A method and system for predicting adverse pregnancy outcomes of preeclampsia by fusing placental microenvironment image features and serum biomarkers

Abstract

The present application relates to the technical field of medical data fusion processing, and discloses a method and system for predicting adverse pregnancy outcomes of preeclampsia by fusing placental microenvironment image features and serum biomarkers, wherein the method comprises: obtaining digital whole field image of placental tissue section and serum biomarker detection data at the same period; constructing a multi-modal sample data set; extracting multi-scale structure features of placental microenvironment; constructing serum biomarker discriminant features; dynamically fusing placental image features and serum marker features; constructing a prediction model for adverse pregnancy outcomes of preeclampsia and outputting a prediction probability. Compared with the prior art which only relies on single image information or single serum biomarker for risk assessment, it is difficult to realize stable prediction of the coordinated changes of placental microenvironment abnormalities and biochemical abnormalities. Due to the multi-scale image structure analysis, the model generalization ability for predicting adverse pregnancy outcomes of preeclampsia is improved.

Description

Technical Field

[0001] This invention relates to the field of medical data fusion processing technology, and in particular to a method and system for predicting adverse pregnancy outcomes in preeclampsia by fusing placental microenvironment image features and serum biomarkers. Background Technology

[0002] Currently, preeclampsia, as a serious complication of pregnancy, has a complex pathogenesis involving placental developmental abnormalities, impaired vascular remodeling, endothelial dysfunction, and abnormal expression of various inflammatory factors and biomarkers. Clinically, the assessment of preeclampsia risk mainly relies on maternal blood pressure, urine protein, past medical history, and serum biomarker test results (such as sFlt-1, PlGF, PAPP-A, etc.), or on ultrasound imaging to observe placental morphology and blood flow. However, these techniques are mostly empirical judgments based on single information sources or local indicator analyses, and overall, there is still a lack of systematic modeling capabilities to understand the intrinsic correlation between the microscopic pathological structure of placental tissue and biochemical indicators.

[0003] In placental pathological analysis, the development of digital pathological scanning technology has enabled placental tissue sections to be scanned into full-field digital images at high resolution, providing conditions for analyzing the placental microenvironment using computer vision techniques. Existing studies have attempted to extract features such as placental vascular density, villous structure, and tissue arrangement through image segmentation and texture analysis methods for diagnostic support. However, placental tissue exhibits significant spatial heterogeneity, with pathological changes often manifesting simultaneously at multiple scales—cellular, capillary, and tissue levels. Traditional single-scale image analysis methods struggle to simultaneously capture both macroscopic structure and microscopic details, easily missing crucial discriminative information and resulting in poor stability of analysis results. Furthermore, in serum biomarker detection, different batches, experimental environments, and sample preservation conditions can all introduce systematic biases to the results. Existing technologies typically employ simple normalization or standardization methods to process serum data, failing to adequately consider the overall bias caused by batch effects. This leads to insufficient comparability of the same biomarker across different batches, consequently affecting the stability and reliability of subsequent model training.

[0004] Furthermore, cases of preeclampsia progressing to adverse pregnancy outcomes account for a relatively small percentage of the overall sample, indicating a significant class imbalance in the clinical data. Existing prediction methods based on machine learning or deep learning often focus more on the majority of normal samples during training, neglecting the feature learning of the minority of high-risk samples. This results in weak identification ability of high-risk cases in real-world applications, with predictions biased towards the normal category and limited generalization ability.

[0005] Therefore, there is an urgent need for a technical solution that can still achieve collaborative fusion modeling of placental microenvironment image features and serum biomarker features even when there is significant spatial heterogeneity in placental tissue, batch bias in serum testing, severe imbalance in sample categories, and independent multimodal information, in order to improve the accuracy, stability, and clinical application value of predicting adverse pregnancy outcomes in preeclampsia. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a method for predicting adverse pregnancy outcomes in preeclampsia that integrates placental microenvironment image features and serum biomarkers. This method aims to solve the technical problem that existing technologies rely solely on single image information or single serum biomarkers for risk assessment, making it difficult to achieve stable prediction of synergistic changes in placental microenvironment and biochemical abnormalities.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers.

[0008] The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers includes:

[0009] Step S10: Obtain digital full-view images of placental tissue sections and concurrent serum biomarker detection data; combine batch effect fitting correction and case-level paired annotation mechanism to perform multimodal sample normalization task; output serum biomarker feature matrix and placental image dataset.

[0010] Step S20: Based on the placental image dataset, a multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanism are used to perform the placental microenvironment structure analysis task, outputting the multi-scale feature vector of the placental microenvironment. ;

[0011] Step S30: Based on the serum biomarker feature matrix, a deep feature selection and discriminative compression preprocessing mechanism is used to perform the serum biomarker discriminative feature construction task, and output the serum biomarker feature vector. ;

[0012] Step S40: Based on the multi-scale feature vector of the placental microenvironment With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. ;

[0013] Step S50: Determine the feature vector based on the fusion. A predictive model for adverse pregnancy outcomes in preeclampsia was constructed using a class imbalance adaptive weighted training and threshold adaptive adjustment mechanism, and the predicted probability of adverse pregnancy outcomes in preeclampsia was output.

