ART offspring circulation system defect prediction method and system and storage medium

Abstract

The invention provides an ART offspring circulatory system defect prediction method and system and a storage medium. The method comprises the following steps: acquiring parental clinical data, gene data and embryo monitoring image data of a to-be-predicted ART offspring; extracting an embryonic development event time node based on the image data, calculating a deviation with a preset standard model, generating a multi-scale space-time distortion parameter matrix, and converting the multi-scale space-time distortion parameter matrix into a development asynchronism scalar; analyzing parental data by using a circulatory system defect related gene-clinical index knowledge graph, constructing a weighted feature network, developing an asynchronous scalar to adjust an edge weight, and calculating initial feature propagation centrality of each node; inputting the space-time distortion parameter matrix into an embryo channel of a double-flow interactive attention network, and inputting the initial feature propagation centrality and other parental clinical data into a parental channel; and fusing the two-channel features through an interactive attention layer to generate context sensing fusion features, and outputting the risk probability of occurrence of circulatory system defects of ART offspring based on the features.

Description

Technical Field

[0001] This application belongs to the field of defect prediction, and in particular relates to a defect prediction method, system and storage medium for ART offspring loop system. Background Technology

[0002] Circulatory system defects (CSDs) are among the most common types of birth defects, involving multiple factors related to genetics, environment, and embryonic development. Risk assessment for CSDs in ART offspring cannot provide early warning before embryo implantation. During embryo selection, clinical assessment methods based on morphology and developmental parameters are used. The subtle spatiotemporal deviations between the embryo's developmental rhythm and the standard pattern—developmental asynchrony—are crucial information reflecting potential developmental abnormalities. While preimplantation genetic testing (PGS) can screen for chromosomal abnormalities and single-gene disorders, it cannot adequately predict CSDs regulated by multiple genes and factors. The occurrence of CSDs is related to the parental genetic background and clinical condition. However, in embryo assessment practice, the use of parental data has failed to establish an intrinsic correlation model between parental risk factors and specific developmental phenotypes in offspring embryos. Existing predictive models are mostly statistical models or simple machine learning methods, which struggle to process and integrate diverse and structurally heterogeneous data, and cannot detect interactions between different modalities. Therefore, a method is needed that can deeply integrate the spatiotemporal characteristics of embryonic development with parental multi-omics risk information to achieve early prediction of the risk of circulatory system defects in ART offspring, and provide a scientific basis for clinical embryo selection. Summary of the Invention

[0003] This invention proposes a defect prediction method for ART offspring recurrent systems to address the problem that existing models struggle to process and fuse data from diverse sources and with heterogeneous structures, and are unable to detect interactions between different modalities. The method includes: Acquire parental clinical data, parental genetic data, and embryo monitoring image data of the ART offspring to be predicted; based on the embryo monitoring image data, extract the time nodes of key embryonic development events, calculate the deviation of the nodes from the preset standard developmental spatiotemporal model, generate a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtain the developmental asynchrony scalar based on the matrix; Parental genes and clinical data are analyzed using a gene-clinical indicator knowledge graph related to circulatory system defects. A weighted feature network is constructed, and the edge weights in the weighted feature network are adjusted using the developmental asynchrony scalar. The initial feature propagation centrality of each node is calculated based on the adjusted weighted feature network. The multi-scale spatiotemporal distortion parameter matrix is ​​input into the embryo channel of the two-stream interactive attention network, and the initial feature propagation centrality and other parental clinical data are input into the parental channel of the network. The dual-stream interactive attention network fuses the feature information of the embryonic channel and the parental channel through an interactive attention layer to generate context-aware fusion features, and outputs the risk probability of circulatory system defects in the ART offspring based on the context-aware fusion features.

