Embryo aneuploidy prediction method and device based on multi-modal fusion and electronic equipment

Abstract

The present application relates to the technical field of biomedical engineering, and discloses an embryo euploidy prediction method and device based on multi-modal fusion and electronic equipment, the method comprising: collecting cfDNA sequencing data, cfRNA expression data and time series image data of the embryo to be tested; extracting structural genetic feature vectors, functional molecular feature vectors and imaging representation vectors of the embryo to be tested based on the cfDNA sequencing data, cfRNA expression data and time series image data respectively; performing cross-modal feature fusion on the structural genetic feature vectors, functional molecular feature vectors and imaging representation vectors to obtain a fusion feature vector; and inputting the fusion feature vector into a pre-trained classification prediction model to obtain a euploidy prediction result of the embryo to be tested. The present application can improve the accuracy of embryo euploidy prediction.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, specifically to a method, device, and electronic device for predicting embryonic euploidy based on multimodal fusion. Background Technology

[0002] In clinical practice of assisted reproductive technologies (ART) such as in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI), the euploidy status of the embryo is a key factor affecting implantation success rate, pregnancy rate, and live birth rate. Generally, euploid embryos have significantly better developmental potential and clinical outcomes. Therefore, pre-transfer euploidy screening to prioritize the transfer of euploid embryos is an important means of improving the clinical success rate of ART.

[0003] Currently, the main clinical method for assessing embryonic euploidy is preimplantation genetic testing (PGT-A). This method typically involves obtaining a small number of cells from the trophectoderm (TE) through biopsy on days 5-6 of blastocyst development, followed by chromosome copy number analysis using whole-genome amplification (WGA) combined with next-generation sequencing (NGS). However, this method has revealed several technical limitations in clinical application. First, TE biopsy is an invasive procedure that may cause mechanical damage to the embryo or adversely affect its subsequent developmental potential. Second, due to the presence of embryonic mosaicism, the genetic status of TE cells may not fully represent the true state of the inner cell mass (ICM), leading to errors in euploidy determination. Furthermore, this method requires high-level laboratory conditions and skilled operators, and is expensive, increasing the economic cost and technical barriers of assisted reproduction.

[0004] To overcome the drawbacks of invasive biopsies, researchers have proposed the concept of non-invasive preimplantation genetic testing (niPGT-A). One technical approach involves analyzing extracellular cell-free nucleic acids (such as cell-free extracellular DNA, cfDNA) in the embryo culture medium to infer the euploidy status of the embryo. cfDNA originates from genomic fragments released by embryonic cells during apoptosis or cell renewal and theoretically can reflect abnormal chromosome copy number. However, in-depth analysis revealed that cfDNA content in the culture medium is extremely low, and it is highly fragmented, leading to a high amplification failure rate. Furthermore, it is susceptible to contamination from maternal DNA or culture medium background, resulting in insufficient consistency between cfDNA detection results and TE biopsy results. Another technical approach involves analyzing extracellular cell-free RNA (cfRNA) in the culture medium to assess the functional status of the embryo. The cfRNA expression patterns of aneuploid embryos may differ from those of euploid embryos in terms of cell cycle regulation and metabolic pathways, possessing the potential to serve as functional biomarkers. However, cfRNA exhibits even worse stability in the culture medium, and detection results are easily affected by degradation, background noise, and amplification bias. Its reliability for independent euploidy determination still requires large-scale clinical validation.

[0005] Furthermore, imaging and time-series developmental analysis based on time-difference imaging systems is also an important direction for non-invasive prenatal testing (NIPT). Continuous imaging and analysis of dynamic developmental indicators such as blastocyst expansion, mitotic rhythm, inner cell mass, and trophectoderm morphology can reflect the developmental dynamics of the embryo to some extent. However, the correlation between morphological parameters and chromosome euploidy is limited, and morphological assessment is easily affected by imaging conditions, analysis algorithms, and operator subjective judgment, making it difficult to use as an accurate sole criterion for euploidy determination.

[0006] Therefore, existing methods for predicting embryonic euploidy generally face problems such as limited information sources, insufficient robustness, and limited accuracy. Summary of the Invention

[0007] This invention provides a method, device, electronic device, storage medium, and computer program product for predicting embryonic euploidy based on multimodal fusion, in order to solve the problem of low accuracy in predicting embryonic euploidy.

[0008] In a first aspect, the present invention provides a method for predicting embryonic euploidy based on multimodal fusion, the method comprising: Collect cfDNA sequencing data, cfRNA expression data, and time-series image data from the embryos to be tested; Based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test are extracted respectively. Cross-modal feature fusion is performed on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector; The fused feature vector is input into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

[0009] In one optional implementation, cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryo to be tested are collected, including: cfDNA and cfRNA nucleic acid samples were extracted from the culture medium samples of the embryos to be tested collected at preset sampling time points. The cfDNA nucleic acid sample was subjected to whole-genome amplification and sequencing to obtain cfDNA sequencing data; The cfRNA nucleic acid sample was subjected to reverse transcription amplification and transcriptome sequencing to obtain cfRNA expression data; Using a time-difference imaging system, embryo images of the embryo under test are continuously acquired from a preset imaging start time point at preset time intervals until a preset prediction time point, to obtain time-series image data; wherein, each embryo image corresponds to an acquisition timestamp, and the preset prediction time point is equal to or later than the preset sampling time point.