[0014] Preferably, step S10, which involves acquiring digital full-view images of placental tissue slices and concurrent serum biomarker detection data, performing multimodal sample normalization tasks by combining batch effect fitting correction and case-level paired annotation mechanisms, and outputting the serum biomarker feature matrix and placental image dataset, specifically includes:

[0015] Step S101: Acquire digital full-view images of placental tissue sections and concurrent serum biomarker detection data, and perform denoising, smoothing and effective tissue region identification processing on the digital full-view images of placental tissue sections, and output effective tissue region image data.

[0016] Step S102: Normalize the serum biomarker detection data of the same period using an improved standardization formula that introduces batch effect factors, and output the normalized data matrix of serum biomarkers.

[0017] Step S103: Based on the effective area image data of the tissue and the normalized data matrix of serum biomarkers, pairing and labeling are performed using a timestamp-based mapping method to construct a serum biomarker feature matrix and a placental image dataset.

[0018] Preferably, in step S20, a multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanism is used to perform the placental microenvironment structure analysis task based on the placental image dataset, outputting a multi-scale feature vector of the placental microenvironment. The steps specifically include:

[0019] Step S201: Construct placental image pyramids of different resolutions based on the placental image dataset. The placental image pyramids include 1×1 scale placental images, 2×2 scale placental images, and 4×4 scale placental images.

[0020] Step S202: Use the U-Net depth segmentation network to segment the vascular villi and tissue structure regions of the 1×1 scale placental image, 2×2 scale placental image, and 4×4 scale placental image respectively, and output the structural region segmentation mask results; based on the structural region segmentation mask results, use the OpenCV library and scikit-image library of Python to perform connected component analysis and texture statistical processing, and output the vascular area features, capillary number features, and texture statistical features.

[0021] Step S203: Based on vascular area features, capillary quantity features, and texture statistical features, a scale-adaptive weight learning method is used to construct and output a multi-scale feature vector of the placental microenvironment. .

[0022] Preferably, the steps of performing connected component analysis and texture statistical processing using Python's OpenCV and scikit-image libraries based on the structural region segmentation mask results, and outputting blood vessel area features, capillary quantity features, and texture statistical features, specifically include:

[0023] The total area of ​​blood vessels is obtained by statistically analyzing the pixel area of ​​the structural region segmentation mask results.

[0024] The number of capillaries is obtained by labeling the connected components of the structural region segmentation mask result.

[0025] The contrast, energy, and entropy values ​​of tissue texture are calculated based on the structural region segmentation mask results and gray-level co-occurrence matrix, and these values ​​are used as texture statistical features.

[0026] Preferably, in step S30, a deep feature selection and discriminative compression preprocessing mechanism is used to perform the serum biomarker discriminative feature construction task based on the serum biomarker feature matrix, and the serum biomarker feature vector is output. The steps specifically include:

[0027] Step S301: Input the serum biomarker feature matrix into a preset fully connected deep neural network for feature representation learning, and output the feature contribution of each serum biomarker;

[0028] Step S302: Mark serum biomarker features whose feature contribution is lower than the preset feature contribution threshold as redundant biomarker features. After removing redundant biomarker features, output the set of filtered serum biomarker features.

[0029] Step S303: Based on the screened serum biomarker feature set, principal component analysis is used to perform discriminative feature compression, outputting serum biomarker feature vectors. .

[0030] Preferably, in step S40, the multi-scale feature vector of the placental microenvironment is used. With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. The steps specifically include:

[0031] Step S401: Convert the multi-scale feature vector of the placental microenvironment With serum biomarker feature vector The vectors are concatenated to output a joint feature vector;

[0032] Step S402: Introduce a gated attention network, learn weights for the joint feature vector based on the gated attention network, and output adaptive fusion weights α;

[0033] Step S403: Multi-scale feature vector of the placental microenvironment based on adaptive fusion weight α With serum biomarker feature vector Weighted fusion is performed using a linear mapping method to obtain the fused discriminative feature vector. .

[0034] Preferably, in step S50, the fused discriminative feature vector is used to... The steps for constructing a predictive model for adverse pregnancy outcomes in preeclampsia using a class-imbalanced adaptive weighted training and threshold adaptive adjustment mechanism, and outputting the predicted probability of adverse pregnancy outcomes in preeclampsia, specifically include:

[0035] Step S501: Construct a weighted cross-entropy loss function with class weights based on the adaptive fusion weights α. The weighted cross-entropy loss function is used to enhance the contribution of minority class samples in model training.

[0036] Step S502: Based on the XGBoost open-source model, a pre-training model for predicting adverse pregnancy outcomes in preeclampsia is pre-set. At the same time, historical fusion discriminant feature vectors and corresponding standard sets for predicting adverse pregnancy outcomes in preeclampsia are obtained. The historical fusion discriminant feature vectors are used as the input of the preeclampsia adverse pregnancy outcome prediction model, and the standard sets for predicting adverse pregnancy outcomes in preeclampsia are used as the output of the preeclampsia adverse pregnancy outcome prediction model. Meanwhile, an adaptive sampling strategy and cross-validation mechanism are used to perform the model pre-training process.

[0037] Step S503: Fuse the discriminative feature vector The input is fed into the pre-trained adverse pregnancy outcome prediction model for preeclampsia, and the model outputs the predicted probability of adverse pregnancy outcome for preeclampsia.

