Obstetrical prenatal complication intelligent prediction system and method based on sclera feature and clinical index multi-modal fusion

Abstract

The invention provides an intelligent prediction system and method for prenatal complications of an obstetrical department based on multi-modal fusion of sclera characteristics and clinical indexes, and is applied to the technical field of data processing. The method comprises the following steps: performing preprocessing and fusion calculation on feature information in multi-modal data to generate fusion calculation data; the prenatal complications (such as gestational diabetes mellitus, gestational hypertension, premature rupture of fetal membranes and premature delivery) of the obstetrics department are dynamically predicted, and prediction characteristic parameters are generated; performing abnormal value filtering processing on the cross-modal feature weight, dynamically optimizing the weight in combination with a clinical and image feature correlation matrix, and generating feature weight data; processing the feature weight data to generate feature importance dynamic change data and abnormal features; processing the feature importance dynamic change data based on a preset risk judgment rule and the abnormal features to generate risk judgment data; and processing the risk judgment data and the prediction characteristic parameters to generate risk early warning information and a complication intervention instruction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-modal fusion obstetric antenatal complication intelligent prediction system and method based on sclera features and clinical indicators. BACKGROUND

[0002] The prior art has many defects in the risk assessment of obstetric antenatal complications, specifically as follows: 1. The clinical data and image feature analysis process are independent of each other, for example, fundus blood vessel parameters cannot be analyzed in conjunction with blood sugar indicators, and the early correlation between sclera blood vessel changes and pregnancy-induced hypertension is also ignored. The independent operation of each complication prediction model makes it difficult to fully explore the correlation between pregnancy-induced hypertension and other complications. 2. The eye image analysis only focuses on simple indicators such as blood vessel diameter, such as arteriovenous ratio (AVR), but ignores the correlation between deep features such as sclera texture and microvessel tortuosity and complications. The spatiotemporal correspondence between clinical indicators and image features has not been fully utilized, and existing eye image analysis only targets a single perspective (such as the fundus), failing to utilize complementary information from multiple eye positions (up, down, left, and right 8 perspectives). 3. Static threshold models cannot adapt to different ethnic groups and gestational age data distributions, for example, there are differences in sclera features between Asian and European pregnant women, and static models are difficult to accurately address. Single algorithms (such as pure logistic regression) have limited ability to express non-linear relationships, making it difficult to fully capture complex pathophysiological changes.

[0003] In addition, there are difficulties in the spatiotemporal alignment of clinical data (tabular) and image features (matrix), such as the mismatch in scale, timing, and semantics between clinical indicators (numeric) and sclera images (high-dimensional matrix), and the lack of a unified feature encoding framework. Dynamic modeling of sclera microvessels with gestational age changes is difficult, and there is a non-linear relationship between their circadian rhythm dilation / constriction and the development of complications. Traditional image processing algorithms are not sensitive enough to low-contrast blood vessels, and traditional models cannot automatically learn the non-linear relationship between modalities. The existing system architecture serially processes data management, feature extraction, model inference, and report generation, making it difficult to achieve an end-to-end prediction delay of <5 seconds in a clinic environment.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an obstetric prenatal complication intelligent prediction system and method based on sclera feature and clinical index multi-modal fusion, at least to some extent, overcome the problems existing in the prior art, integrate 8 eye position sclera feature images and 13 clinical indicators, use improved U-Net++ preprocessing, multi-objective dynamic framework prediction, adaptive mechanism to optimize weight, fuse SHAP and Grad-CAM to determine risk, linkage guide to generate intervention instructions, and improve AUC-ROC.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0007] According to one aspect of the present application, an obstetric prenatal complication intelligent prediction method based on multi-modal fusion of sclera features and clinical indicators is provided, comprising: acquiring multi-modal data and clinical basic information, wherein the multi-modal data includes sclera feature images and clinical indicator data, and the clinical basic information is composed of 13 structured clinical features, 22 manual features of sclera images and 1024-dimensional depth features of sclera images; pre-processing and fusion calculation are performed on the feature information in the multi-modal data to generate fusion calculation data, wherein the fusion calculation data is composed of cross-modal feature point sets, feature probability distribution, abnormal feature value filtering results, distribution weight feature values and centroid weight feature values; based on a multi-objective dynamic selection algorithm, obstetric prenatal complications are dynamically predicted, the bottom layer features are shared and the decision boundary is independently optimized, the related complications are dynamically prioritized according to the patient's medical history, and the prediction feature parameters are generated; based on an adaptive weight calculation method, cross-modal feature weight is filtered for abnormal values, the weight is dynamically optimized in combination with the clinical and image feature correlation matrix, and feature weight data is generated; the feature weight data is processed to generate feature importance dynamic change data and abnormal features; based on a preset risk determination rule and the abnormal features, the feature importance dynamic change data is processed to generate risk determination data; the risk determination data and the prediction feature parameters are processed to generate risk warning information and complication intervention instructions.

[0008] In another aspect of the present application, an intelligent prediction device for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators comprises: an acquisition module configured to acquire multi-modal data and clinical basic information, wherein the multi-modal data comprises a sclera feature image and clinical indicator data, and the clinical basic information is composed of 13 structured clinical features, 22 manual features of the sclera image, and 1024-dimensional deep features of the sclera image; a processing module configured to pre-process and fuse the feature information in the multi-modal data to generate fusion calculation data, wherein the fusion calculation data is composed of cross-modal feature point sets, feature probability distributions, abnormal feature value filtering results, distribution weight feature values, and centroid weight feature values; based on a multi-objective dynamic selection algorithm, the obstetric prenatal complications are dynamically predicted, the bottom features are shared and the decision boundaries are independently optimized, the related complications are dynamically prioritized according to the patient's medical history, and prediction feature parameters are generated; based on an adaptive weight calculation method, the cross-modal feature weights are subjected to abnormal value filtering processing, the weights are dynamically optimized in combination with a clinical and image feature correlation matrix, and feature weight data are generated; the feature weight data are processed to generate feature importance dynamic change data and abnormal features; based on a preset risk judgment rule and the abnormal features, the feature importance dynamic change data are processed to generate risk judgment data; and the risk judgment data and the prediction feature parameters are processed to generate risk warning information and complication intervention instructions.

[0009] According to still another aspect of the present application, an electronic device comprises: a first processor; and a memory configured to store executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned intelligent prediction method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators.

[0010] According to still another aspect of the present application, a computer readable storage medium having a computer program stored thereon is provided, wherein the computer program is executed by a second processor to implement the above-mentioned intelligent prediction method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators.