[0010] In one optional implementation, cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryo to be tested are collected, including: cfDNA and cfRNA nucleic acid samples were extracted from the culture medium samples of the embryos to be tested collected at preset sampling time points. The cfDNA nucleic acid sample was subjected to whole-genome amplification and sequencing to obtain cfDNA sequencing data; The cfRNA nucleic acid sample was subjected to reverse transcription amplification and transcriptome sequencing to obtain cfRNA expression data; Using a time-difference imaging system, embryo images of the embryo under test are continuously acquired from a preset imaging start time point at preset time intervals until a preset prediction time point, to obtain time-series image data; wherein, each embryo image corresponds to an acquisition timestamp, and the preset prediction time point is equal to or later than the preset sampling time point.

[0011] In one optional implementation, based on the cfDNA sequencing data, the cfRNA expression data, and the time-series imaging data, the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test are extracted, including: Based on the cfDNA sequencing data, a structural genetic feature vector is extracted from the embryo to be tested; wherein, the structural genetic feature vector is used to characterize the difference between euploidy and aneuploidy; Based on the cfRNA expression data, the functional molecular feature vector of the embryo under test is extracted; wherein, the functional molecular feature vector is used to characterize the functional state of the embryo under test. Based on the time-series image data, the imaging representation vector of the embryo under test is extracted; wherein, the imaging representation vector is used to characterize the dynamic developmental features and spatial morphological features of the blastocyst of the embryo under test.

[0012] In one optional implementation, based on the cfDNA sequencing data, the structural genetic feature vector of the embryo to be tested is extracted, including: The cfDNA sequencing data is cleaned to obtain the normalized sequencing coverage of each chromosome in the embryo to be tested; wherein, the data cleaning includes sequentially performing adapter removal, quality filtering, reference genome alignment, repetitive sequence removal, GC content correction and sequencing coverage normalization. The deviation between the normalized sequencing coverage of each chromosome and the reference coverage of the corresponding chromosome in the standard euploid control sample was calculated to obtain the coverage deviation index for each chromosome. The deviation between the normalized sequencing coverage of each chromosome and the mean normalized coverage of the corresponding chromosome in the normal euploid control sample set is calculated to obtain the coverage abnormality index of each chromosome; wherein, the normal euploid control sample set includes multiple normal euploid control samples. The structural genetic feature vector of the embryo under test is obtained by combining the coverage deviation index, the coverage abnormality index, and the quality index of the cfDNA sequencing data of each chromosome; wherein the quality index includes at least one of the following: effective alignment rate, repetitive sequence ratio, genome coverage, coverage uniformity, average fragment length, and amplification quality score.

[0013] In one optional implementation, based on the cfRNA expression data, the functional molecular feature vector of the embryo to be tested is extracted, including: The cfRNA expression data were sequentially subjected to low-quality sequence filtering, reference transcriptome alignment, transcript quantification, batch effect correction, and expression level standardization to obtain a gene expression dataset. Logarithmically transform the expression levels of target genes in the gene expression dataset to obtain the target expression levels of each target gene. The target expression levels of target genes belonging to the same biological pathway are weighted and summed according to preset weights to obtain the pathway activity index of the corresponding biological pathway. The gene expression levels of each target gene and the pathway activity indicators of each biological pathway are combined to obtain the functional molecular feature vector of the embryo under test.

[0014] In one optional implementation, based on the time-series image data, the imaging representation vector of the embryo to be tested is extracted, including: The region of interest of each embryo image in the time series image data is preprocessed to obtain a target image sequence; wherein, the target image sequence includes target images corresponding to each embryo image; the preprocessing includes at least one of the following: background removal, noise suppression, position alignment, size unification and pixel value normalization. Feature extraction is performed on each target image to obtain the spatial morphological features of each target image; The spatial morphological features of each target image and its corresponding temporal location code are fused to obtain a spatiotemporal representation vector. Based on the time-series image data, the developmental dynamics parameter vector of the embryo under test is extracted; The spatiotemporal representation vector is concatenated with the developmental dynamics parameter vector to obtain the imaging representation vector of the embryo under test.

[0015] In one optional implementation, cross-modal feature fusion is performed on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector, including: The structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector are mapped to the latent space to obtain the target structural feature, the target molecular feature, and the target imaging feature; Structural feature encoding is added to the target structural features, molecular feature encoding is added to the target molecular features, and image feature encoding is added to the target imaging features, and an initial multimodal feature matrix is ​​constructed; Based on the initial multimodal feature matrix, cross-modal feature fusion is performed to obtain an enhanced feature matrix; wherein, the enhanced feature matrix includes enhanced structural feature vectors, enhanced molecular feature vectors, and enhanced imaging feature vectors; A fused feature vector is generated based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector.