[0004] Furthermore, this invention also relates to an ART offspring loop system defect prediction system, comprising the following modules: The generation module is used to acquire parental clinical data, parental genetic data, and embryo monitoring image data of the offspring to be predicted for ART; based on the embryo monitoring image data, it extracts the time nodes of key events in embryonic development, calculates the deviation of the nodes from the preset standard developmental spatiotemporal model, generates a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtains the developmental asynchrony scalar based on the matrix; The computation module is used to analyze parental genes and clinical data using a gene-clinical indicator knowledge graph related to circulatory system defects, construct a weighted feature network, adjust the edge weights in the weighted feature network using the developmental asynchrony scalar, and calculate the initial feature propagation centrality of each node based on the adjusted weighted feature network. The input module is used to input the multi-scale spatiotemporal distortion parameter matrix into the embryo channel of the two-stream interactive attention network, and to input the initial feature propagation centrality and other parental clinical data into the parental channel of the network. The output module is used by the dual-stream interactive attention network to fuse the feature information of the embryonic channel and the parental channel through the interactive attention layer, generate context-aware fusion features, and output the risk probability of the ART offspring having circulatory system defects based on the context-aware fusion features.

[0005] This invention proposes a multi-scale spatiotemporal distortion parameter to represent the overall asynchronicity of embryonic development, enabling the detection of subtle developmental rhythm abnormalities that predict defect risk. This developmental asynchronicity feature is used to adjust a parental gene-clinical weighted network constructed based on a knowledge graph, thereby establishing a correlation between specific embryonic developmental phenotypes and parental genetic and clinical risk factors, solving the technical challenge of information separation between parents and offspring. By fusing embryonic and parental features through a two-stream interactive attention network, deep relationships between different modalities of data are uncovered, improving the reliability of the predictive model and providing a scientific basis for clinical embryo implantation decisions. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of a two-stream network structure; Figure 3This is a schematic diagram of the interactive attention fusion mechanism; Figure 4 This is a schematic diagram of the risk probability output module. Detailed Implementation

[0007] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0008] The term "multiple" in this application refers to two or more. Furthermore, it should be understood that the terms "first," "second," etc., used in the description of this application are only used to distinguish the purpose of the description and should not be construed as indicating or implying relative importance, nor as indicating or implying order. If the technical solution of this application involves personal information, the product applying the technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product applying the technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent." For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected; if an individual voluntarily enters the collection scope, it is deemed that they have consented to the collection of their personal information. Alternatively, on personal information processing devices, with clear signs / information informing users of the personal information processing rules, authorization is obtained through pop-up messages or by asking users to upload their personal information themselves. The personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0009] In the first embodiment, the present invention proposes a method for predicting defects in ART offspring loop systems, such as... Figure 1 ,include: S1. Acquire parental clinical data, parental genetic data, and embryo monitoring image data of the ART offspring to be predicted; based on the embryo monitoring image data, extract the time nodes of key embryonic development events, calculate the deviation of the nodes from the preset standard developmental spatiotemporal model, generate a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtain the developmental asynchrony scalar based on the matrix. Parental clinical data were obtained through the hospital's electronic medical record system, including the mother's age, body mass index, blood glucose level, and history of heart disease. Parental genetic data were obtained through whole-exome sequencing of the parental peripheral blood, including known gene locus variation information related to circulatory system development, such as single nucleotide polymorphisms of GATA4 and NKX2-5 genes. Embryo monitoring imaging data were acquired through a time-lapse photography system in the embryo incubator, consisting of a continuous image sequence taken every 10 minutes from fertilization to the blastocyst stage.

[0010] A convolutional neural network was used to segment the image sequence, identifying and recording the absolute time points of key events such as pronuclear appearance, 2-cell, 4-cell, 8-cell, morula, and blastocyst formation. A pre-defined standard developmental spatiotemporal model was used, based on the average time point sequence obtained from statistical analysis of embryonic development data from over ten thousand normal live births. A time warping algorithm was employed to compare the event time series of the tested embryo with the standard model sequence, calculating the degree of stretching or compression of the time axis at different scales: cell division, morula, and blastocyst stages. The parameters of the warped path constituted a multi-scale spatiotemporal distortion parameter matrix. The square root of the sum of squares of all elements in this matrix was calculated to obtain a scalar value representing the overall degree of developmental asynchrony.