[0011] The application provides an intelligent prediction system and method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators, which integrates 8 eye position sclera feature images and 13 structured clinical indicators. In the preprocessing, the improved U-Net++ is used to segment the sclera to extract features, and after three-level verification and CNN screening, data fusion is realized through softmax weighting and centroid distance adjustment. The prediction adopts a multi-target dynamic framework, integrates the gestational week features by using LSTM, combines Grad-CAM to locate the abnormal area, compares the feature library to generate prediction parameters. The weight optimization introduces an adaptive mechanism, and realizes light weight through quantitative perception training. The risk judgment fuses the SHAP and Grad-CAM results, links the clinical guidelines to generate intervention instructions, and multi-modal fusion improves the AUC-ROC.

[0012] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flow chart of an intelligent prediction method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 2 A schematic diagram of an intelligent prediction system architecture for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 3 A simplified diagram of an intelligent prediction system architecture for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 4 A schematic diagram of a cross-modal feature dynamic fusion based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 5 A schematic diagram of an improved U-Net++ sclera segmentation network based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 6 A schematic diagram of a multi-index aggregation model cross-validation based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 7 A schematic diagram of disease distribution at different gestational stages based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 8 A schematic diagram of sclera extraction from 8 eye position eye image data based on multi-modal fusion of sclera features and clinical indicators provided by an embodiment of the application is shown; Figure 9A schematic diagram of a sclera feature and clinical index multi-modal fusion-based eye image heat map, eye position importance, and sclera manual feature comparison for simultaneously predicting four different complications of an individual pregnant woman is shown according to an embodiment of the present application. FIG. 10 shows a schematic diagram of a multi-complication feature importance analysis across a pregnancy period according to an embodiment of the present application. Figure 11 A structural schematic diagram of a device for intelligent prediction of obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are merely intended to illustrate and explain the present application, and are not intended to limit the present application.

[0015] In one embodiment, the present application also provides an intelligent prediction system and method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators. Figure 1 A flowchart of an intelligent prediction method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators according to an embodiment of the present application is shown schematically. The method is applied to a server and includes: S101, obtaining multi-modal data and clinical basic information.

[0016] In one embodiment, the multi-modal data includes sclera feature images and clinical indicator data, which are specifically as follows, as shown in Figure 8 The sclera feature images are standardized collected through 8 eye positions (up, down, left, right, left up, left down, right up, and right down), covering multi-view information of the sclera (the data includes 8 eye position original images, labeled images, and predicted images). The collection process needs to filter out invalid images such as motion blur, eyelash obstruction, and corneal reflection through intelligent quality screening CNN, to ensure data quality.

[0017] As shown in Figure 2 The clinical indicator data includes 13 structured features, specifically including eye image measurement gestational age, delivery gestational age, eye image perception gestational age, age, pre-pregnancy BMI, weight gain, weight gain rate, eye image measurement gestational period, single or multiple births, assisted reproduction or not, PCOS (polycystic ovary syndrome), family history of diabetes, and family history of hypertension. The 13 structured clinical features directly correspond to the clinical indicator data, which are used to quantify the basic physiological state and medical history of the pregnant woman.

[0018] 22 hand-crafted features of sclera images, extracted by traditional image processing methods, cover the following dimensions: vascular features such as tortuosity, wall shear stress, microaneurysm count, etc.; morphological features such as area, perimeter, circularity, etc.; texture features such as contrast, energy, homogeneity, etc.; blood flow features such as flow variation, vascular density, etc.; color features: RGB mean and standard deviation (6 items).

[0019] 1024-dimensional deep features of sclera images, extracted by SE-ResNet50 neural network, focus on key sclera regions (such as abnormal vascular regions) with channel attention mechanism, and output high-dimensional deep feature vectors.

[0020] S102, pre-process and fuse the feature information in the multi-modal data to generate fused calculation data.

[0021] In one embodiment, cross-modal feature point sets are extracted from 8 eye position sclera feature images based on an improved U-Net++ segmentation network. As shown in Figure 5 , the improved U-Net++ uses depth separable convolution and multi-scale attention gate structure, fuses shallow edge features and deep semantic features through skip connection, and improves the accuracy of sclera blood vessel boundary segmentation. Edge enhancement loss function: introduce composite loss function , strengthen the gradient consistency of the edge pixels of the blood vessels, and make the segmentation Dice coefficient reach 0.98.

[0022] Vascular connectivity verification: through morphological operation to eliminate noise areas less than 50px, ensure the continuity of the extracted vascular features. An example of 8 eye position sclera image feature extraction is as follows: multi-view acquisition, through 8 eye positions (upper, lower, left, right, upper left, lower left, upper right, lower right) to standardize the acquisition of sclera images, for example, the left eye position image focuses on the nasal side of the blood vessel distribution, and the right eye position focuses on the temporal side of the blood vessel morphology.

[0023] As shown in Figure 2 , cross-modal feature point set generation: 22 hand-crafted features such as tortuosity, area, and perimeter are extracted from the left eye position image; 1024-dimensional deep features are extracted by SE-ResNet50 network to capture microvascular texture and collagen fiber changes.

[0024] The cross-modal feature point set is calculated to generate a feature probability distribution. The cross-modal feature point set (such as 8 eye position vascular density, RGB mean value) is histogram counted, and the formula is where k is a feature value, is the number of occurrences, is the total number of pixels. The probability distribution of 8 eye position vascular density is calculated, and it is found that the probability of late pregnancy vascular density concentrated in the interval of 0.2-0.4 reaches 73%, which is negatively correlated with the risk of premature birth.

[0025] The feature probability distribution is filtered for abnormal values, combined with the three-level verification of clinical data and the intelligent quality screening CNN filtering of eye image data to generate abnormal feature value filtering results. The double verification mechanism of abnormal value filtering is as follows, three-level verification of clinical data: logical rule check, eliminating abnormal clinical records with BMI>40 or gestational age<12 weeks. NLP semantic verification, checking the consistency of medical history description through natural language processing, such as the relevance of “family history of diabetes” and blood glucose monitoring data. Expert review, marking the samples that fail the verification (such as sudden blood pressure rise without clinical explanation) as pending review. Eye image data quality screening, using CNN model to automatically filter images with motion blur and eyelash obstruction, with a filtering efficiency of 92%.