[0016] In one optional implementation, based on the initial multimodal feature matrix, cross-modal feature fusion is performed to obtain an enhanced feature matrix, including: An effectiveness matrix is ​​generated based on the quality indices of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector; wherein the effectiveness matrix is ​​a binarized diagonal matrix. Based on the validity matrix, generate attention masks for each modality; Based on the attention mask and the initial multimodal feature matrix, an initial attention feature matrix is ​​generated; Multiply the effectiveness matrix by the initial attention feature matrix to obtain the target attention feature matrix; Using the effectiveness matrix, the residual connections and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as the output of the feedforward neural network, are modified to obtain the enhanced feature matrix.

[0017] In one optional implementation, the validity matrix is ​​used to mask invalid data in the residual connections and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as in the output of the feedforward neural network, to obtain an enhanced feature matrix, including: The initial enhancement feature matrix is ​​obtained by multiplying the effectiveness matrix by the residual concatenation and layer normalization results of the initial attention feature matrix and the target attention feature matrix; The enhanced feature matrix is ​​obtained by multiplying the validity matrix by the initial enhanced feature matrix and passing the result through the feedforward neural network.

[0018] In one optional implementation, a fused feature vector is generated based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector, including: Multiplying the enhanced feature matrix with the preset modality mapping weight vector yields the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the imaging modality contribution weight corresponding to the enhanced imaging feature vector. According to the genetic modality contribution weight, the molecular modality contribution weight, and the imaging modality contribution weight, the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector are weighted and aggregated to obtain a weighted aggregated feature vector; The enhanced structural feature vector, the enhanced molecular feature vector, the enhanced imaging feature vector, and the weighted aggregated feature vector are concatenated to obtain a fused feature vector.

[0019] In one optional implementation, the enhanced feature matrix is ​​multiplied by a preset modality mapping weight vector to obtain the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the imaging modality contribution weight corresponding to the enhanced imaging feature vector, including: Based on trainable modality weight mapping parameters, a modality mapping vector is generated; wherein, the modality mapping vector includes initial weights for genetic modalities, initial weights for molecular modalities, and initial weights for image modalities; Based on the quality indices of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector, a binary mask value vector is generated; wherein, the mask value vector is used to characterize the effectiveness of each modality; The result of adding the modality mapping vector and the mask value vector is normalized to obtain the genetic modality contribution weight, the molecular modality contribution weight, and the image modality contribution weight; wherein the sum of the genetic modality contribution weight, the molecular modality contribution weight, and the image modality contribution weight is 1.

[0020] In one optional implementation, the fused feature vector is input into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested, including: The fused feature vector is compressed to obtain a compressed fused feature vector; The compressed and fused feature vector is input into a preset classification model to obtain the euploidy prediction probability of the embryo under test; If the euploid prediction probability is greater than or equal to a preset threshold, the prediction result of the embryo to be tested is determined to be euploid.

[0021] In one optional implementation, the number of embryos to be tested is multiple, and the method further includes: The euploidy prediction probabilities of multiple embryos to be tested are sorted in descending order to generate an embryo prediction sequence; Output the embryo prediction sequence, and the embryo prediction result corresponding to each prediction probability in the embryo prediction sequence.

[0022] Secondly, the present invention provides an embryonic euploidy prediction device based on multimodal fusion, the device comprising: The data acquisition module is used to acquire cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryos to be tested; The feature extraction module is used to extract the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, respectively. The feature fusion module is used to perform cross-modal feature fusion on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector; The embryo prediction module is used to input the fused feature vector into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

[0023] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the multimodal fusion-based embryo euploidy prediction method described in the first aspect or any corresponding embodiment thereof.

[0024] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multimodal fusion-based embryo euploidy prediction method of the first aspect or any corresponding embodiment described above.

[0025] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the multimodal fusion-based embryo euploidy prediction method of the first aspect or any corresponding embodiment described above.

[0026] The embryo euploidy prediction method based on multimodal feature fusion provided in this invention extracts structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors from cfDNA sequencing data, cfRNA expression data, and time-series image data, respectively. It then performs cross-modal fusion of these feature vectors, resulting in a fused feature vector that simultaneously contains information on chromosome copy number variation, gene expression regulation, and morphological developmental dynamics. This comprehensively reflects the euploidy status of the embryo from genetic, molecular, and morphological dimensions. Compared to using only single-modal data, the multimodal fusion-based prediction method can more fully utilize the complementary information between different modalities, improving the accuracy and reliability of embryo euploidy prediction. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1This is a flowchart illustrating an embryo euploidy prediction method based on multimodal fusion provided in an embodiment of the present invention.