[0038] This invention also provides a system for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers, comprising:

[0039] The data preparation module is used to acquire digital full-view images of placental tissue slices and concurrent serum biomarker detection data. It combines batch effect fitting correction and case-level paired annotation mechanism to perform multimodal sample preparation tasks and output serum biomarker feature matrix and placental image dataset.

[0040] The multi-scale image parsing module is used to perform placental microenvironment structure analysis based on a placental image dataset, employing multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanisms to output multi-scale feature vectors of the placental microenvironment. ;

[0041] The serum feature construction module is used to perform the task of constructing serum biomarker discriminative features based on the serum biomarker feature matrix using a deep feature selection and discriminative compression preprocessing mechanism, and outputs serum biomarker feature vectors. ;

[0042] Gated fusion module for use with multi-scale feature vectors of the placental microenvironment With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. ;

[0043] The prediction output module is used to determine the feature vector based on the fusion. A predictive model for adverse pregnancy outcomes in preeclampsia was constructed using a class imbalance adaptive weighted training and threshold adaptive adjustment mechanism, and the predicted probability of adverse pregnancy outcomes in preeclampsia was output.

[0044] This invention also provides a device for predicting preeclampsia with integrated placental microenvironment image features and serum biomarkers, comprising: a memory, a processor, and a preeclampsia with integrated placental microenvironment image features and serum biomarkers prediction program stored in the memory and executable on the processor. When the preeclampsia with integrated placental microenvironment image features and serum biomarkers prediction program is executed by the processor, it realizes the method for predicting preeclampsia with integrated placental microenvironment image features and serum biomarkers prediction.

[0045] The present invention also provides a computer program product, including a preeclampsia adverse pregnancy outcome prediction program that integrates placental microenvironment image features and serum biomarkers. When the preeclampsia adverse pregnancy outcome prediction program that integrates placental microenvironment image features and serum biomarkers is executed by a processor, it implements the preeclampsia adverse pregnancy outcome prediction method that integrates placental microenvironment image features and serum biomarkers.

[0046] The beneficial effects of this invention are as follows: By constructing a multi-scale placental microenvironment image analysis mechanism and a serum biomarker deep discrimination feature construction mechanism, and introducing a gated adaptive fusion weight learning strategy, this invention achieves collaborative modeling of placental morphological abnormality information and biochemical abnormality information, breaking through the technical bottleneck that traditional single-modal analysis cannot characterize the nonlinear correlation between the two types of information, and significantly improving the prediction accuracy and discrimination ability of adverse pregnancy outcomes in preeclampsia.

[0047] This invention addresses the problem of significant batch bias and extremely low proportion of high-risk samples leading to model instability in real clinical data by introducing a batch effect correction standardization processing mechanism and a class imbalance adaptive weighted training control mechanism. This effectively improves the model's recognition sensitivity and generalization stability in scenarios with a small number of high-risk samples, and enhances the reliability of the method for practical application in real clinical environments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating the first embodiment of a method for predicting adverse pregnancy outcomes in preeclampsia that integrates placental microenvironment image features and serum biomarkers according to the present invention.

[0050] Figure 2 This diagram illustrates the distribution differences of the same serum biomarker in different testing batches in the first embodiment of a method for predicting adverse pregnancy outcomes in preeclampsia that integrates placental microenvironment image features and serum biomarkers according to the present invention.

[0051] Figure 3 This is a schematic diagram of the estimation curves of vascular density at different scales for a first embodiment of the method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers according to the present invention.

[0052] Figure 4 This diagram illustrates the ranking of the feature contribution of serum biomarkers in a first embodiment of the method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers according to the present invention.

[0053] Figure 5 This is a schematic diagram of a device for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers, according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the method for predicting adverse pregnancy outcomes in preeclampsia that integrates placental microenvironment image features and serum biomarkers according to the present invention. The first embodiment of the method for predicting adverse pregnancy outcomes in preeclampsia that integrates placental microenvironment image features and serum biomarkers according to the present invention is presented.

[0056] In the first embodiment, the method for predicting adverse pregnancy outcomes in preeclampsia by fusing placental microenvironment image features and serum biomarkers includes:

[0057] Step S10: Obtain digital full-view images of placental tissue sections and concurrent serum biomarker detection data; combine batch effect fitting correction and case-level paired annotation mechanism to perform multimodal sample normalization task; output serum biomarker feature matrix and placental image dataset.

[0058] It should be noted that "digital full-view image of placental tissue sections" refers to digital image data generated by high-resolution full-view scanning of placental tissue sections after fixation, dehydration, paraffin embedding, and staining using a pathological section scanning system. This image fully preserves the morphology of placental villi, vascular distribution, tissue arrangement, and grayscale, color, and texture statistics of different regions, possessing high-dimensional structural information suitable for computer vision analysis. "Concurrent serum biomarker detection data" refers to biochemical indicator values ​​obtained through laboratory testing from pregnant women corresponding to the placental sample during the same gestational stage or adjacent time periods before and after delivery. These indicators include, but are not limited to, sFlt-1 and PlGF. The study identifies biomarkers such as PAPP-A, which are highly correlated with placental developmental abnormalities and endothelial dysfunction. "Batch effect fitting correction" refers to the systematic overall shift phenomenon occurring under different time periods, reagent batches, testing equipment, and experimental conditions. This is achieved by establishing a batch shift fitting model to uniformly correct serum test results, ensuring comparability between different batches of samples. "Case-level paired labeling mechanism" refers to strictly matching and labeling placental image data with serum biomarker data one-to-one based on case number, sample collection timestamp, and electronic medical record information. This ensures complete consistency between the two types of data in both individual and temporal dimensions, thereby guaranteeing semantic consistency in subsequent feature fusion.

