled and randomly split into 10 sub-datasets using stratified random sampling to prepare for the subsequent training and validation of the prediction model through manual 10-fold cross validation (CV). Specifically, the entire dataset was randomly divided into 10 sub-datasets of almost equal size, and the distribution of PE cases in each sub-dataset was ensured to be almost the same. During the 10-fold cross validation process, 9 sub-datasets were merged into the training set each time to train the prediction model, and the remaining 1 sub-dataset was used as the validation set to evaluate the prediction model. This process was repeated 10 times.

[0060] 5. Prediction Model Construction and Development Module

[0061] During the construction and development of the prediction model in this embodiment, five machine learning algorithms were tested, including logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and gradient boosting method (GBM).

[0062] Models for five machine learning algorithms were trained and evaluated using manual 10-fold cross-validation. In each 10-fold iteration, a training dataset consisting of nine subsets was randomly sampled from PE-positive samples with replacement until the number of positive samples in the dataset equaled the number of negative samples. The model was then trained using the optimal hyperparameters on the balanced training dataset, while the remaining subset served as the validation dataset for model validation. Model performance on both the training and validation sets was calculated and recorded. Model performance was evaluated using the area under the curve (AUC) and compared with LR. As shown in Table 2, the five models performed comparably on the validation set, with AUCs ranging from 0.858 to 0.867.

[0063] Table 2. Comparison of the effects of different machine learning algorithm models

[0064]

[0065] In this example, the logistic regression (LR) algorithm was selected to construct and develop the prediction model, taking into account the model performance and clinical interpretability.

[0066] Based on the dataset generated by the dataset preparation module and the logistic regression (LR) algorithm, this embodiment constructs four prediction models: a baseline model, an eye model, a MAP (mean arterial pressure) model, and a PROMPT model. The baseline model is constructed based on traditional clinical risk factors, the eye model is constructed using only fundus parameters, the MAP model is constructed based on clinical risk factors and mean arterial pressure, and the PROMPT model is constructed using a combination of clinical risk factors, mean arterial pressure, and fundus parameters.

[0067] During the PROMPT modeling process, the LASSO logistic regression algorithm combined with 10-fold cross-validation was first used to select feature variables. Specifically, the LASSO logistic regression algorithm was used to screen feature variables from the dataset, ultimately identifying nine feature variables: MAP, overweight, history of kidney disease, history of PE, and extracted fundus parameters: arteriovenous retinal ratio (AVR), vascular bone density (VSD), tortuosity (tortD), angle-based tortuosity (angtort), and standard deviation (std) of tortuosity. The LASSO logistic regression algorithm adds the L1 norm (the sum of the absolute values of the regression parameters) to the logistic regression algorithm. Its loss function is: logistic regression loss function + λ*L1 norm, where λ is a hyperparameter representing the penalty term for the L1 norm. The optimization goal of Lasso logistic regression is to minimize the loss function. By adjusting the hyperparameter λ, the model complexity can be controlled. When λ is small, the model tends to fit more feature variables, while when λ is large, the model compresses more parameters to zero, thus optimizing the selection of feature variables. Therefore, the process of feature variable selection is the process of tuning the hyperparameter λ of the LASSO logistic regression, that is, the selection of feature variables is achieved by tuning the hyperparameter λ of the LASSO logistic regression. This embodiment uses random search and 10-fold cross-validation tuning to select the hyperparameter when the average AUC on the 10 validation sets is the largest as the optimal hyperparameter. When the optimal hyperparameter λ is taken, the feature whose model parameter is not equal to 0 is used as the selected feature variable. Considering that non-significant features are not beneficial to the performance of the logistic regression (LR) model, and to ensure the simplicity of the final model, a two-way stepwise regression method is further used for feature screening.

[0068] Then, based on the selected feature variables, 10 sub-datasets randomly split by the dataset preparation module are used to build a model and evaluate model performance based on the logistic regression LR algorithm and 10-fold cross-validation. The model performance is mainly evaluated by AUC, and the comprehensive performance of the model on the 10 validation sets is taken as the model performance.

[0069] In practical applications, the paired DeLong test can be used to evaluate the improvement in model prediction performance after adding fundus parameters; decision cruve analysis (DCA) is used to demonstrate the clinical utility of each model. Lift charts, forest plots, and nomograms are used to intuitively display the model risk score, performance, and importance of characteristic variables. Please refer to Figure 2 、 Figure 3 , Figure 2 The importance of variables in the PROMPT model provided by the embodiment of the present invention is illustrated; Figure 3 Nomogram of the PROMPT model provided for the implementation of the present invention.