[0029] Figure 2 This is a flowchart illustrating the implementation of an embryonic euploidy prediction method based on multimodal fusion, provided in an embodiment of the present invention.

[0030] Figure 3 This is a structural block diagram of an embryo euploidy prediction device based on multimodal fusion provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0032] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0033] In the clinical practice of assisted reproductive technology, the euploidy status of embryos is one of the key biological factors determining the success rate of in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI). The implantation rate, clinical pregnancy rate, and live birth rate of euploid embryos are generally significantly higher than those of aneuploid embryos. Therefore, assessing the euploidy status of candidate embryos before embryo transfer provides an objective basis for clinical embryo selection and transfer decisions, and is an important step in improving the efficiency of assisted reproductive treatment.

[0034] Preimplantation genetic testing (PGT-A) is currently the main technique for clinically assessing the euploidy status of embryos. This technique typically involves obtaining a small number of cells from the trophectoderm (TE) through biopsy on days 5 to 6 of blastocyst development. After whole-genome amplification, chromosome copy number analysis is performed to determine whether the embryo is euploid.

[0035] However, after in-depth analysis, it was found that the PGT-A method based on TE biopsy has several limitations in clinical application and detection principles. Specifically, TE biopsy is an invasive procedure that may cause mechanical damage to the embryo or adversely affect its subsequent developmental potential. Furthermore, due to the prevalence of embryonic mosaicism, the genetic status of TE cells may not fully represent the true genetic composition of the inner cell mass, leading to some bias in euploidy determination. In addition, this method requires high-level laboratory conditions and skilled operators, and is expensive, posing both technical and economic challenges to its wider application in assisted reproductive centers.

[0036] To address the aforementioned technical challenges, researchers have proposed a non-invasive preimplantation genetic testing (niPGT-A) approach. One approach analyzes cell-free extracellular DNA (cfDNA) in the embryo culture medium to infer the euploidy status of the embryo. cfDNA originates from genomic fragments released during embryonic cell apoptosis or cell renewal and theoretically can directly reflect chromosomal copy number abnormalities. However, cfDNA concentrations in the culture medium are extremely low and highly fragmented, resulting in a high amplification failure rate. Furthermore, it is susceptible to contamination from maternal DNA or culture medium background, leading to insufficient consistency between the test results and TE biopsy results. Another approach analyzes cell-free extracellular RNA (cfRNA) in the culture medium to assess the functional molecular status of the embryo. Previous studies have suggested differences in cfRNA expression patterns between aneuploid and euploid embryos in pathways such as cell cycle regulation, energy metabolism, and apoptosis signaling. However, cfRNA exhibits even lower stability in the culture medium, and the test results are easily affected by degradation, background noise, and amplification bias. Its reliability as an independent criterion for euploidy has not yet been fully validated.

[0037] Furthermore, imaging and time-series developmental analysis based on time-difference imaging systems is also an important direction for non-invasive embryo assessment. By continuously imaging and extracting parameters from dynamic developmental indicators such as blastocyst expansion, cleavage rhythm, inner cell mass, and trophectoderm morphology, the developmental dynamics of the embryo can be reflected. However, due to the limited intrinsic correlation between morphological parameters and chromosome euploidy, and the susceptibility of morphological assessment to the influence of imaging conditions, analysis algorithms, and operator subjective judgment, relying solely on imaging analysis makes it difficult to achieve highly accurate euploidy determination.

[0038] It can be seen that current euploidy prediction schemes, whether relying solely on the structural genetic information provided by cfDNA, the functional molecular characteristics reflected by cfRNA, or the developmental dynamics revealed by imaging data, all face technical challenges such as limited information sources, insufficient robustness, and limited predictive accuracy. Single-modality detection methods, due to their inherent technical limitations, cannot independently meet the comprehensive clinical needs for accuracy, stability, and safety in euploidy determination.

[0039] Based on this, embodiments of the present invention provide a method for predicting embryonic euploidy based on multimodal feature fusion. This method extracts structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors from cfDNA sequencing data, cfRNA expression data, and time-series image data, respectively. The feature vectors from these three modalities are then fused across modalities. The resulting fused feature vector simultaneously contains information on chromosome copy number variation, gene expression regulation, and morphological developmental dynamics, comprehensively reflecting the euploidy state of the embryo from genetic, molecular, and morphological dimensions. Compared to using only single-modal data, the prediction method based on multimodal fusion can more fully utilize the complementary information between different modalities, improving the accuracy and reliability of embryonic euploidy prediction.

[0040] Please see Figure 1 This invention provides a method for predicting embryonic euploidy based on multimodal fusion.

[0041] The euploidy prediction method is a non-invasive assessment of the chromosomal euploidy status of embryos cultured in vitro in assisted reproductive technologies. Specifically, euploidy refers to an embryo with a normal number of chromosomes, while aneuploidy refers to an embryo with an abnormal number of chromosomes. Euploid embryos typically have higher implantation potential and better clinical pregnancy outcomes. The prediction method provided in this invention integrates multiple biological information from the embryo culture process non-invasively and utilizes an artificial intelligence model to automatically determine the euploidy status, providing an objective basis for embryo selection and implantation decisions.