[0059] Understandably, placental pathological images are typically stored in pathology information systems, while serum test data is stored in laboratory information systems. The sources of these two data points are dispersed, and their acquisition times differ. Directly using them for model training could easily introduce noisy features due to time mismatches, batch differences, and systematic biases. This step introduces a case-level paired annotation mechanism and batch effect fitting correction, ensuring that placental image data and serum biomarker data have undergone time alignment, individual alignment, and data scale alignment before feature extraction and fusion analysis. This reduces interference from data heterogeneity at the source, providing a unified, stable, and semantically consistent data foundation for subsequent multimodal feature learning, significantly improving the reliability of the model input data.

[0060] It should be understood that traditional techniques often only perform simple mean standardization on serum data, or simply perform routine denoising and cropping on images before direct analysis, without considering the overall shift in serum test results across different batches, nor conducting rigorous case-level matching between images and serum data. This approach can easily lead to the model learning spurious correlation features derived from differences in the detection environment or time, rather than truly reflecting the intrinsic relationship between placental pathological abnormalities and biomarker abnormalities. This proposed solution effectively eliminates these spurious correlation factors by performing batch correction and case-level matching before the data enters the model, enabling the model to focus on learning feature associations with practical medical significance, thereby improving the stability and generalization ability of the predictive model.

[0061] For example, in practical applications, placental tissue slices come from postpartum pathological scans. If these two types of data are directly used for model training, the model may incorrectly treat differences in detection time or batches as discriminative features. Through case-level pairing and batch effect correction in this step, the two types of data corresponding to the pregnant woman can be precisely aligned in the feature space, allowing abnormal fluctuations in serum indicators and placental tissue structural abnormalities to be identified by the model at the same feature scale. Furthermore, serum sFlt-1 test results from different batches may generally be 1 to 2 units higher. After batch effect fitting correction, all samples are mapped to a uniform distribution space, thus avoiding the model's misjudgment of batch differences as pathological differences, ultimately significantly improving the accuracy and reliability of predicting adverse pregnancy outcomes in preeclampsia.

[0062] For example, such as Figure 2As shown in the figure, before batch effect correction, the same serum biomarker exhibits a significant overall distribution shift across different testing batches, with a systematic misalignment of the distribution centers across different batches. This misalignment does not originate from genuine physiological differences but is caused by variations in testing equipment, reagent batches, or experimental conditions. If directly used for model training, it is easily mislearned as a discriminant feature related to preeclampsia, thus introducing spurious correlations.

[0063] After introducing a batch effect fitting correction mechanism, the distributions of samples from different batches were statistically aligned, systematic biases between different batches were effectively eliminated, and relative differences within samples were preserved. This result demonstrates that this step can significantly improve the comparability of cross-batch serological data without diminishing real-world individual differences.

[0064] Therefore, this invention provides a unified, stable and semantically consistent data foundation for subsequent multimodal feature extraction and fusion analysis by batch-correcting serum biomarker data and accurately pairing it with placental images at the case level, thus avoiding the problem of model instability caused by heterogeneous data sources in traditional methods.

[0065] Step S20: Based on the placental image dataset, a multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanism are used to perform the placental microenvironment structure analysis task, outputting the multi-scale feature vector of the placental microenvironment. ;

[0066] It should be noted that "multi-scale image pyramid construction" refers to constructing multiple resolution-level image sets for digital full-view images of placental tissue slices by downsampling, while maintaining the original structural information without distortion. This allows the same placental slice to be analyzed simultaneously at different spatial scales, including low-magnification images that reflect the overall vascular distribution pattern and high-magnification images that depict capillary, villus details, and cell arrangement structures. "Scale-adaptive weighted segmentation mechanism" refers to performing depth segmentation processing on vascular regions, villus regions, and tissue structure regions separately for images at different scales. By introducing scale weight factors, the structural features extracted at different scales are weighted and fused, so that the most discriminative structural information in each scale image is preserved and enhanced.

[0067] Understandably, placental tissue exhibits significant spatial heterogeneity, with abnormalities often present simultaneously at both the macroscopic tissue structure and the microscopic capillary level. Analyzing using only a single magnification or resolution can easily miss key discriminative features at a particular level. This step utilizes parallel multi-scale image analysis, enabling the model to simultaneously perceive tissue-level and cellular-level abnormalities within the same case. This allows for the construction of a more complete representation of the placental microenvironment features, enhancing the discriminative capabilities of subsequent fusion modeling.

[0068] For example, in some cases of preeclampsia, placental abnormalities are mainly manifested as sparse overall vascular distribution, a feature more easily identified in low-magnification images. In other cases, abnormalities are primarily characterized by disordered capillary structures and destruction of villous details, features that require high-magnification images for clear capture. Through the multi-scale construction and weighted segmentation mechanism in this step, abnormal features at both scales can be extracted simultaneously, and the weights are automatically adjusted based on the actual image appearance. This ensures that the final output multi-scale feature vector of the placental microenvironment comprehensively and stably reflects the pathological state of the placenta, providing more sufficient discriminative basis for subsequent prediction models.