[0011] In an optional embodiment, the extraction of key embryonic development events includes: Using image segmentation and time series analysis algorithms, the number of hours for pronuclear disappearance time tPNf, 2-cell division time t2, 3-cell division time t3, 4-cell division time t4, 5-cell division time t5, morula formation time tM, and blastocyst formation time tB are identified and recorded from the embryo monitoring image data.

[0012] Specifically, a deep learning model based on the U-Net architecture is used to process each frame of the embryo monitoring image data. This model is pre-trained on a large number of labeled human embryo images and is able to identify and segment the embryo's outline and the internal blastomeres (cells). For example, for frame 520 of the video, the model outputs a mask image where the embryonic region is labeled as 1 and the background region as 0, while also determining the boundary between the two cells present at that moment.

[0013] The segmentation results of all frames are concatenated into a time series. By analyzing the change in the number of segmented cells over time, the occurrence of key events is determined. For example, a frame where the number of cells changes from 1 to 2 is recorded as cell division time t2, with the corresponding timestamp, such as 26.8 hours after fertilization. Similarly, tPNf is recorded when the segmentation regions of two pronuclei disappear; tM is recorded when cells tightly aggregate to form cell clusters indistinguishable from individual cells; and tB is recorded when a blastocoel appears and expands. The entire process requires no manual intervention and achieves the extraction of seven key time points.

[0014] In an optional embodiment, generating a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtaining a developmental asynchrony scalar based on the matrix, includes: The absolute values ​​of the differences between the seven extracted key event time points and the corresponding reference time points [24.5h, 27.0h, 38.0h, 40.0h, 50.0h, 92.0h, 112.0h] in the standard developmental model are calculated to obtain the deviation vector. The deviation vector is then window-smoothed and averaged at three time scales of 1 hour, 4 hours, and 12 hours to generate a 7×3 multi-scale spatiotemporal distortion parameter matrix. The developmental asynchrony scalar is obtained by calculating the square root of the sum of the squares of all elements in the matrix.

[0015] Suppose that the seven key event time points of a certain embryo extracted above are [25.0h, 26.8h, 39.5h, 41.0h, 52.0h, 90.0h, 115.0h]. Subtract the corresponding elements of this vector from the standard reference time point vector [24.5h, 27.0h, 38.0h, 40.0h, 50.0h, 92.0h, 112.0h] and take the absolute value to obtain the original deviation vector V=[0.5,0.2,1.5,1.0,2.0,2.0,3.0]. This vector reflects the instantaneous deviation degree of each event.

[0016] To detect developmental rhythm disturbances at different time scales, the deviation vector V was smoothed using multi-scale methods. At the 1-hour scale, due to the small time window, the smoothed vector is essentially equal to the original deviation vector; this is the first column of the matrix. At the 4-hour scale, a suitable moving average window was used to smooth V, reflecting short-term developmental fluctuations; this is the second column of the matrix. At the 12-hour scale, a large moving average window was used for smoothing, revealing long-term developmental trend deviations; this is the third column of the matrix. This constitutes a 7x3 multi-scale spatiotemporal distortion parameter matrix M. The square root of the sum of the squares of all 21 elements in the matrix yields a scalar value D, for example, 5.7. This value comprehensively represents the degree of overall asynchrony between embryonic development and the standard pattern across multiple time dimensions.

[0017] S2, use the gene-clinical indicator knowledge graph related to circulatory system defects to analyze parental genes and clinical data, construct a weighted feature network, and use the developmental asynchrony scalar to adjust the edge weights in the weighted feature network, and calculate the initial feature propagation centrality of each node based on the adjusted weighted feature network; The knowledge graph is pre-constructed by integrating the OMIM, ClinVar public databases, and relevant medical literature. Nodes represent genes, clinical indicators, and diseases, while edges represent the strength of associations between them. For a test sample, the gene variation information and abnormal clinical indicators of the sample's parents are mapped to nodes in the graph, and these nodes and their first-order neighbors are extracted to form a weighted feature network for the sample. An adjustment factor is generated using a hyperbolic tangent function with a developmental asynchrony scalar as input. This factor is multiplied by the original weights of all edges in the network to achieve adjustment, thereby amplifying the influence of parental risk factors corresponding to highly asynchronous embryos. A personalized Page ranking algorithm is used to iteratively calculate the score for each gene and clinical indicator node on the adjusted network, which serves as the initial feature propagation centrality.