[0042] Step S110: Collect cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryos to be tested.

[0043] In this embodiment, the embryo to be tested is an embryo for which euploidy status assessment is required during in vitro culture for assisted reproduction. Specifically, the embryo to be tested may be an embryo that has been cultured to the blastocyst stage in a culture medium after in vitro fertilization or intracytoplasmic sperm injection.

[0044] cfDNA sequencing data refers to the genomic sequence information obtained by amplifying and sequencing the entire genome of extracellular cell-free DNA in embryonic culture medium. cfDNA originates from genomic fragments released by embryonic cells during apoptosis or cell turnover and theoretically can reflect the chromosome copy number status of the embryo.

[0045] cfRNA expression data refers to gene expression information obtained by reverse transcription amplification and transcriptome sequencing of extracellular cell-free RNA in the same embryo culture medium. cfRNA can reflect the functional molecular state of the embryo at the transcriptional level, such as the activity of pathways related to cell cycle regulation, energy metabolism, and apoptosis.

[0046] Time-series imaging data refers to a set of embryonic images with acquisition timestamps obtained by continuously photographing the same embryo at preset time intervals during in vitro embryo culture using a time-difference imaging system. This imaging data can record the complete developmental process of the embryo from cleavage to blastocyst formation, reflecting dynamic developmental characteristics such as blastocyst expansion, division rhythm, inner cell mass, and trophectoderm morphology.

[0047] All three types of data are derived from the same embryo and are linked by a unique embryo identifier to ensure that different modal data correspond to the same individual embryo. Obtaining this data through non-invasive collection of embryo culture medium and continuous imaging avoids the potential mechanical damage caused by trophoblastic ectoderm biopsy, thus enabling non-invasive euploidy assessment of the embryo.

[0048] Step S120: Based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, extract the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo to be tested.

[0049] In this embodiment, the structural genetic feature vector is a feature vector extracted from cfDNA sequencing data to characterize the chromosomal structural abnormalities of the embryo under test. Specifically, the structural genetic feature vector may include coverage deviation indicators, coverage abnormality indicators, and sequencing data quality indicators for each chromosome. This feature vector can quantitatively characterize the differences between euploidy and aneuploidy at the genomic level, capturing structural genetic information such as chromosome copy number variations.

[0050] Functional molecular feature vectors (FMVs) are feature vectors extracted from cfRNA expression data to characterize the functional state of an embryo at the transcriptional level. Specifically, a FMV may include the expression levels of each gene in a pre-defined set of target genes, as well as pathway activity indicators obtained by weighted aggregation of the expression levels of multiple target genes within the same biological pathway. This feature vector can characterize the activity of biological processes in the embryo, such as cell cycle regulation, energy metabolism, and apoptosis signaling, at the molecular functional level.

[0051] The imaging representation vector is a feature vector extracted from time-series image data to characterize the dynamic developmental features and spatial morphological features of the blastocyst in the embryo under test. Specifically, the imaging representation vector can include a spatiotemporal representation vector learned from time-series images using a deep learning model, and an explicit developmental dynamics parameter vector calculated based on continuous images. This feature vector can simultaneously capture the spatial morphological information and temporal dynamic information of the embryo during development.

[0052] Step S130: Perform cross-modal feature fusion on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector.

[0053] In this embodiment, cross-modal feature fusion refers to the process of information interaction and aggregation of feature vectors from different modalities in a unified feature space. Specifically, structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors can be mapped to latent spaces of the same dimension, and information exchange and complementary enhancement between the features of different modalities can be achieved through a multimodal self-attention mechanism, thereby generating a fused feature vector containing multimodal complementary information.

[0054] Unlike simple splicing or summation, cross-modal feature fusion can establish associations between effective modalities, enabling fused features to comprehensively utilize chromosomal structural information provided by cfDNA, molecular functional state information reflected by cfRNA, and developmental dynamics information revealed by imaging data. At the same time, it can adaptively handle situations where some modalities are missing or the data quality does not meet the requirements.

[0055] Step S140: Input the fused feature vector into the pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

[0056] In this embodiment, the pre-trained classification prediction model refers to a machine learning model whose parameters have been optimized using training data. This model outputs euploidy prediction results based on the input fused feature vector. Specifically, the classification prediction model can be a binary classification model based on a deep neural network. During the model training phase, PGT-A detection results obtained from trophoblastic ectoderm biopsy can be used as supervisory labels for euploidy classification, and the model is trained using a classification cross-entropy loss function. After model training is complete, in the actual prediction phase, there is no need to perform a biopsy on the embryos to be tested. Only non-invasively acquired cfDNA sequencing data, cfRNA expression data, and time-series image data need to be input to output euploidy prediction results.