[0069] For example, such as Figure 3 As shown in the figure, the estimation results of microenvironment indicators such as placental vascular density at different spatial scales are presented, along with the comprehensive feature curves after scale-adaptive weighted fusion. It can be observed that the estimation results at a single scale fluctuate significantly in different cases: while low-magnification scales can reflect the overall vascular distribution pattern, they are not sensitive to local details; high-magnification scales can capture fine structures such as capillaries, but are easily affected by the quality of local tissue sections and the limitation of the field of view.

[0070] By introducing a multi-scale image pyramid and assigning adaptive weights to features at different scales, the resulting microenvironment feature curve is smoother and more stable overall. It retains the ability of low-scale images to depict the overall structure while incorporating the sensitivity of high-scale images to detail anomalies. This result demonstrates that the scale-adaptive weighting mechanism can automatically adjust the contribution ratio of features at each scale based on the distribution of effective information in images from different cases.

[0071] Therefore, this step effectively solves the problem that traditional single-scale image analysis cannot simultaneously take into account both macroscopic structure and microscopic details, enabling the output multi-scale feature vector of the placental microenvironment to more comprehensively and robustly reflect the true pathological state of the placenta.

[0072] Step S30: Based on the serum biomarker feature matrix, a deep feature selection and discriminative compression preprocessing mechanism is used to perform the serum biomarker discriminative feature construction task, and output the serum biomarker feature vector. ;

[0073] It should be noted that the "serum biomarker feature matrix" refers to the structured serum data set formed after batch effect correction and case-level paired labeling in step S10. This feature matrix is ​​arranged with cases as rows and different serum biomarkers as columns, reflecting the multi-indicator biochemical status of each case at the same data scale. "Deep feature selection" refers to modeling the nonlinear correlation between serum biomarkers by introducing a deep neural network, and evaluating the discriminative ability of the original serum biomarkers based on the feature contribution or importance score learned by the network. "Discriminative compression preprocessing mechanism" refers to compressing or reducing the dimensionality of biomarker features with high redundancy, strong correlation, or low discriminative contribution while retaining the information of biomarkers with high discriminative power, thereby forming a more compact and more discriminative serum biomarker feature vector.

[0074] Understandably, in real-world clinical testing scenarios, there are numerous serum biomarkers, and these markers often exhibit strong correlations or information overlap. Directly inputting all biomarkers into the prediction model would not only increase model complexity but also easily introduce noisy features, affecting the stability of model training. This step utilizes deep feature selection and discriminative compression to enable the model to automatically filter out key biomarker combinations highly correlated with adverse pregnancy outcomes associated with preeclampsia, reducing the interference of redundant information on the model's learning process and thus improving overall predictive performance.

[0075] It should be understood that, compared to traditional methods that screen serum biomarkers based on human experience or simple statistical tests, this step uses deep networks to learn the nonlinear mapping relationship between serum biomarkers and pregnancy outcomes. This makes the feature selection process no longer dependent on fixed thresholds or single statistical indicators, but dynamically determines the features to be retained based on the global discrimination effect, thereby more accurately characterizing complex biological processes and improving the model's generalization ability in different populations and under different testing conditions.

[0076] For example, in a certain dataset, sFlt-1 and PlGF are both considered important biomarkers associated with preeclampsia, but they are highly correlated in some cases. If they are directly input into the model at the same time, it may lead to feature redundancy. Through the deep feature selection mechanism in this step, the model can learn the relative discriminative contributions of the two in different cases, and integrate or compress the highly redundant information in the discriminative compression stage. This ensures that the final output serum biomarker feature vector retains key discriminative information while avoiding unnecessary feature duplication, thereby improving the accuracy and stability of predicting adverse pregnancy outcomes in preeclampsia.

[0077] For example, such as Figure 4As shown in the figure, the ranking of the feature contributions of each serum biomarker in the deep feature selection model is presented in bar chart form, and a threshold line is used to distinguish high-discriminative biomarkers from low-contribution redundant biomarkers. It can be seen that the contributions of different biomarkers to the prediction task vary significantly. Some biomarkers have significant discriminative value for adverse pregnancy outcomes in molecular preeclampsia, while others contribute little or even exhibit information redundancy.

[0078] By setting a discrimination threshold and removing or compressing low-contribution features, a serum biomarker feature vector with lower dimensionality and stronger discriminative power is ultimately formed. This process not only reduces feature dimensionality and model complexity but also effectively suppresses the collinearity problem among biomarkers.

[0079] It can be seen that the present invention can significantly improve the compactness and discriminative power of serum features while retaining key information, providing high-quality input for subsequent multimodal fusion and prediction model training.

[0080] Step S40: Based on the multi-scale feature vector of the placental microenvironment With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. ;

[0081] It should be noted that the "gated adaptive fusion weight learning mechanism" refers to the joint input processing of multi-scale feature vectors of the microenvironment from the placental image channel and biomarker feature vectors from the serum channel by constructing a gated network structure. The fusion weight parameters used to adjust the contribution ratio of the two types of features are then learned through the gated unit, so that different cases can automatically adjust the degree of influence of image information and biochemical information according to their own feature distribution during the fusion process. The "fusion discriminant feature vector" refers to the unified discriminant feature expression formed by the adaptive combination of the two types of features under the gated weight adjustment. This feature vector can simultaneously reflect the comprehensive information of placental morphological abnormalities and serum molecular abnormalities.