[0026] As shown in Figure 4 , the filtered feature data is weighted and calculated to generate a distribution weight feature value, and the calculation formula for generating the distribution weight feature value is feature, is a deep feature. The input quality sensitive mechanism is as follows: if the eye image is blurred (determined by PSNR<25), automatically reduce to 40% of the original value; if the clinical data lacks family history, adaptively reduce by 20%.

[0027] The feature value is weighted and adjusted based on the distance of the pixel point to the target centroid to generate a centroid weight feature value, and the distance calculation data is composed of the cross-modal feature point set, the feature probability distribution, the abnormal feature value filtering result, the distribution weight feature value and the centroid weight feature value. Taking the complication key feature centroid as the reference (such as the vascular density centroid corresponding to GDM), the pixel point weight is calculated by Euclidean distance: where is the distance of the pixel point to the centroid, is a scale parameter. In the left lower eye position image of Figure 9 , the pixel points with a distance of <50px to the vascular centroid have a vascular density feature weight increased by 30%, strengthening the local features related to gestational hypertension.

[0028] 8 Eye position image covers the whole sclera blood vessels, combined with clinical history to realize "structure-function" joint analysis, and the single mode prediction AUC-ROC is improved. The weight adaptive mechanism makes the model maintain the prediction stability when the image quality fluctuates, and the end-to-end delay is controlled within 3 seconds. The centroid weight calculation provides spatial explanation basis for Grad-CAM heat map, such as the weight proportion of the nasal side blood vessel area in GDM risk prediction is 41%.

[0029] S103, based on the multi-objective dynamic selection algorithm, dynamically predicts obstetric antenatal complications, shares the bottom layer features and independently optimizes the decision boundary, dynamically prioritizes the prediction of related complications according to the patient's medical history, and generates predicted feature parameters.

[0030] In an embodiment, based on the multi-objective dynamic selection algorithm, a cross-modal dynamic weighted fusion technology is introduced, combined with the gestational age time sequence feature, the clinical features and sclera image features in the cross-modal feature point set and feature probability distribution data generated by preprocessing and fusion calculation are extracted to generate a multi-dimensional feature vector set containing gestational age dynamic weights. As shown in Figure 3 , the architecture of shared bottom layer features + independent task head is adopted to support the joint prediction of GDM, GH, PROM, PTB and other complications. For example, diabetic patients automatically activate the GDM / PTB double task head, sharing 1024-dimensional deep features extracted by SE-ResNet50.

[0031] Gestational age time sequence feature integration, through LSTM network to capture the dynamic change of gestational age, such as the correlation between the decline rate of blood vessel density in late pregnancy (>36 weeks) and PTB. The weight formula is: , where t is the gestational age, when t=28 weeks, , the weight of late pregnancy features is strengthened (Figure 10 gestational age trend line). An example of generating cross-modal feature vectors is as follows: data input: clinical features (BMI=26.5, family history of diabetes=1) and sclera image features (temporal blood vessel tortuosity=1.8, blood vessel density=0.32) of a patient at 28 weeks of gestation. Weighted extraction: according to the gestational age weight , generate feature vector: .

[0032] The feature vector set of different complications was compared, and the spatial correlation of the sclera abnormal area located by Grad-CAM heat map was determined to determine the change rule of the target feature of the complication and the shared feature at the bottom layer with the case data and the gestational age stage, and to generate the complication correlation feature mapping relationship with the location of the sclera abnormal area. The Grad-CAM technology was used to generate a heat map of the sclera image of the GDM patient, and it was found that the activation value of the temporal blood vessel area (eye position RRight) was the highest (weight 0.37), and the Pearson correlation coefficient between the tortuosity of the blood vessels in this area and the fasting blood glucose was 0.62. Dynamic feature mapping: a complication-feature correlation matrix was established, for example: . As shown in Figure 7 , among the 605 GDM patients analyzed, 78% of the patients with sclera temporal blood vessel tortuosity >1.5 were accompanied by abnormal blood glucose, and this feature was included in the complication correlation mapping relationship.

[0033] The feature mapping relationship related to the complications was compared with the preset normal pregnancy feature library and the complication risk feature library, and the feature deviation evaluation result was generated by 10-fold cross-validation. As shown in Figure 6 , the preset feature library includes a normal pregnancy library and a complication risk library. Specifically, the normal pregnancy library is based on the feature mean ± standard deviation of 1064 normal pregnant women; the complication risk library includes the feature threshold of GDM, PTB and other cases (such as blood vessel density >0.45 as high risk). The process of deviation calculation is as follows: the blood vessel density of the GDM patient is 0.52, the mean value of the normal library is 0.31, and the standard deviation is 0.09. The deviation , which is determined as high risk.

[0034] Combined with the family history of diabetes in the patient's medical history, the clinical data of assisted reproduction and the quality of the sclera image, the feature deviation evaluation result was weighted by using the eye position heat map weight distribution, and the complication prediction risk feature vector fused with the target feature contribution degree was generated. The process of eye position heat map weighting and risk vector generation is as follows: the weight adjustment rule is that the blood vessel feature weight of the left eye position (LLeft) of the patient with a family history of diabetes is increased by 40%, and the formula is: , wherein is the family history of diabetes (1 / 0). The risk vector is as follows: after integrating the deviation (2.33) and the eye position weight, the vector RiskVec=[2.3 1.4,0.82,0.67] is generated.

[0035] The feature deviation degree evaluation result and the complication prediction risk feature vector are normalized and synthesized, the real-time prediction efficiency of the model after light weight optimization is combined, and a prediction feature parameter set containing feature correlation coefficient, risk probability value and clinical warning line is generated. The FP32 model is converted to INT8, the model size is reduced from 1.2GB to 480MB, and the inference delay is reduced from 5.6 seconds to 2.8 seconds. Normalization and synthesis operation: Z-Score normalization of deviation degree, combined with logistic regression to generate risk probability: . Wherein d is the deviation degree, is the weighted coefficient.

[0036] The clinical warning line and the prediction report generation dynamic warning line process is as follows: according to the gestational age, the risk growth threshold is automatically labeled, such as the risk growth >5% / week after 36 weeks triggers an early warning (GDM trend line in FIG. 10). Prediction parameter output: generate a report containing feature correlation coefficient (such as blood glucose-blood vessel tortuosity=0.72), risk probability (0.82, 95% CI 0.79-0.85) and warning line (36 weeks).