[0057] The euploidy prediction results can include the predicted probability that the embryo is euploid, the predicted probability that it is aneuploid, and the euploidy or aneuploidy classification result based on a preset classification threshold. When there are multiple candidate embryos for the same patient, they can also be ranked and selected according to the euploidy prediction probability of each candidate embryo, providing an auxiliary basis for clinical transplantation decisions.

[0058] In some cases, because single-modal data can only reflect the embryo’s state information in one aspect, it is difficult to fully capture the multidimensional characteristics of embryonic chromosomal abnormalities, resulting in insufficient accuracy in euploidy prediction.

[0059] The above embodiments can solve the technical problems of insufficient robustness and limited predictive accuracy of existing single-modality non-invasive detection methods due to the single information source. Specifically, the use of cfDNA detection alone is limited by the extremely low nucleic acid content in the culture medium, severe fragmentation, high amplification failure rate, and susceptibility to maternal background contamination, resulting in insufficient consistency between the detection results and TE biopsy results. The use of cfRNA detection alone is limited by the poor stability of cfRNA in the culture medium, its susceptibility to degradation, and its susceptibility to background noise interference. The use of imaging analysis alone faces the problem that the intrinsic correlation between morphological parameters and chromosome euploidy is limited, and morphological assessment is easily affected by imaging conditions, analysis algorithms, and operator subjective judgment.

[0060] Therefore, by fusing structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors across modalities, the feature information from different modalities can complement each other. When the data quality of a certain modality fluctuates due to amplification bias, background noise, or changes in imaging conditions, other effective modalities can still provide stable prediction information, thereby reducing the risk of misjudgment caused by noise or poor data quality in a single modality. Inputting the fused comprehensive feature vector into a pre-trained classification prediction model can improve the accuracy and robustness of euploidy prediction, achieving reliable determination of euploidy status while avoiding biopsy damage to embryos.

[0061] In some embodiments, step S110, collecting cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryo to be tested, may include the following steps.

[0062] Step S111: Extract cfDNA and cfRNA nucleic acid samples from the culture medium samples of the embryos to be tested collected at preset sampling time points.

[0063] Step S112: Perform whole-genome amplification and sequencing on the cfDNA nucleic acid sample to obtain cfDNA sequencing data.

[0064] Step S113: Perform reverse transcription amplification and transcriptome sequencing on the cfRNA nucleic acid sample to obtain cfRNA expression data.

[0065] Step S114: Using a time-difference imaging system, starting from a preset imaging start time point, continuously acquire embryo images of the embryo to be tested at preset time intervals until a preset prediction time point, to obtain time-series image data; wherein, each embryo image corresponds to an acquisition timestamp, and the preset prediction time point is equal to or later than the preset sampling time point.

[0066] In this embodiment, by obtaining cfDNA and cfRNA nucleic acid samples from the culture medium of the same embryo at the same preset sampling time point, and simultaneously using a time-difference imaging system to acquire continuous embryo images of the embryo, it can be ensured that the three types of data used for feature extraction and fusion analysis all come from the same individual embryo, thereby avoiding prediction bias caused by data mismatch between different embryos.

[0067] In this embodiment, the preset sampling time point refers to a pre-selected time point for collecting culture medium samples during the in vitro embryo culture process. Specifically, the preset sampling time point can be set according to the embryonic development stage and clinical operating procedures. The preset sampling time point can be any time point on day 3, day 5, or day 6 of embryo culture. In one specific embodiment, the preset sampling time point is set to day 5 of embryo culture, at which time the embryo has usually developed to the blastocyst stage, and a sufficient amount of extracellular free nucleic acid has accumulated in the culture medium. At the same time, the time-difference imaging system has completed the key developmental imaging recording of the early cleavage to blastocyst formation process of the embryo.

[0068] A culture medium sample refers to a sample of the liquid culture medium in which the embryo under test is cultured in vitro. Because embryos release extracellular cell-free nucleic acids into the culture medium during culture through apoptosis, renewal, and secretion, the culture medium sample contains cfDNA and cfRNA derived from the embryo under test. Linking the culture medium sample to the corresponding embryo under test using a unique embryo identifier ensures that subsequent nucleic acid and imaging data correspond to the same individual embryo.

[0069] cfDNA nucleic acid samples are extracellular cell-free DNA samples extracted from culture medium samples. Specifically, the culture medium sample can be divided into a first subsample and a second subsample. cfDNA nucleic acid samples are extracted from the first subsample, and cfRNA nucleic acid samples are extracted from the second subsample. Alternatively, a nucleic acid extraction method capable of simultaneously extracting DNA and RNA from the same culture medium sample can be used to obtain cfDNA and cfRNA nucleic acid samples corresponding to the same embryo under test. The cfDNA nucleic acid samples are used for subsequent whole-genome amplification and sequencing to obtain structural genetic information reflecting the chromosome copy number status of the embryo.