[0082] Understandably, the manifestations of abnormal phenotypes vary significantly among different preeclampsia cases. Some cases primarily exhibit placental structural abnormalities, while others show more abnormalities in serum biomarkers. Using a fixed ratio or simple splicing method for fusion can easily lead to the model over-relying on one type of feature in certain cases while neglecting another crucial type. This step employs a gating mechanism to dynamically adjust the fusion weights according to changes in case characteristics, thereby ensuring that the model consistently captures the most discriminative information sources across different case scenarios.

[0083] It should be understood that, compared to the feature concatenation or fixed weighted averaging methods commonly used in traditional multimodal fusion methods, this step introduces a trainable gating network, so that the fusion weights are no longer manually set, but are automatically learned through data-driven methods. This can characterize the complex nonlinear relationship between placental microenvironment features and serum biomarker features, thereby significantly improving the discriminative ability of multimodal feature expression and the generalization performance of the model.

[0084] For example, in some cases, placental images show significant vascular sparseness and disordered villous structure, while serum marker changes are relatively small. In these cases, the gating mechanism automatically increases the fusion weight of placental microenvironment features, making the fused discriminative features more dependent on image information. In other cases, placental image changes are not obvious, but serum sFlt-1 and PlGF show significant abnormalities. In these cases, the gating mechanism decreases the weight of image features and increases the weight of serum features. Through this adaptive adjustment, the fused discriminative feature vector can more accurately reflect the true pathological state of different cases, thereby improving the accuracy and stability of predicting adverse pregnancy outcomes in preeclampsia.

[0085] Step S50: Determine the feature vector based on the fusion. A predictive model for adverse pregnancy outcomes in preeclampsia was constructed using a class imbalance adaptive weighted training and threshold adaptive adjustment mechanism, and the predicted probability of adverse pregnancy outcomes in preeclampsia was output.

[0086] It should be noted that the "class imbalance adaptive weighted training" addresses the objective fact that the number of preeclampsia pregnancy outcome samples is far less than the number of normal pregnancy samples in real clinical data. By introducing a weight coefficient related to class frequency into the model's loss function, a small number of high-risk samples have a greater influence during model parameter updates, thus preventing the model from tending to learn the features of the majority class samples while ignoring the features of key high-risk samples during training. The "threshold adaptive adjustment mechanism" means that after the model training is completed and the predicted probability is output, instead of using a fixed classification threshold, the optimal decision threshold is dynamically found based on the sensitivity, specificity, F1 value, and ROC curve performance in the validation set, so that the model can maintain a better risk identification ability under different sample distributions. The "preeclampsia pregnancy outcome prediction model" is a classification model trained based on the fused discriminative feature vectors obtained in the above steps. Its output is the probability value of adverse pregnancy outcome for each case, rather than a simple binary decision result, thus providing clinicians with more continuous and interpretable risk assessment information.

[0087] Understandably, in medical data scenarios, high-risk cases are inherently scarce. Without addressing this imbalance in class distribution, models may easily achieve lower losses during training by "predicting the majority as normal," thus ignoring the few truly clinically valuable high-risk samples. This step, through class weight adjustment, forces the model to focus on the characteristic patterns of these high-risk samples during learning, thereby improving its ability to identify adverse pregnancy outcomes in preeclampsia. Simultaneously, through an adaptive threshold mechanism, the model's final judgment criteria are no longer fixed but dynamically adjusted based on its performance during the validation phase, ensuring stable discrimination capabilities across different data environments.

[0088] It should be understood that, compared to traditional training methods that only use a fixed cross-entropy loss function and a fixed decision threshold, this step introduces a dual control mechanism of class weighting and dynamic threshold adjustment. This allows the model to automatically adjust its training and discrimination strategies in response to changes in sample distribution, detection environment, and differences in case characteristics, thereby significantly improving the model's generalization ability and reliability in real clinical applications. This design not only enhances the model's sensitivity to minority class samples but also avoids misjudgments caused by improperly set thresholds.

[0089] For example, in a training dataset containing 900 normal pregnancies and 100 adverse pregnancy cases, traditional training methods might only require the model to predict most samples as normal to achieve high accuracy, while the recognition rate for adverse pregnancy cases would be extremely low. Through the class-weighted training mechanism in this step, the weight of adverse pregnancy cases in the loss function is significantly amplified, allowing the model to learn more about the characteristic patterns of these high-risk samples during training. Furthermore, during the validation phase, by analyzing the model's recall and specificity, the default decision threshold was adjusted from 0.5 to 0.35. This allows the model to maintain overall accuracy while improving the recognition rate of adverse pregnancy cases, thus significantly enhancing the model's practical application value in clinical high-risk screening.