[0018] In an optional embodiment, the step of using a gene-clinical indicator knowledge graph related to circulatory system defects to analyze parental genes and clinical data and construct a weighted feature network includes: The MTHFRC677T and VLeiden mutation status from parental genetic data, as well as pre-pregnancy homocysteine ​​levels, body mass index (BMI), D-dimer, and anticardiolipin antibodies from parental clinical data, were used as network nodes. Based on the "gene-influence-indicator" or "indicator-synergy-indicator" relationship defined in the knowledge graph, connection edges were established between the corresponding nodes, and each edge was assigned an initial weight in the range of [1.0, 5.0] according to the association strength ratio reported in evidence-based medicine literature.

[0019] Specifically, six nodes are established for a couple's data, representing the MTHFRC677T genotype, coagulation factor VLeiden genotype, homocysteine ​​level, BMI, D-dimer level, and anticardiolipin antibody status, respectively. For example, a mother's data might show a heterozygous MTHFRC677T mutation and a homocysteine ​​level of 13 μmol / L; these two indicators would then become two specific nodes in the network.

[0020] A pre-constructed knowledge graph of extensive medical knowledge is used to determine the relationships between nodes. This knowledge graph stores triples of knowledge, such as the impact of the MTHRC677T mutation and elevated homocysteine ​​levels. When both an MTHRC677T and a homocysteine ​​node exist, a connection is established between them based on this knowledge. Similarly, if the knowledge graph contains relationships such as high BMI, synergistic effects, and high D-dimer levels, an edge is also established between the BMI and D-dimer nodes. Each edge is assigned a weight. For example, an analysis of authoritative medical literature indicates that the odds ratio of MTHRC677T heterozygous mutation leading to hyperhomocysteinemia is 4.2; therefore, 4.2 is used as the initial weight for this edge. All weight values ​​are normalized to the range of 1.0 to 5.0, resulting in a weighted network representing the strength of interactions between parental intrinsic risk factors.

[0021] To integrate the characteristics of embryonic development, namely the developmental asynchrony scalar D, into the parental risk assessment network for individualized adjustment, in an optional embodiment, adjusting the edge weights in the weighted feature network using the developmental asynchrony scalar includes: For each edge in the weighted feature network, the adjusted weight Through formula The calculation yields W, where W is the initial weight of the edge and D is the developmental asynchrony scalar.

[0022] Taking the aforementioned example, assuming the initial weight W between the MTHFRC677T node and the homocysteine ​​node is 4.2, and the corresponding embryo's calculated developmental asynchrony scalar D is 5.7.

[0023] The exponential value is calculated as -0.285. The exponential function value of this value is then calculated, approximately 0.752. This 0.752 is the regulatory factor based on the embryonic development status. The initial weight W is multiplied by this regulatory factor to obtain the adjusted weight W', approximately 3.16. This process is performed on every edge in the network. If an embryo develops very synchronously, the D value will be small, for example, close to 0, and the regulatory factor will be close to 1, with the parental network weights remaining essentially unchanged. Conversely, if embryonic development is severely asynchronous, the D value will be large, and the regulatory factor will become very small, thus weakening the influence of parental genetics and clinical factors in the overall risk assessment model, reflecting the importance of the embryo's own developmental state in prediction.

[0024] S3, input the multi-scale spatiotemporal distortion parameter matrix into the embryo channel of the dual-stream interactive attention network, and input the initial feature propagation centrality and other parental clinical data into the parental channel of the network; In one embodiment, the embryonic channel of the dual-stream interactive attention network consists of multiple one-dimensional convolutional layers and self-attention layers, used to receive and process a multi-scale spatiotemporal distortion parameter matrix representing the rhythm of embryonic development. The parental channel of the network consists of a multilayer perceptron, and the input is a concatenated vector, which is a vector composed of the initial feature propagation centrality of each node calculated in the previous step, concatenated with standardized maternal age, body mass index, and other parental clinical data not included in the knowledge graph.