[0037] Multi-modal dynamic fusion improves AUC-ROC over single-modal prediction, especially when image quality is poor (β down-regulated by 60%) while maintaining prediction stability. Through Grad-CAM heat map and eye position weight distribution, the "black box" prediction is converted into clinically understandable feature correlation (such as "temporal blood vessel tortuosity Real-time clinical decision support, model light weight supports outpatient real-time prediction, from image acquisition to report generation <3 seconds, meets the rapid screening needs of prenatal examination.

[0038] S104, based on the adaptive weight calculation method, the cross-modal feature weight is filtered and processed, the weight is dynamically optimized combined with the correlation matrix of clinical and image features, and the feature weight data is generated.

[0039] In one embodiment, based on the adaptive weight calculation method, the dynamic optimization technology of clinical and image feature correlation matrix is introduced, combined with the gestational age time series feature, the cross-modal feature weight is filtered and processed, and a preliminary weight matrix is generated. Dynamic correlation analysis, the spatiotemporal correlation matrix of clinical features (such as BMI, gestational age) and sclera features (such as blood vessel density, tortuosity) is calculated by Pearson correlation coefficient. For example, the correlation coefficient of BMI and sclera nasal blood vessel density at 32 weeks of pregnancy is 0.58, which increases to 0.72 at 36 weeks of pregnancy. Gestational age time series feature integration, introduce LSTM network to capture the dynamic change of weight with gestational age, the formula is: , wherein t is the gestational age, is the basic correlation coefficient, which reflects the increasing trend of correlation from the second trimester to the late trimester.

[0040] The outlier determination rule is as follows: the weight outliers are filtered by the 3σ principle, and the weight of the patient with 32 weeks of pregnancy is marked as abnormal when the weight of BMI and blood vessel density deviates from the mean value by 2.5σ. For a certain GDM patient, the initial value of the clinical BMI weight is 0.35, and after the outlier filtering, it is adjusted to 0.28, and the preliminary weight matrix is generated: Combined with the spatial correlation of the abnormal sclera region located by Grad-CAM heat map, the spatiotemporal correlation matrix of clinical indicators and sclera features is compared to determine the dynamic change law of feature weight with case data and gestational age stage, and a dynamic weight adjustment mapping relationship with sclera abnormal region positioning is generated. Sclera abnormal region positioning: using Grad-CAM technology to generate a heat map of the sclera image of a GDM patient, it is found that the activation value of the temporal blood vessel area (eye position RRight) is the highest (weight 0.41), and the spatiotemporal correlation coefficient between the blood glucose and the tortuosity of the blood vessels in this area reaches 0.67 in the third trimester of pregnancy. Dynamic weight adjustment model, establish the mapping function of weight change with gestational age, for example: . Wherein is the weight of 24 weeks of pregnancy, t is the current gestational age, which reflects the increasing trend of weight in the third trimester of pregnancy.

[0041] The dynamic correlation of blood glucose-blood vessel weight is as follows: after analyzing 331 GH patients, it is found that the weight of blood pressure and upper scleral blood vessel diameter increases by 2.3% per week after 30 weeks of pregnancy, and this law is included in the dynamic weight mapping relationship. Compare the dynamic weight adjustment mapping relationship with the preset normal weight distribution library and the complication risk weight library, and generate weight deviation degree evaluation results combined with 10-fold cross-validation. The preset weight library includes normal pregnancy library and complication risk library, wherein the normal pregnancy library is based on the weight mean ± standard deviation of 1064 normal pregnant women; the complication risk library includes the weight threshold of GDM, PTB and other cases (such as blood glucose-blood vessel weight>0.45 as high risk).

[0042] The deviation degree is calculated as follows: the blood glucose-blood vessel weight of a certain GDM patient is 0.53, the mean value of the normal library is 0.35, and the standard deviation is 0.09. The deviation degree , combined with the AUC-ROC=0.93 of 10-fold cross-validation, it is determined that the weight is abnormal.

[0043] Eye position heat map weighting and risk vector construction are as follows: weight adjustment rule: the weight of the blood vessel features of the left eye position (LLeft) of the patient with a family history of diabetes is increased by 30%, and the formula is: , wherein is the family history of diabetes (1 / 0). After integrating the deviation degree (2.0) and the weight of the eye position, the vector is generated: RiskVec=[2.0 1.3,0.53,0.41].

[0044] With the diabetes family history in the patient's medical history, assisted reproductive clinical data and sclera image quality, the eye position heat map weight distribution and TOP3 feature contribution degree analysis are used to evaluate the weight deviation degree, and the feature weight risk vector is generated by weighting the fusion of eye position heat map weight and feature importance. The multi-source data fusion weighting logic is as follows: the medical history such as diabetes family history (binary variable) and assisted reproduction (binary variable) is converted into weight adjustment factor. For example, the clinical feature weight adjustment formula of patients with diabetes family history is: , wherein Diabetes family history and assisted reproduction markers (1 / 0) are respectively. The sclera image quality is evaluated as follows: the image quality is quantified by PSNR (peak signal-to-noise ratio) and SSIM (structural similarity), and when PSNR<25, the manual feature weight is automatically reduced by 60%. Based on the Grad-CAM heat map, the weight distribution of 8 eye positions is determined, such as the weight proportion of the temporal eye position (RRight) of GDM patients reaching 37%, which is 15%-20% higher than that of other eye positions. Feature contribution degree ranking: TOP3 feature contribution degree is calculated by SHAP value, for example, in a certain GDM case: blood glucose (contribution degree 32%); tortuosity of temporal blood vessels (28%); pre-pregnancy BMI (22%) weight adjustment increases the weight of the first three features by 10%.

[0045] The deviation degree weighting process is as follows: the blood glucose-blood vessel weight deviation degree of a GDM patient is , the normal library mean is 0.35, and the standard deviation is 0.09. The weight of the history and eye position is superimposed: the diabetes family history adjustment factor 1.3; the weight of the left eye position (LLeft) is increased eye position factor 1.3; the image quality is good without reducing the final deviation degree: . After integrating the deviation degree and the feature contribution degree, the risk vector is generated: RiskVec=[3.04,0.32,0.28,0.22,0.37] corresponding to the elements of the adjusted deviation degree, TOP3 feature contribution degree, and the highest eye position weight.

[0046] The weight deviation degree evaluation result and the feature weight risk vector are normalized and integrated, the real-time calculation efficiency is combined with the model lightweight and the quantitative perception training optimization, and the feature weight data containing dynamic weight coefficient, feature correlation coefficient and gestational age weight factor are generated. Min-Max normalization: the deviation degree and the risk vector elements are mapped to the [0,1] interval, such as the deviation degree is normalized to 0.92 (assuming the maximum deviation degree is . The integrated operation formula is: wherein​ SHAP is a feature contribution vector, and EyeWeight is an eye position weight vector.