[0070] Finally, the model is trained on all the datasets generated by the dataset preparation module to obtain the final prediction model, which will be used for future new data prediction and model interpretation, such as nomograms and variable importance diagrams.

[0071] The variables of the baseline model are based on the high-risk risk factors for PE in the obstetric clinical guidelines and are suitable for scenarios with limited medical resources, including elderly pregnant women (over 35 years old), overweight, primiparas, history of chronic hypertension, history of kidney disease, history of autoimmune disease, history of pre-gestational diabetes, and history of PE. The LR algorithm is used to directly incorporate these high-risk risk factor variables for PE into the training model. The variables of the eye model only include the screened fundus parameters (AVR, VSD, tortD, angtort, and std of angtort), and there is no clinical information data for pregnant women. The MAP model combines the screened maternal risk factors with MAP (overweight, history of kidney disease, history of PE, and MAP). Except for the baseline model, the remaining models are constructed using the same method as the PROMPT model.

[0072] Statistical analysis and performance evaluation of the model were performed as follows: For sample size calculation, assuming a 5% prevalence of PE, a logistic regression model was constructed using the nine selected characteristic variables. Based on the "10 events per variable" principle, 1,800 samples were required for model development. For normally distributed continuous characteristics, data were presented as mean ± standard deviation and compared using the t-test; otherwise, data were presented as median and interquartile range, and compared using the Mann-Whitney U test. Categorical variables were presented as frequencies and percentages, and compared using the chi-square test or Fisher's exact test. The discriminatory ability of the model was assessed using receiver operating characteristic (ROC) curves and area under the curve (AUC) with 95% DeLong confidence intervals (CIs). The paired DeLong test was used to assess the improvement in predicting PE by adding fundus parameters. Decision curve analysis (DCA) was used to demonstrate the net benefit of the model at different risk thresholds to assess its practicality in decision making. For the PROMPT model, lift charts were used to demonstrate its discriminatory performance, and forest plots and nomograms were used to enhance model interpretability. A p value < 0.05 in a two-tailed test was considered statistically significant.

[0073] In this embodiment, the model effects and selections can be: Please refer to Figure 4 , Figure 4 The performance and subgroup analysis results of the PE prediction model provided in the embodiment of the present invention. In this embodiment, the AUCs of the four models, namely the baseline model, the eye model, the MAP model and the PROMPT model, reached 0.69 (0.63-0.76), 0.74 (0.69-0.79), 0.82 (0.78-0.87) and 0.87 (0.83-0.90), respectively. Among them, the AUCs of the PROMPT model in predicting premature PE and full-term PE were 0.91 (0.85-0.97) and 0.84 (0.80-0.89), respectively; the decision curve analysis showed that the PROMPT model performed best over a wide range of risk thresholds. The lift chart shows the discriminant performance of the PROMPT model. Therefore, this embodiment ultimately selected the PROMPT model as the prediction model.

[0074] 6. Prediction Module

[0075] The fundus parameter extraction module and PE prediction model (i.e., PROMPT model) used for retinal microvascular segmentation and quantification of fundus images are deployed on portable terminal devices and / or cloud platforms to realize real-time collection of clinical information data and fundus photograph data of pregnant women and online risk prediction.

[0076] Based on the prediction results of the PROMPT model, intervention measures can be implemented for high-risk pregnant women, such as recommending prophylactic use of aspirin, increasing the frequency of monitoring, or adjusting the delivery plan to reduce the risk of preeclampsia and adverse pregnancy outcomes.

[0077] The prediction system presented in this paper is an objective and reliable AI-based solution for the early prediction of PE, offering a new solution for improving the early identification and management of PE and related adverse pregnancy outcomes. By integrating advanced AI technology with clinical medicine through fundus photographs, MAP, and maternal characteristic data, this prediction system enables efficient early prediction of PE and adverse neonatal pregnancy outcomes. This provides a non-invasive, accessible, and cost-effective approach to pregnancy health management in resource-limited areas, supporting clinical decision-making.