[0025] In an optional embodiment, the step of inputting the multi-scale spatiotemporal distortion parameter matrix into the embryonic channel of the two-stream interactive attention network, and inputting the initial feature propagation centrality and other parental clinical data into the parental channel of the network, includes: The embryonic channel consists of two one-dimensional convolutional layers with a kernel size of 3 and channel numbers of 32 and 64, respectively; the parental channel consists of a three-layer fully connected feedforward network with 128, 64, and 64 neurons in each layer; other parental clinical data include values ​​representing maternal age, paternal age, smoking history, and drinking history.

[0026] Specifically, the model comprises two parallel information processing streams. The embryo channel specifically processes a 7×3 dimensional multi-scale spatiotemporal distortion parameter matrix representing the rhythm of embryonic development. This matrix is ​​fed into a one-dimensional convolutional neural network. The first convolutional layer uses 32 kernels of size 3, sliding along the 7 time points to detect local features of deviation patterns between adjacent developmental events. The output is fed into a second convolutional layer, which uses 64 kernels of the same size to extract higher-level, more abstract features of developmental asynchrony. This channel outputs a feature vector that encapsulates information about embryonic development.

[0027] Simultaneously, the parental pathway processes data from both parents. Input data includes the initial feature propagation centrality of each node calculated from the network before adjustment, as well as other important clinical variables, such as mother's age (35 years), father's age (38 years), mother's smoking history (coded as 1), and drinking history (coded as 0). These values ​​are combined into a long vector. This vector is input into a feedforward neural network consisting of three fully connected layers. The first layer of the network has 128 neurons, and the second and third layers each have 64 neurons. Through layer-by-layer transformation, the raw, discrete parental data is converted into a parental risk feature vector, such as... Figure 2 .

[0028] S4, the dual-stream interactive attention network fuses the feature information of the embryonic channel and the parental channel through the interactive attention layer to generate context-aware fusion features, and outputs the risk probability of the ART offspring having circulatory system defects based on the context-aware fusion features.

[0029] After extracting high-dimensional features from the embryonic and parental channels respectively, the interactive attention layer performs cross-attention calculations. For example, using embryonic features as the query vector and parental features as the key and value vectors, a parental information-weighted embryonic feature is calculated; conversely, the parental features are used for the other two. These two weighted feature vectors are then concatenated to form a context-aware fusion feature that incorporates both embryonic developmental rhythm information and parental risk background, such as... Figure 3 The fused feature is fed into an output layer containing two fully connected layers and a sigmoid activation function, which outputs a value between 0 and 1, representing the predicted risk probability.

[0030] In an optional embodiment, the dual-stream interactive attention network fuses feature information from the embryonic channel and the parental channel through an interactive attention layer to generate context-aware fusion features, including: The feature vector output from the embryo channel is used as the query, and the feature vector output from the parent channel is used as the key and value. The parent-information-weighted embryo features are calculated using a dot product attention mechanism. At the same time, the feature vector output from the parent channel is used as the query, and the feature vector output from the embryo channel is used as the key and value. The embryo-information-weighted parent features are calculated. The two weighted feature vectors are added element-wise and then concatenated with the original features to obtain the context-aware fusion features.

[0031] To ensure that embryonic features focus on the most relevant parental information, the embryonic feature vector is used as the query, and the parental feature vectors are used as the key and value. A set of attention weights is obtained by calculating the dot product of the embryonic query vector and the parental key vector, and then normalizing it using a softmax function. This set of weights is used to perform a weighted summation of the parental value vectors, generating a context vector of parental information. This context vector is then added to the original embryonic feature vector to obtain an enhanced version of the embryonic features that incorporates parental background information.