[0047] The implementation process of model lightweight and real-time optimization is shown below. Quantitative perception training converts FP32 weight matrix to INT8. Pseudo-quantization simulates quantization error to ensure accuracy loss <3%. For example: Original weight matrix elements: 0.65, 0.71, 1.3, INT8 quantization: 104, 114, 208 (corresponding to 0.65x128, 0.71x128, 1.3x128).

[0048] The real-time calculation efficiency is as follows. After lightweight, the weight calculation delay is reduced from 4.1 seconds to 1.7 seconds, supporting 8 patient data per second. The dynamic weight coefficient is as follows, = 0.65 (clinical feature weight): up-regulated by 30% due to family history of diabetes; (hand feature weight): maintain the original value for good image quality; (gestational age weight factor): automatically activate the late pregnancy coefficient at 32 weeks of gestation. The correlation coefficient between blood glucose and temporal vessel tortuosity is 0.71, calculated by Spearman correlation coefficient. After integrating the weight data into the electronic medical record system, a visual report is generated: risk prompt: "GDM risk probability 0.85 (0.82-0.88), due to family history of diabetes and abnormal temporal vessel weight up-regulation by 30%". Intervention suggestion: "Pay attention to blood glucose monitoring after 32 weeks of gestation, corresponding to scleral vessel feature weight γ = 1.3, which needs to be strengthened follow-up".

[0049] S105, processing feature weight data to generate feature importance dynamic change data and abnormal features.

[0050] In one embodiment, the feature weight data is processed based on SHAP value and Grad-CAM heat map to generate feature importance dynamic change data. SHAP value is used to analyze feature weight data (such as clinical feature weight , scleral hand feature weight ), calculate the contribution of each feature to the prediction result, and generate feature importance dynamic change curve. For example, generate blood glucose feature importance change curve from 28 weeks to 36 weeks of gestation (from 0.28 to 0.41), corresponding Figure 9 feature importance TOP3 column chart.

[0051] The feature weight data is processed in combination with the preset abnormal threshold and spatiotemporal correlation to generate abnormal features. By comparing the feature weight data with the preset abnormal threshold (such as weight change rate > 20%), in combination with the Grad-CAM heat map (e.g. Figure 9) positioning the abnormal region of the sclera, identifying abnormal features. For example, it is found that the weight of the temporal blood vessels of the sclera increases by 25% at 30 weeks of pregnancy, exceeding the threshold and corresponding to the high activation area of the heat map, generating an abnormal feature {temporal blood vessel weight} Figure 9 color mutation region of the heat map).

[0052] In S106, the feature importance dynamic change data is processed based on the preset risk judgment rule and the abnormal feature to generate risk judgment data.

[0053] In one embodiment, based on the preset risk judgment rule, the spatial correlation of the abnormal region of the sclera positioned by Grad-CAM heat map is introduced, the abnormal features of the clinical features and the sclera image features in the feature importance dynamic change data are identified, and an abnormal feature set is generated. The preset rule covers the threshold of clinical features (such as BMI>30, blood glucose>5.1mmol / L), the mutation of sclera features (such as the change rate of blood vessel density>15%), and the spatiotemporal correlation rule (such as the correlation between blood glucose and blood vessel tortuosity after 28 weeks of pregnancy<0.3). The rule is stored in the knowledge base ( Figure 9 the "feature importance analysis" module on the right). The Grad-CAM generates a heat map of the abnormal region of the sclera ( Figure 9 ), for example, when the activation value of the temporal blood vessel area (RRight eye position) of the GDM patient is greater than 0.7, it is marked as an abnormal region, and the threshold of the feature importance mutation of the blood vessel tortuosity is set to +20%.

[0054] The abnormal feature identification process is as follows: the data input is the clinical features (blood glucose 5.8mmol / L) and the sclera features (temporal blood vessel density 0.45, change rate 22%) of a patient at 30 weeks of pregnancy. Grad-CAM locates the temporal blood vessel area as an abnormal region (heat map color depth>0.7); the feature importance of blood vessel density increases by 25% compared with the previous week, exceeding the preset threshold of 15%. The abnormal feature set is generated as {blood glucose, temporal blood vessel density, change rate of blood vessel density}( Figure 9 abnormal region annotation box). By comparing the feature importance dynamic change data with the preset normal feature importance distribution library and the complication risk feature importance library, and combining 10-fold cross-validation, the feature importance deviation degree is determined, and the feature deviation degree evaluation result is generated. The preset feature importance library: normal pregnancy library: based on the mean±standard deviation of the feature importance of 1064 normal pregnant women, such as the mean of blood glucose feature importance 0.28, the standard deviation 0.05; and the complication risk library: the threshold of blood glucose feature importance in GDM cases is greater than 0.35. The 10-fold cross-validation process: divide the data set into 10 parts, use 9 parts to train the feature importance distribution model each time; calculate the deviation degree of the feature importance of the test set from the library mean, the formula is: where x is the measured value, the library mean and standard deviation.

[0055] GDM feature deviation degree The importance of blood glucose characteristics of a GDM patient is 0.39, the mean of the normal library is 0.28, the standard deviation is 0.05, and the deviation degree is: ; The AUC-ROC of 10-fold cross-validation is 0.93, which is determined as high deviation. The feature deviation degree evaluation result is associated with the abnormal feature set, and the clinical data such as family history of diabetes and assisted reproduction in the patient's medical history and the image quality of the sclera are combined to weight the feature deviation degree evaluation result, and a risk feature vector fused with eye position heat map weight is generated. The intersection feature (such as blood glucose and blood vessel density) weight of the abnormal feature set and the deviation degree is improved; the history factor (diabetes family history = 1) makes the related feature weight up-regulated by 25%; the image quality (PSNR = 22) makes the sclera feature weight down-regulated by 40%.

[0056] Eye position heat map weight integration: Based on the Grad-CAM result, the weight proportion of the temporal eye position (RRight) of the GDM patient is 35%, which is 15% higher than that of other eye positions, and the formula is: The risk feature vector construction example is as follows: integrate the following parameters to generate the vector: adjusted deviation degree: (diabetes family history up-regulated by 25%, image quality down-regulated by 40%); eye position weight: RRight eye position weight 0.35, left eye position 0.20; TOP3 feature contribution: blood glucose (0.39), blood vessel density (0.32), and pre-pregnancy BMI (0.21). The final vector is [1.65, 0.35, 0.39, 0.32, 0.21].