[0070] cfRNA nucleic acid samples are extracellular cell-free RNA samples extracted from culture medium samples. cfRNA nucleic acid samples are used for subsequent reverse transcription amplification and transcriptome sequencing to obtain information reflecting the functional molecular state of the embryo at the transcriptional level.

[0071] Whole-genome amplification and sequencing of cfDNA nucleic acid samples yield cfDNA sequencing data. Whole-genome amplification is a technique used for unbiased amplification of cfDNA nucleic acid samples with very low starting amounts. By amplifying cfDNA nucleic acid samples into a whole genome, trace amounts of cfDNA can be amplified to the amount required for sequencing library construction. Sequencing can be low-coverage whole-genome sequencing, and high-throughput sequencing of the amplified products yields the genomic sequence information of the embryo under test. cfDNA sequencing data reflects the sequencing coverage distribution of each chromosome in the embryo under test, providing a data foundation for subsequent extraction of structural genetic feature vectors.

[0072] cfRNA nucleic acid samples were subjected to reverse transcription amplification and transcriptome sequencing to obtain cfRNA expression data. Reverse transcription amplification was used to reverse transcribe RNA molecules in the cfRNA nucleic acid samples into complementary DNA, and then amplified the complementary DNA. Transcriptome sequencing was performed on the amplified complementary DNA using high-throughput sequencing to obtain gene expression information of the embryos under test. cfRNA expression data can reflect the gene expression profile of the embryos under test at the predetermined sampling time point, providing a data foundation for subsequent extraction of functional molecular feature vectors.

[0073] Time-lapse imaging systems are imaging devices used for continuous, non-invasive imaging of embryos during in vitro embryo culture. These systems are typically integrated into or used in conjunction with an embryo incubator, automatically acquiring microscopic images of embryos at preset time intervals without affecting the embryo culture environment. The preset imaging start point refers to the point at which the time-lapse imaging system begins continuous imaging of the embryo being examined. This preset start point can be the moment the embryo is placed in the time-lapse imaging incubator after fertilization, or a specific time point during embryo culture.

[0074] The preset time interval refers to the time interval between two consecutive frames of embryo images. The preset time interval can be set according to the embryonic development speed and the required temporal resolution. In one specific embodiment, the preset time interval can be set to acquire images every 10 minutes, thereby obtaining sufficiently dense temporal image data over several days of blastocyst development to capture key developmental events such as cleavage, compaction, blastocyst formation, blastocyst expansion, and collapse.

[0075] The preset prediction time point refers to the pre-defined time point for performing euploidy prediction. The preset prediction time point is no earlier than the preset sampling time point to ensure that complete culture medium nucleic acid data and time-series image data have been acquired by the time prediction is performed. In one specific embodiment, the preset sampling time point is set to day 5 of embryo culture, and the preset prediction time point is set to day 6 of embryo culture. Thus, the time-difference imaging system acquires embryo images every 10 minutes starting from the preset imaging start time point, continuing until day 6 of embryo culture, thereby obtaining complete late-stage blastocyst development image data.

[0076] The acquisition timestamp is used to record the actual acquisition time of each embryo image. The acquisition timestamp serves as the temporal basis for subsequent time series analysis and temporal location encoding. Since the developmental rhythms of different embryos may vary, the acquisition timestamp allows for accurate reconstruction of the position of each image frame on the developmental timeline, providing a precise temporal reference for extracting developmental dynamics parameters and constructing spatiotemporal representation vectors.

[0077] Through the above embodiments, it is possible to achieve synchronous and standardized acquisition of cfDNA sequencing data, cfRNA expression data and time-series image data under a unified preset sampling time point and preset prediction time point framework. This ensures that the three modalities of data correspond to the same embryo to be tested, and that the three have a clear temporal relationship in the time dimension, laying a data foundation for subsequent feature extraction, multimodal fusion and euploidy prediction.

[0078] Please see Figure 2In some embodiments, step S120, based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, extracts the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo to be tested, which may include the following steps.

[0079] S121: Based on the cfDNA sequencing data, extract the structural genetic feature vector of the embryo to be tested; wherein the structural genetic feature vector is used to characterize the difference between euploidy and aneuploidy.

[0080] S122: Based on the cfRNA expression data, extract the functional molecular feature vector of the embryo to be tested; wherein the functional molecular feature vector is used to characterize the functional state of the embryo to be tested.

[0081] S123: Based on the time-series image data, extract the imaging representation vector of the embryo to be tested; wherein, the imaging representation vector is used to characterize the dynamic developmental features and spatial morphological features of the blastocyst of the embryo to be tested.

[0082] In this embodiment, feature extraction is performed on the raw data of the same embryo from three different modalities, transforming the raw high-throughput sequencing data and image data into structured feature vectors that can be used as input for machine learning models. These three feature vectors characterize the same embryo from different biological dimensions, providing complementary information sources for subsequent multimodal feature fusion.