[0032] At this point, the parental feature vector serves as the query, and the embryonic feature vector serves as the key and value, repeating the dot product attention calculation process described above. A context vector of embryonic information is generated, which is added to the original parental feature vector to obtain an enhanced parental feature that incorporates the embryonic developmental state. The two enhanced feature vectors, namely the parent-information-weighted embryonic feature and the embryo-information-weighted parental feature, are then added element-wise. The result of this addition is further concatenated with the original feature vectors output from the two channels to form a context-aware fusion feature with higher dimensions and more comprehensive information. This feature simultaneously detects information from two sources and deep interaction relationships.

[0033] In an optional embodiment, the risk probability of the ART offspring experiencing a cyclic system defect is output based on the context-aware fusion features, including: The context-aware fusion features are input into a two-layer fully connected network consisting of 64 neurons and 16 neurons for feature extraction. A sigmoid activation function is used to output a value between 0 and 1, which is the risk probability.

[0034] Specifically, after obtaining the context-aware fused feature vector through the interactive attention layer, this high-dimensional vector is fed into a prediction module. This module is a relatively simple two-layer fully connected feedforward network. The first layer contains 64 neurons, which receives the fused feature vector and performs a transformation to extract key combined features for the prediction target from the interactive information. The output of the first layer is fed into the second layer, which contains 16 neurons for feature compression and refinement. The outputs of these 16 neurons are aggregated into a single output neuron. This neuron uses the Sigmoid activation function. The expression for the Sigmoid function is: Its characteristic is that it can map any real number input to an open interval between 0 and 1. Therefore, the output value of the neuron, for example, 0.735, can be interpreted as a predicted probability. This value indicates that, based on the model's comprehensive analysis of all information from embryonic development and parental data, the risk of the offspring corresponding to this embryo having circulatory system defects in the future is assessed to be 73.5%. Figure 4 .

[0035] In the second embodiment, the present invention also proposes an ART offspring loop system defect prediction system, comprising the following modules: The generation module is used to acquire parental clinical data, parental genetic data, and embryo monitoring image data of the offspring to be predicted for ART; based on the embryo monitoring image data, it extracts the time nodes of key events in embryonic development, calculates the deviation of the nodes from the preset standard developmental spatiotemporal model, generates a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtains the developmental asynchrony scalar based on the matrix; The computation module is used to analyze parental genes and clinical data using a gene-clinical indicator knowledge graph related to circulatory system defects, construct a weighted feature network, adjust the edge weights in the weighted feature network using the developmental asynchrony scalar, and calculate the initial feature propagation centrality of each node based on the adjusted weighted feature network. The input module is used to input the multi-scale spatiotemporal distortion parameter matrix into the embryo channel of the two-stream interactive attention network, and to input the initial feature propagation centrality and other parental clinical data into the parental channel of the network. The output module is used by the dual-stream interactive attention network to fuse the feature information of the embryonic channel and the parental channel through the interactive attention layer, generate context-aware fusion features, and output the risk probability of the ART offspring having circulatory system defects based on the context-aware fusion features.

[0036] In an optional embodiment, the extraction of key embryonic development events includes: Using image segmentation and time series analysis algorithms, the number of hours for pronuclear disappearance time tPNf, 2-cell division time t2, 3-cell division time t3, 4-cell division time t4, 5-cell division time t5, morula formation time tM, and blastocyst formation time tB are identified and recorded from the embryo monitoring image data.

[0037] In an optional embodiment, generating a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtaining a developmental asynchrony scalar based on the matrix, includes: The absolute values ​​of the differences between the seven extracted key event time points and the corresponding reference time points [24.5h, 27.0h, 38.0h, 40.0h, 50.0h, 92.0h, 112.0h] in the standard developmental model are calculated to obtain the deviation vector. The deviation vector is then window-smoothed and averaged at three time scales of 1 hour, 4 hours, and 12 hours to generate a 7×3 multi-scale spatiotemporal distortion parameter matrix. The developmental asynchrony scalar is obtained by calculating the square root of the sum of the squares of all elements in the matrix.