[0057] The feature deviation degree evaluation result and the risk feature vector are normalized and comprehensively operated, combined with the real-time calculation efficiency of the model after lightweight optimization, to generate a risk judgment data set containing risk probability value, feature deviation coefficient and clinical warning line. Multi-dimensional normalization, the specific process is as follows, deviation degree: normalized to [0, 1] by Min-Max normalization, (set the maximum deviation ); feature contribution: L2 normalized to a unit vector. Comprehensive risk probability calculation: , wherein is the normalized deviation degree, is the risk vector, and W is the weight coefficient.

[0058] The risk judgment result is as follows: risk probability: 0.88 (95% CI 0.85-0.91); feature deviation coefficient: 1.65 (corresponding to the adjusted value of 2.2σ); clinical warning line: labeled 36 weeks (risk increase > 5% / week in the third trimester, the inflection point of the GDM trend line in Figure 10). Model lightweight Afterwards, the risk calculation latency was reduced from 3.2 seconds to 1.5 seconds, and batch processing was supported.

[0059] S107 processes the risk assessment data and predictive characteristic parameters to generate risk warning information and complication intervention instructions.

[0060] In one implementation, the risk assessment data undergoes risk level analysis, combining the AUC-ROC index obtained through 10-fold cross-validation with dynamic gestational age characteristics to generate a risk level assessment result containing risk probability values. Combining the AUC-ROC index obtained through 10-fold cross-validation (e.g., GDM predicted AUC = 0.93 ± 0.02) with dynamic gestational age characteristics (e.g., the rate of decrease in vascular density in late pregnancy), a logistic regression model maps the risk assessment data to risk probability values. The formula is: , where d is the feature deviation coefficient, t is the gestational age, and a / b / c are model parameters.

[0061] Gestational Week Dynamic Feature Embedding: An LSTM network is used to capture the gestational week-risk relationship. For example, when the risk probability increases by more than 5% per week after 28 weeks of gestation, the risk level is automatically upgraded (marked by the gestational week trend line in Figure 10). Specifically, the feature deviation coefficient d = 2.1 for a patient at 32 weeks of gestation. AUC-ROC=0.91, gestational age dynamic factor t=32; P=1 / (1+exp(-(0.6)) 2.1 + 0.3 32+0.1 0.91)))=0.88; the risk level assessment result is "high risk", with a probability value of 0.88 (95% CI 0.85-0.91).

[0062] Based on the feature deviation coefficients in the risk assessment data and the preset risk thresholds, abnormal features are compared and processed. Combined with the spatial correlation of scleral abnormalities located by clinical warning lines and Grad-CAM thermal mapping, complication risk assessment results are generated. The feature deviation coefficients are then compared with the preset risk thresholds (e.g., d>2). Comparison, combined with clinical warning lines (such as 36 weeks) and Grad-CAM thermograms ( Figure 9 The system locates abnormal scleral regions (such as the temporal vascular region activation value > 0.7 in GDM patients) and generates a binary judgment result.

[0063] The spatial-clinical linkage rule is as follows: if the deviation coefficient d > 2 Furthermore, if the overlap between the abnormal area and the high-activation area on the heatmap is >50%, it is judged as "having a risk of complications." The patient characteristic deviation coefficient d = 2.3. Exceeding the threshold 2 ; Grad-CAM shows that the activation value of the upper sclera blood vessel area is 0.81, and the coincidence degree with the high blood pressure related feature space is 65%; the complication risk determination result is "high risk of gestational hypertension", which corresponds to Figure 9 The middle red thermal region is marked. The risk level evaluation result and the complication risk determination result are processed, the sclera abnormal region located by the Grad-CAM heat map and the top three feature importance are fused, and the risk warning information containing the feature importance and the eye position heat map weight is generated. The feature importance is sorted: the top 3 features (such as blood glucose, temporal blood vessel tortuosity, pre-pregnancy BMI) are calculated by SHAP value, combined with the abnormal region located by the Grad-CAM heat map (Fig. 9), and visual warning information is generated. The eye position heat map weight is integrated, and the weight distribution of 8 eye positions (such as the weight of temporal eye position in GDM patients is 35%) is weighted and fused with the feature importance, and the formula is: Figure 9 . .

[0064] The warning information is as follows: "high risk of GDM (probability 0.88), key features: blood glucose (contribution 38%), temporal blood vessel tortuosity (32%), pre-pregnancy BMI (20%)"; the eye position heat map shows that the weight of the temporal eye position (RRight) is the highest (35%, heat map color depth); the 36-week warning line in the correlation diagram 10 prompts "risk increase speed exceeds threshold in late pregnancy". Figure 9

[0065] Real-time gestational age data and clinical guideline standard intervention process information of the patient are obtained, and a mapping relationship between the features and the intervention measures is established based on a clinical semantic mapper. The feature-intervention mapping rule is as follows: the AI features are mapped to the clinical intervention measures through a pre-defined dictionary, for example: "temporal blood vessel tortuosity "recommendation for ophthalmic consultation"; "blood glucose feature importance "start insulin treatment plan". Real-time gestational age data linkage: dynamically adjust the intervention intensity according to the gestational age, such as lowering the blood glucose control threshold from 5.1 mmol / L to 4.8 mmol / L after 34 weeks of pregnancy (Fig. 10 gestational age-intervention threshold curve).

[0066] Based on the real-time gestational age data and the risk warning information, the intervention process information is processed, the real-time calculation efficiency and the dynamic weight coefficient of the model after light weight optimization are combined, and the complication intervention instructions containing the gestational age weight factor and the multi-modal feature weight are generated. The intervention instruction generation example is as follows: the input data is the real-time gestational age of 34 weeks, the risk warning information shows "blood glucose feature importance 0.41, temporal blood vessel tortuosity 1.9"; the clinical semantic mapping is as follows, "monitor blood glucose 4 times a day", "ophthalmic ultrasound examination"; the gestational age weight factor =1.2 (intensive intervention in late pregnancy), the multi-modal feature weight​ (Clinical feature dominant); intervention instruction is "daily blood glucose monitoring + ophthalmic consultation, every 2 weeks review sclera image".