[0078] In the process of predicting PE, the present invention adopts a retinal microvascular health assessment technology based on deep learning, and uses a convolutional neural network architecture to automatically segment and quantify retinal blood vessels to accurately extract multiple key retinal parameters, such as arteriovenous ratio, vascular density, vascular tortuosity, etc. These retinal parameters can quantify the morphological characteristics of retinal blood vessels and are related to the progression of systemic vascular lesions. This AI-assisted vascular quantification method has several significant technical advantages: first, it can process large-scale fundus image data and achieve high-throughput analysis; second, the system has sub-pixel measurement accuracy and can detect subtle vascular changes that are difficult to identify with traditional methods; third, through machine learning algorithms, the system can automatically identify and quantify a variety of vascular characteristics, reducing the subjectivity of human judgment. These technical advantages make the AI-assisted vascular quantification method have important application value in the early prediction of PE.

[0079] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A preeclampsia prediction system based on retinal parameters, characterized in that: include: Data acquisition module, used to collect clinical information data and fundus photograph data of pregnant women; The fundus parameter extraction module is used to segment retinal microvessels from fundus photograph data and then quantify fundus parameters after segmentation. Fundus parameters include arteriovenous ratio, vascular bone density, tortuosity, angle-based tortuosity, and angle-based tortuosity standard deviation. The dataset preparation module is used to match fundus parameters with clinical information data according to maternal status, perform data cleaning and standardization, and generate a dataset for predictive model construction and development. It is also used to shuffle the samples in the dataset and randomly split the dataset into multiple sub-datasets, with the distribution of preeclampsia cases in each sub-dataset being the same. A prediction model construction and development module is used to construct and train a prediction model based on the data set, the logistic regression LR algorithm, and the comprehensive clinical risk factors, MAP and fundus parameters; The prediction module deploys the fundus parameter extraction module and prediction model on a portable terminal device and / or a cloud platform to collect clinical information data and fundus photograph data of pregnant women in real time and perform online risk prediction.

2. The preeclampsia prediction system according to claim 1, wherein: The data collection module also marks the pregnancy outcome in combination with the pregnancy follow-up data.

3. The preeclampsia prediction system according to claim 1, wherein: Before segmenting the retinal microvessels in the fundus photograph data, the fundus parameter extraction module first evaluates the overall image quality of the fundus photograph data, eliminates unqualified images, and selects qualified left and / or right eye fundus images.

4. The preeclampsia prediction system according to claim 3, wherein: The fundus parameter extraction module includes an image quality control module, a segmentation module and a measurement module connected to each other; During the fundus parameter extraction process, the image quality control module first evaluates the overall image quality of the fundus photograph data. Subsequently, the segmentation module divides the fundus image after quality assessment to generate a segmentation map of arteries, veins, and optic discs. Next, the measurement module calculates the measurement values of the corresponding areas in the segmentation map to extract the fundus parameters.

5. The preeclampsia prediction system according to claim 4, wherein: The segmentation module uses a multi-branch U-Net architecture based on deep learning to perform fully automatic segmentation of retinal microvessels in fundus images after quality assessment.

6. The preeclampsia prediction system according to claim 1, wherein: In the prediction model construction and development module, the prediction model construction process includes: The LASSO logistic regression algorithm combined with 10-fold cross-validation was used to select characteristic variables; the selected characteristic variables included MAP, overweight, history of renal disease, history of PE in the clinical information data, and the extracted fundus parameters; Based on the selected feature variables, the dataset preparation module randomly splits multiple sub-datasets, builds models based on the logistic regression LR algorithm and 10-fold cross validation, and evaluates model performance; The model is trained on all the datasets generated by the dataset preparation module to obtain the final prediction model.

7. The preeclampsia prediction system according to claim 6, wherein: Feature variables are selected by tuning the hyperparameters of LASSO logistic regression. During the hyperparameter tuning process, the hyperparameter with the largest average area under the curve (AUC) on the validation set is selected as the optimal hyperparameter. When the optimal hyperparameter is selected, the feature whose model parameter is not equal to 0 is selected as the feature variable.

8. The preeclampsia prediction system according to claim 7, wherein: Random search and 10-fold cross validation were used for hyperparameter tuning.

9. The preeclampsia prediction system according to claim 6, wherein: The LASSO logistic regression algorithm adds the L1 norm to the logistic regression algorithm. The loss function of the LASSO logistic regression algorithm is: Loss = Logistic Regression Loss Function + λ*L1 Norm Where λ is a hyperparameter used to characterize the penalty term of the L1 norm.

10. The preeclampsia prediction system according to claim 1, wherein: Also includes: The data preprocessing module is used to preprocess the collected clinical information data and fundus photograph data respectively.

Images