[0038] In an optional embodiment, the step of using a gene-clinical indicator knowledge graph related to circulatory system defects to analyze parental genes and clinical data and construct a weighted feature network includes: The MTHFRC677T and VLeiden mutation status from parental genetic data, as well as pre-pregnancy homocysteine ​​levels, body mass index (BMI), D-dimer, and anticardiolipin antibodies from parental clinical data, were used as network nodes. Based on the "gene-influence-indicator" or "indicator-synergy-indicator" relationship defined in the knowledge graph, connection edges were established between the corresponding nodes, and each edge was assigned an initial weight in the range of [1.0, 5.0] according to the association strength ratio reported in evidence-based medicine literature.

[0039] In an optional embodiment, adjusting the edge weights in the weighted feature network using the developmental asynchrony scalar includes: For each edge in the weighted feature network, the adjusted weight Through formula The calculation yields W, where W is the initial weight of the edge and D is the developmental asynchrony scalar.

[0040] In an optional embodiment, the step of inputting the multi-scale spatiotemporal distortion parameter matrix into the embryonic channel of the two-stream interactive attention network, and inputting the initial feature propagation centrality and other parental clinical data into the parental channel of the network, includes: The embryonic channel consists of two one-dimensional convolutional layers with a kernel size of 3 and channel numbers of 32 and 64, respectively; the parental channel consists of a three-layer fully connected feedforward network with 128, 64, and 64 neurons in each layer; other parental clinical data include values ​​representing maternal age, paternal age, smoking history, and drinking history.

[0041] In an optional embodiment, the dual-stream interactive attention network fuses feature information from the embryonic channel and the parental channel through an interactive attention layer to generate context-aware fusion features, including: The feature vector output from the embryo channel is used as the query, and the feature vector output from the parent channel is used as the key and value. The parent-information-weighted embryo features are calculated using a dot product attention mechanism. At the same time, the feature vector output from the parent channel is used as the query, and the feature vector output from the embryo channel is used as the key and value. The embryo-information-weighted parent features are calculated. The two weighted feature vectors are added element-wise and then concatenated with the original features to obtain the context-aware fusion features.

[0042] In an optional embodiment, the risk probability of the ART offspring experiencing a cyclic system defect is output based on the context-aware fusion features, including: The context-aware fusion features are input into a two-layer fully connected network consisting of 64 neurons and 16 neurons for feature extraction. A sigmoid activation function is used to output a value between 0 and 1, which is the risk probability.

[0043] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0044] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0045] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. An ART progeny circulatory system defect prediction method, characterized by, include: Acquire parental clinical data, parental genetic data, and embryo monitoring imaging data of the ART offspring to be predicted; Based on the embryo monitoring image data, time nodes of key embryonic development events are extracted, and the deviation of the nodes from the preset standard developmental spatiotemporal model is calculated. A multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony is generated, and a developmental asynchrony scalar is obtained based on the matrix. Parental genes and clinical data are analyzed using a gene-clinical indicator knowledge graph related to circulatory system defects. A weighted feature network is constructed, and the edge weights in the weighted feature network are adjusted using the developmental asynchrony scalar. The initial feature propagation centrality of each node is calculated based on the adjusted weighted feature network. The multi-scale spatiotemporal distortion parameter matrix is ​​input into the embryo channel of the two-stream interactive attention network, and the initial feature propagation centrality and other parental clinical data are input into the parental channel of the network. The dual-stream interactive attention network fuses the feature information of the embryonic channel and the parental channel through an interactive attention layer to generate context-aware fusion features, and outputs the risk probability of circulatory system defects in the ART offspring based on the context-aware fusion features.

2. The method of claim 1, wherein, The time points for extracting key events in embryonic development include: Using image segmentation and time series analysis algorithms, the number of hours for pronuclear disappearance time tPNf, 2-cell division time t2, 3-cell division time t3, 4-cell division time t4, 5-cell division time t5, morula formation time tM, and blastocyst formation time tB are identified and recorded from the embryo monitoring image data.

3. The method of claim 1, wherein, The process of generating a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtaining a developmental asynchrony scalar based on the matrix, includes: The absolute values ​​of the differences between the seven extracted key event time points and the corresponding reference time points [24.5h, 27.0h, 38.0h, 40.0h, 50.0h, 92.0h, 112.0h] in the standard developmental model are calculated to obtain the deviation vector. The deviation vector is then window-smoothed and averaged at three time scales of 1 hour, 4 hours, and 12 hours to generate a 7×3 multi-scale spatiotemporal distortion parameter matrix. The developmental asynchrony scalar is obtained by calculating the square root of the sum of the squares of all elements in the matrix.