[0067] In the present application, through multi-dimensional data fusion and dynamic algorithm optimization, precise risk early warning and intervention are realized. At the data level, the system integrates 8 eye position sclera feature images (including 22 manual features and 1024-dimensional deep features) and 13 structured clinical indicators (such as gestational age, BMI, family history, etc.). In the preprocessing stage, the improved U-Net++ network is used to extract cross-modal feature points, combined with three-level verification and CNN quality screening to filter abnormal values, and the distribution weight is generated by the softmax function weighting clinical features (f_clin), manual features (f_hand) and deep features (f_deep), and the centroid weight is adjusted based on the pixel to centroid distance, realizing the preliminary fusion of multi-source data.

[0068] The prediction algorithm adopts a multi-objective dynamic selection framework, sharing the SE-ResNet50 bottom features, integrating the gestational age time series features (such as the dynamic weight formula of γ(t)) through LSTM, locating the sclera abnormal area (such as the association of temporal blood vessels and GDM) combined with Grad-CAM heat map, comparing the deviation degree evaluation of normal pregnancy and complication feature library, and finally outputting the prediction parameters containing risk probability and clinical warning line. The weight optimization module introduces an adaptive mechanism, constructs a clinical-image feature dynamic matrix through the Pearson correlation coefficient, adjusts the weight combined with the patient's medical history (such as family history of diabetes) and eye position heat map, realizes model lightweight through quantitative perception training, and the end-to-end delay is controlled within 3 seconds.

[0069] In the risk judgment stage, the SHAP value and Grad-CAM result are fused to generate a dynamic curve of feature importance, and the abnormal features (weight change rate > 20%) are identified through preset rules, and finally the intervention instruction (such as blood glucose monitoring, ophthalmic consultation) is generated by linking the clinical guidelines. This scheme improves the AUC-ROC through multi-modal fusion compared with single modal, and provides an intelligent solution for early screening and precise intervention of prenatal complications.

[0070] In one embodiment, as shown in Figure 11 The present application also provides a kind of obstetric prenatal complication intelligent prediction device based on sclera feature and clinical index multi-modal fusion, comprising: The acquisition module 1101 is used to acquire multi-modal data and clinical basic information, wherein the multi-modal data includes sclera feature image and clinical index data, and the clinical basic information is composed of 13 structured clinical features, 22 sclera image manual features and 1024-dimensional sclera image deep features; The processing module 1102 is configured to preprocess and fusion calculate feature information in the multi-modal data to generate fusion calculation data, wherein the fusion calculation data is composed of cross-modal feature point sets, feature probability distribution, abnormal feature value filtering result, distribution weight feature value and centroid weight feature value; dynamically predict obstetric antenatal complications based on a multi-objective dynamic selection algorithm, share bottom layer features and independently optimize decision boundaries, dynamically prioritize related complications according to patient medical history to generate prediction feature parameters; perform abnormal value filtering processing on cross-modal feature weights based on an adaptive weight calculation method, dynamically optimize weights in combination with clinical and image feature correlation matrix to generate feature weight data; process the feature weight data to generate feature importance dynamic change data and abnormal features; process the feature importance dynamic change data based on a preset risk judgment rule and the abnormal features to generate risk judgment data; and process the risk judgment data and the prediction feature parameters to generate risk warning information and complication intervention instructions.

[0071] The computer readable storage medium provided by the above embodiments of the present application and the obstetric antenatal complication intelligent prediction method based on multi-modal fusion of sclera features and clinical indicators provided by the embodiments of the present application have the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0072] Each of the embodiments in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments. In particular, the evaluation of the obstetric antenatal complication intelligent prediction method based on multi-modal fusion of sclera features and clinical indicators, the electronic device, the electronic equipment and the readable storage medium embodiments are basically similar to the above-mentioned embodiments of the obstetric antenatal complication intelligent prediction method based on multi-modal fusion of sclera features and clinical indicators, so the description is relatively simple, and the related parts can be referred to the above-mentioned embodiments of the obstetric antenatal complication intelligent prediction method based on multi-modal fusion of sclera features and clinical indicators.

Claims

1. An intelligent prediction method for obstetric prenatal complications based on multi-modal fusion of sclera features and clinical indicators, characterized in that, The method comprises the following steps: acquiring multi-modal data and clinical basic information, wherein the multi-modal data comprises sclera feature images and clinical index data, and the clinical basic information is composed of 13 structured clinical features, 22 manual features of the sclera images, and 1024-dimensional deep features of the sclera images; preprocessing and fusion calculating feature information in the multi-modal data to generate fusion calculation data, wherein the fusion calculation data is composed of cross-modal feature point sets, feature probability distributions, abnormal feature value filtering results, distribution weight feature values, and centroid weight feature values; dynamically predicting obstetric antenatal complications based on a multi-objective dynamic selection algorithm, sharing bottom layer features and independently optimizing decision boundaries, dynamically prioritizing related complications according to patient medical history, and generating prediction feature parameters; performing abnormal value filtering processing on cross-modal feature weights based on an adaptive weight calculation method, dynamically optimizing weights in combination with clinical and image feature correlation matrices, and generating feature weight data; processing the feature weight data to generate feature importance dynamic change data and abnormal features; processing the feature importance dynamic change data based on preset risk judgment rules and abnormal features to generate risk judgment data; processing the risk judgment data and the prediction feature parameters to generate risk warning information and complication intervention instructions.

2. The method of claim 1, wherein, The method comprises the following steps of preprocessing and fusion calculating feature information in the multi-modal data to generate fusion calculation data: extracting cross-modal feature point sets from 8 eye position sclera feature images based on an improved U-Net++ segmentation network; calculating the cross-modal feature point sets to generate feature probability distributions; performing abnormal value filtering processing on the feature probability distributions in combination with three-level verification of clinical data and intelligent quality screening CNN filtering of eye image data to generate abnormal feature value filtering results; performing weighted calculation on the filtered feature data to generate distribution weight feature values, wherein the calculation formula for generating the distribution weight feature values is performing weighted adjustment on the feature values based on the distance between the pixel points and the target centroid to generate centroid weight feature values, wherein the distance calculation data is composed of the cross-modal feature point sets, the feature probability distributions, the abnormal feature value filtering results, the distribution weight feature values, and the centroid weight feature values.