[0090] Example 2: Furthermore, the present invention provides a preeclampsia adverse pregnancy outcome prediction system that integrates placental microenvironment image features and serum biomarkers. This system employs a preeclampsia adverse pregnancy outcome prediction method that integrates placental microenvironment image features and serum biomarkers as described in the above embodiments, and can solve the technical problem of predicting preeclampsia adverse pregnancy outcomes using this integrated system. The beneficial effects of the preeclampsia adverse pregnancy outcome prediction system that integrates placental microenvironment image features and serum biomarkers provided by the present invention are the same as those of the preeclampsia adverse pregnancy outcome prediction method that integrates placental microenvironment image features and serum biomarkers provided in the above embodiments. Other technical features of the preeclampsia adverse pregnancy outcome prediction system that integrates placental microenvironment image features and serum biomarkers are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0091] Example 3: This invention provides a device for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers. Please refer to... Figure 5A device for predicting preeclampsia with adverse pregnancy outcomes by fusing placental microenvironment image features and serum biomarkers includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the method for predicting preeclampsia with fusing placental microenvironment image features and serum biomarkers described in Embodiment 1 above. The device for predicting preeclampsia with fusing placental microenvironment image features and serum biomarkers in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This device for predicting preeclampsia with fusing placental microenvironment image features and serum biomarkers is merely an example and should not be construed as limiting the functionality or scope of the embodiments of this invention. A device for predicting preeclampsia with adverse pregnancy outcomes by integrating placental microenvironment image features and serum biomarkers may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows a preeclampsia adverse pregnancy outcome prediction device that integrates placental microenvironment image features and serum biomarkers to wirelessly or wiredly communicate with other devices to exchange data.While the figure illustrates a device for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features with serum biomarkers using various systems, it should be understood that implementation of or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0092] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for predicting adverse pregnancy outcomes in preeclampsia by fusing placental microenvironment image features and serum biomarkers. The computer program product provided by this invention can solve the technical problem of predicting adverse pregnancy outcomes in preeclampsia by fusing placental microenvironment image features and serum biomarkers. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the method for predicting adverse pregnancy outcomes in preeclampsia by fusing placental microenvironment image features and serum biomarkers provided in the above embodiments, and will not be repeated here.

[0093] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0094] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers, characterized in that, The methods include: Step S10: Obtain digital full-view images of placental tissue sections and concurrent serum biomarker detection data; combine batch effect fitting correction and case-level paired annotation mechanism to perform multimodal sample normalization task; output serum biomarker feature matrix and placental image dataset. Step S20: Based on the placental image dataset, a multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanism are used to perform the placental microenvironment structure analysis task, outputting the multi-scale feature vector of the placental microenvironment. ; Step S30: Based on the serum biomarker feature matrix, a deep feature selection and discriminative compression preprocessing mechanism is used to perform the serum biomarker discriminative feature construction task, and output the serum biomarker feature vector. ; Step S40: Based on the multi-scale feature vector of the placental microenvironment With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. ; Step S50: Determine the feature vector based on the fusion. A predictive model for adverse pregnancy outcomes in preeclampsia was constructed using a class imbalance adaptive weighted training and threshold adaptive adjustment mechanism, and the predicted probability of adverse pregnancy outcomes in preeclampsia was output.

2. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 1, characterized in that, Step S10 involves acquiring digital full-view images of placental tissue slices and concurrent serum biomarker detection data. It also includes performing a multimodal sample normalization task by combining batch effect fitting correction and case-level paired annotation mechanisms to output the serum biomarker feature matrix and placental image dataset. Specifically, this includes: Step S101: Acquire digital full-view images of placental tissue sections and concurrent serum biomarker detection data, and perform denoising, smoothing and effective tissue region identification processing on the digital full-view images of placental tissue sections, and output effective tissue region image data. Step S102: Normalize the serum biomarker detection data of the same period using an improved standardization formula that introduces batch effect factors, and output the normalized data matrix of serum biomarkers. Step S103: Based on the effective area image data of the tissue and the normalized data matrix of serum biomarkers, pairing and labeling are performed using a timestamp-based mapping method to construct a serum biomarker feature matrix and a placental image dataset.

3. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 1, characterized in that, In step S20, based on the placental image dataset, a multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanism are used to perform the placental microenvironment structure analysis task, outputting the multi-scale feature vector of the placental microenvironment. The steps specifically include: Step S201: Construct placental image pyramids of different resolutions based on the placental image dataset. The placental image pyramids include 1×1 scale placental images, 2×2 scale placental images, and 4×4 scale placental images. Step S202: Use the U-Net depth segmentation network to segment the vascular villi and tissue structure regions of the 1×1 scale placental image, 2×2 scale placental image, and 4×4 scale placental image respectively, and output the structural region segmentation mask results; based on the structural region segmentation mask results, use the OpenCV library and scikit-image library of Python to perform connected component analysis and texture statistical processing, and output the vascular area features, capillary number features, and texture statistical features. Step S203: Based on vascular area features, capillary quantity features, and texture statistical features, a scale-adaptive weight learning method is used to construct and output a multi-scale feature vector of the placental microenvironment. .

4. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 1, characterized in that... Based on the structural region segmentation mask results, connected component analysis and texture statistical processing are performed using Python's OpenCV and scikit-image libraries to output blood vessel area features, capillary quantity features, and texture statistical features. Specifically, this includes: The total area of ​​blood vessels is obtained by statistically analyzing the pixel area of ​​the structural region segmentation mask results. The number of capillaries is obtained by labeling the connected components of the structural region segmentation mask result. The contrast, energy, and entropy values ​​of tissue texture are calculated based on the structural region segmentation mask results and gray-level co-occurrence matrix, and these values ​​are used as texture statistical features.

5. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 1, characterized in that, In step S30, a deep feature selection and discriminative compression preprocessing mechanism is used to construct serum biomarker discriminative features based on the serum biomarker feature matrix, and the serum biomarker feature vector is output. The steps specifically include: Step S301: Input the serum biomarker feature matrix into a preset fully connected deep neural network for feature representation learning, and output the feature contribution of each serum biomarker; Step S302: Mark serum biomarker features whose feature contribution is lower than the preset feature contribution threshold as redundant biomarker features. After removing redundant biomarker features, output the set of filtered serum biomarker features. Step S303: Based on the screened serum biomarker feature set, principal component analysis is used to perform discriminative feature compression, outputting serum biomarker feature vectors. .

6. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 1, characterized in that, In step S40, based on the multi-scale feature vector of the placental microenvironment... With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. The steps specifically include: Step S401: Convert the multi-scale feature vector of the placental microenvironment With serum biomarker feature vector The vectors are concatenated to output a joint feature vector; Step S402: Introduce a gated attention network, learn weights for the joint feature vector based on the gated attention network, and output adaptive fusion weights α; Step S403: Multi-scale feature vector of the placental microenvironment based on adaptive fusion weight α With serum biomarker feature vector Weighted fusion is performed using a linear mapping method to obtain the fused discriminative feature vector. .

7. The method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in claim 5, characterized in that, In step S50, the fused discriminative feature vector is used as a basis for further analysis. The steps for constructing a predictive model for adverse pregnancy outcomes in preeclampsia using a class-imbalanced adaptive weighted training and threshold adaptive adjustment mechanism, and outputting the predicted probability of adverse pregnancy outcomes in preeclampsia, specifically include: Step S501: Construct a weighted cross-entropy loss function with class weights based on the adaptive fusion weights α. The weighted cross-entropy loss function is used to enhance the contribution of minority class samples in model training. Step S502: Based on the XGBoost open-source model, a pre-training model for predicting adverse pregnancy outcomes in preeclampsia is pre-set. At the same time, historical fusion discriminant feature vectors and corresponding standard sets for predicting adverse pregnancy outcomes in preeclampsia are obtained. The historical fusion discriminant feature vectors are used as the input of the preeclampsia adverse pregnancy outcome prediction model, and the standard sets for predicting adverse pregnancy outcomes in preeclampsia are used as the output of the preeclampsia adverse pregnancy outcome prediction model. Meanwhile, an adaptive sampling strategy and cross-validation mechanism are used to perform the model pre-training process. Step S503: Fuse the discriminative feature vector The input is fed into the pre-trained adverse pregnancy outcome prediction model for preeclampsia, and the model outputs the predicted probability of adverse pregnancy outcome for preeclampsia.

8. A system for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers, applied to the method for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers as described in any one of claims 1 to 7, characterized in that, The preeclampsia adverse pregnancy outcome prediction system, which integrates placental microenvironment image features and serum biomarkers, includes: The data preparation module is used to acquire digital full-view images of placental tissue slices and concurrent serum biomarker detection data. It combines batch effect fitting correction and case-level paired annotation mechanism to perform multimodal sample preparation tasks and output serum biomarker feature matrix and placental image dataset. The multi-scale image parsing module is used to perform placental microenvironment structure analysis based on a placental image dataset, employing multi-scale image pyramid construction and scale-adaptive weighted segmentation mechanisms to output multi-scale feature vectors of the placental microenvironment. ; The serum feature construction module is used to perform the task of constructing serum biomarker discriminative features based on the serum biomarker feature matrix using a deep feature selection and discriminative compression preprocessing mechanism, and outputs serum biomarker feature vectors. ; Gated fusion module for use with multi-scale feature vectors of the placental microenvironment With serum biomarker feature vector A gated adaptive fusion weight learning mechanism is used to perform a multimodal feature dynamic fusion task, and the output is a fused discriminative feature vector. ; The prediction output module is used to determine the feature vector based on the fusion. A predictive model for adverse pregnancy outcomes in preeclampsia was constructed using a class imbalance adaptive weighted training and threshold adaptive adjustment mechanism, and the predicted probability of adverse pregnancy outcomes in preeclampsia was output.

9. A device for predicting adverse pregnancy outcomes in preeclampsia by integrating placental microenvironment image features and serum biomarkers, characterized in that, The device for predicting adverse preeclampsia outcomes by integrating placental microenvironment image features and serum biomarkers includes: a memory, a processor, and a program for predicting adverse preeclampsia outcomes by integrating placental microenvironment image features and serum biomarkers stored in the memory and executable on the processor. When the program for predicting adverse preeclampsia outcomes by integrating placental microenvironment image features and serum biomarkers is executed by the processor, it implements a method for predicting adverse preeclampsia outcomes by integrating placental microenvironment image features and serum biomarkers according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a preeclampsia adverse pregnancy outcome prediction program that integrates placental microenvironment image features and serum biomarkers. When the preeclampsia adverse pregnancy outcome prediction program that integrates placental microenvironment image features and serum biomarkers is executed by a processor, it implements a preeclampsia adverse pregnancy outcome prediction method that integrates placental microenvironment image features and serum biomarkers as described in any one of claims 1 to 7.

Images