4. The method according to claim 1, characterized in that, The method of using a gene-clinical indicator knowledge graph related to circulatory system defects to analyze parental genes and clinical data, and constructing a weighted feature network, includes: The MTHFRC677T and VLeiden mutation status from parental genetic data, as well as pre-pregnancy homocysteine ​​levels, body mass index (BMI), D-dimer, and anticardiolipin antibodies from parental clinical data, were used as network nodes. Based on the "gene-influence-indicator" or "indicator-synergy-indicator" relationship defined in the knowledge graph, connection edges were established between the corresponding nodes, and each edge was assigned an initial weight in the range of [1.0, 5.0] according to the correlation strength ratio reported in evidence-based medicine literature.

5. The method according to claim 1, characterized in that, The adjustment of edge weights in the weighted feature network using the developmental asynchrony scalar includes: For each edge in the weighted feature network, the adjusted weight Through formula The calculation yields W, where W is the initial weight of the edge and D is the developmental asynchrony scalar.

6. The method according to claim 1, characterized in that, The step of inputting the multi-scale spatiotemporal distortion parameter matrix into the embryonic channel of the two-stream interactive attention network, and inputting the initial feature propagation centrality and other parental clinical data into the parental channel of the network, includes: The embryonic channel consists of two one-dimensional convolutional layers with a kernel size of 3 and channel numbers of 32 and 64, respectively; the parental channel consists of a three-layer fully connected feedforward network with 128, 64, and 64 neurons in each layer; other parental clinical data include values ​​representing maternal age, paternal age, smoking history, and drinking history.

7. The method according to claim 1, characterized in that, The dual-stream interactive attention network fuses feature information from the embryonic channel and the parental channel through an interactive attention layer to generate context-aware fusion features, including: The feature vector output from the embryo channel is used as the query, and the feature vector output from the parent channel is used as the key and value. The parent-information-weighted embryo features are calculated using a dot product attention mechanism. At the same time, the feature vector output from the parent channel is used as the query, and the feature vector output from the embryo channel is used as the key and value. The embryo-information-weighted parent features are calculated. The two weighted feature vectors are added element-wise and then concatenated with the original features to obtain the context-aware fusion features.

8. The method according to claim 1, characterized in that, The probability of the ART offspring experiencing a cyclic system defect is output based on the context-aware fusion features, including: The context-aware fusion features are input into a two-layer fully connected network consisting of 64 neurons and 16 neurons for feature extraction. A sigmoid activation function is used to output a value between 0 and 1, which is the risk probability.

9. A defect prediction system for ART offspring loop systems, characterized in that, Includes the following modules: The generation module is used to acquire parental clinical data, parental genetic data, and embryo monitoring image data of the offspring to be predicted for ART; based on the embryo monitoring image data, it extracts the time nodes of key events in embryonic development, calculates the deviation of the nodes from the preset standard developmental spatiotemporal model, generates a multi-scale spatiotemporal distortion parameter matrix representing developmental asynchrony, and obtains the developmental asynchrony scalar based on the matrix; The computation module is used to analyze parental genes and clinical data using a gene-clinical indicator knowledge graph related to circulatory system defects, construct a weighted feature network, adjust the edge weights in the weighted feature network using the developmental asynchrony scalar, and calculate the initial feature propagation centrality of each node based on the adjusted weighted feature network. The input module is used to input the multi-scale spatiotemporal distortion parameter matrix into the embryo channel of the two-stream interactive attention network, and to input the initial feature propagation centrality and other parental clinical data into the parental channel of the network. The output module is used by the dual-stream interactive attention network to fuse the feature information of the embryonic channel and the parental channel through the interactive attention layer, generate context-aware fusion features, and output the risk probability of the ART offspring having circulatory system defects based on the context-aware fusion features.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-8.

Images