3. The method of claim 1, wherein, The method comprises the following steps of dynamically predicting obstetric antenatal complications based on a multi-objective dynamic selection algorithm, sharing bottom layer features and independently optimizing decision boundaries, dynamically prioritizing related complications according to patient medical history, and generating prediction feature parameters: introducing a cross-modal dynamic weighted fusion technology based on the multi-objective dynamic selection algorithm, combining time sequence features of gestational weeks, performing feature extraction on clinical features and sclera image features in the cross-modal feature point sets and the feature probability distribution data generated by preprocessing and fusion calculation, and generating a multi-dimensional feature vector set containing dynamic weights of gestational weeks; comparing feature vector sets of different complications, combining spatial correlation of sclera abnormal regions positioned by Grad-CAM heat maps, determining the change law of complication target features and bottom layer shared features with case data and gestational week stages, and generating a complication correlation feature mapping relationship with positioning of sclera abnormal regions. The complication correlation feature mapping relationship is compared with a preset normal pregnancy feature library and a complication risk feature library, and a feature deviation degree evaluation result is generated by combining 10-fold cross validation; The feature deviation degree evaluation result is weighted processed by using the eye position heat map weight distribution in combination with the diabetes family history in the patient history, the assisted reproduction clinical data and the sclera image quality, and a complication prediction risk feature vector fused with a target feature contribution degree is generated; The feature deviation degree evaluation result and the complication prediction risk feature vector are normalized and comprehensively operated, a prediction feature parameter set including a feature correlation coefficient, a risk probability value and a clinical warning line is generated by combining the real-time prediction efficiency of the model after lightweight optimization.

4. The method of claim 1, wherein, Based on the adaptive weight calculation method, the cross-modal feature weight is subjected to abnormal value filtering processing, and the feature weight data is generated by combining the dynamic optimization weight of the clinical and image feature correlation matrix, including: Based on the adaptive weight calculation method, the cross-modal feature weight is subjected to abnormal value filtering processing by introducing the dynamic optimization technology of the clinical and image feature correlation matrix in combination with the gestational age time sequence feature, and a preliminary weight matrix is generated; The dynamic change law of the feature weight with the case data and the gestational age stage is determined by comparing the space correlation matrix of the clinical index and the sclera feature in combination with the space correlation of the sclera abnormal region located by the Grad-CAM heat map, and a dynamic weight adjustment mapping relationship with the sclera abnormal region positioning is generated; The dynamic weight adjustment mapping relationship is compared with a preset normal weight distribution library and a complication risk weight library, and a weight deviation degree evaluation result is generated by combining 10-fold cross validation; The weight deviation degree evaluation result is weighted processed by using the eye position heat map weight distribution and the TOP3 feature contribution degree analysis in combination with the diabetes family history in the patient history, the assisted reproduction clinical data and the sclera image quality, and a feature weight risk vector fused with the eye position heat map weight and the feature importance is generated; The weight deviation degree evaluation result and the feature weight risk vector are normalized and comprehensively operated, and the feature weight data including a dynamic weight coefficient, a feature correlation coefficient and a gestational age weight factor is generated by combining the real-time calculation efficiency of the model after lightweight and quantitative perception training optimization.

5. The method of claim 1, wherein, Based on the preset risk judgment rule and the abnormal feature, the feature importance dynamic change data is processed to generate risk judgment data, including: The abnormal feature in the feature importance dynamic change data is identified by introducing the space correlation of the sclera abnormal region located by the Grad-CAM heat map based on the preset risk judgment rule, and an abnormal feature set is generated; The feature importance deviation degree is determined by comparing the feature importance dynamic change data with a preset normal feature importance distribution library and a complication risk feature importance library, and a feature deviation degree evaluation result is generated by combining 10-fold cross validation; The feature deviation degree evaluation result is associated with the abnormal feature set, and the feature deviation degree evaluation result is weighted processed by combining the diabetes family history in the patient history, the assisted reproduction and the sclera image quality, and a risk feature vector fused with the eye position heat map weight is generated. The feature deviation degree evaluation result and the risk feature vector are normalized and comprehensively calculated, a risk judgment data set including a risk probability value, a feature deviation coefficient and a clinical warning line is generated in combination with real-time calculation efficiency of the model after lightweight optimization.

6. The method of claim 5, wherein, The risk judgment data and the prediction feature parameters are processed to generate risk warning information and complication intervention instructions, including: The risk judgment data is processed to generate risk level evaluation results including risk probability values in combination with AUC-ROC indexes of 10-fold cross validation and dynamic features of gestational weeks; Abnormal feature comparison processing is performed on the feature deviation coefficient in the risk judgment data and a preset risk threshold, and complication risk judgment results are generated in combination with clinical warning lines and spatial correlation of the sclera abnormal region positioned by the Grad-CAM heat map; The risk level evaluation results and the complication risk judgment results are processed to generate risk warning information including feature importance and eye position heat map weight by fusing the Grad-CAM heat map positioned abnormal region of the sclera and the top three feature importance. Real-time gestational week data and clinical guideline standard intervention process information of the patient are obtained, and a mapping relationship between features and intervention measures is established based on a clinical semantic mapper; The intervention process information is processed based on the real-time gestational week data and the risk warning information, and complication intervention instructions including gestational week weight factors and multi-modal feature weights are generated in combination with real-time calculation efficiency and dynamic weight coefficients of the model after lightweight optimization.

7. An intelligent prediction device for obstetric antenatal complications based on multi-modal fusion of sclera features and clinical indicators, characterized in that, The device comprises: An acquisition module is configured to acquire multi-modal data and clinical basic information, wherein the multi-modal data includes a sclera feature image and clinical index data, and the clinical basic information is composed of 13 structured clinical features, 22 manual features of the sclera image and 1024-dimensional deep features of the sclera image; A processing module is configured to pre-process and fusion-calculate feature information in the multi-modal data to generate fusion calculation data, wherein the fusion calculation data is composed of cross-modal feature point sets, feature probability distributions, abnormal feature value filtering results, distribution weight feature values and centroid weight feature values; dynamically predict obstetric antenatal complications based on a multi-objective dynamic selection algorithm, share bottom features and independently optimize decision boundaries, dynamically preferentially predict related complications according to patient medical histories, generate prediction feature parameters; filter abnormal values of cross-modal feature weights based on an adaptive weight calculation method, dynamically optimize weights in combination with clinical and image feature correlation matrices, generate feature weight data; process the feature weight data to generate feature importance dynamic change data and abnormal features; process the feature importance dynamic change data based on preset risk judgment rules and the abnormal features to generate risk judgment data; process the risk judgment data and the prediction feature parameters to generate risk warning information and complication intervention instructions.

8. An electronic device, comprising: It comprises: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the executable instructions to perform the method for intelligent prediction of obstetric antenatal complications based on multi-modal fusion of sclera features and clinical indicators according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the second processor to implement the method for intelligent prediction of obstetric antenatal complications based on multi-modal fusion of sclera features and clinical indicators according to any one of claims 1-6.

Images