[0083] Structural genetic feature vectors (SGMs) are feature vectors extracted from cfDNA sequencing data to characterize the differences between euploidy and aneuploidy. Specifically, SGMs can quantify the differences between euploidy and aneuploidy at the chromosome level. Euploid embryos have a normal chromosome number, and the sequencing coverage of each chromosome is similar to that of normal control samples. Aneuploid embryos, on the other hand, have abnormal chromosome numbers, such as trisomy or monosomy, causing the sequencing coverage of that chromosome to deviate from that of normal control samples. By calculating the deviation between the normalized coverage of each chromosome and that of normal control samples, SGMs can transform these chromosome-level structural abnormalities into numerical feature representations, thereby providing classification and prediction models with discriminative information directly related to chromosome copy number abnormalities.

[0084] Functional molecular feature vectors (FMVs) are feature vectors extracted from cfRNA expression data to characterize the functional state of an embryo under test. Specifically, FMVs can characterize the biological activity state of an embryo at the molecular functional level. The functional state of an embryo involves the activity of multiple biological pathways, including cell cycle regulation, energy metabolism, and apoptosis. Due to chromosomal dosage effects and disordered gene expression regulatory networks, aneuploid embryos typically exhibit functional differences compared to euploid embryos. By extracting the expression levels of target genes and integrative pathway activity indicators, FMVs can capture these functional differences at the transcriptional level, thereby providing classification and prediction models with discriminative information reflecting the normality or absence of embryonic molecular function.

[0085] Imaging representation vectors are feature vectors extracted from time-series image data to characterize the dynamic developmental features and spatial morphological features of the embryo's blastocyst stage. Specifically, imaging representation vectors can simultaneously characterize the developmental process of the embryo from both temporal and spatial dimensions. Dynamic developmental features refer to the temporal developmental behaviors and rhythmic characteristics exhibited by the embryo throughout its in vitro culture process from fertilized egg to blastocyst, such as cleavage time, cleavage synchronicity, compaction time, blastocyst formation time, blastocyst expansion rate, and collapse and re-expansion behavior. Spatial morphological features refer to the static morphological attributes of the embryo at a specific time point, such as the size and shape of the blastocyst, the area and compactness of the inner cell mass, and the thickness and integrity of the trophectoderm. By integrating these dynamic developmental features and spatial morphological features into a single vector representation, imaging representation vectors can comprehensively reflect the developmental dynamics of the embryo, thereby providing discriminative information related to embryonic developmental quality and potential for classification and prediction models.

[0086] Through the above embodiments, the extraction of the three types of feature vectors can transform the original heterogeneous data of different modalities into feature representations with a unified mathematical form, laying the foundation for subsequent cross-modal feature fusion. Furthermore, each type of feature vector represents the embryo under test from three complementary dimensions: genetic structure, molecular function, and developmental morphology. This allows the fused feature vector to comprehensively utilize multi-source information, effectively overcoming the problem of insufficient prediction accuracy caused by limited signal content or noise interference from a single modal information source.

[0087] In some embodiments, step S121, based on the cfDNA sequencing data, extracts the structural genetic feature vector of the embryo to be tested, which may include the following steps.

[0088] Step S1211: Perform data cleaning on the cfDNA sequencing data to obtain the normalized sequencing coverage of each chromosome in the embryo to be tested; wherein, the data cleaning includes sequentially performing adapter removal, quality filtering, reference genome alignment, repetitive sequence removal, GC content correction and sequencing coverage normalization processing.

[0089] Step S1212: Calculate the deviation between the normalized sequencing coverage of each chromosome and the reference coverage of the corresponding chromosome in the standard euploid control sample to obtain the coverage deviation index for each chromosome.

[0090] Step S1213: Calculate the deviation between the normalized sequencing coverage of each chromosome and the mean normalized coverage of the corresponding chromosome in the normal euploid control sample set, to obtain the coverage abnormality index of each chromosome; wherein, the normal euploid control sample set includes multiple normal euploid control samples.

[0091] Step S1214: Combine the coverage deviation index, the coverage abnormality index, and the quality index of the cfDNA sequencing data of each chromosome to obtain the structural genetic feature vector of the embryo to be tested; wherein the quality index includes at least one of the following: effective alignment rate, repetitive sequence ratio, genome coverage, coverage uniformity, average fragment length, and amplification quality score.

[0092] In this embodiment, by systematically cleaning the cfDNA sequencing data, quantifying the two-dimensional chromosome coverage deviation, and integrating it with sequencing quality indicators, the raw sequencing data can be transformed into feature vectors with clear biological and statistical significance, thereby providing a robust structural genetic discrimination basis for classification prediction models.

[0093] Data cleaning is a process used to eliminate technical noise and systematic biases in cfDNA sequencing data, obtaining normalized coverage information that accurately reflects the copy number status of chromosomes in the embryo being tested. Data cleaning includes, in sequence, adapter removal, quality filtering, reference genome alignment, repetitive sequence removal, GC content correction, and sequencing coverage normalization. Adapter removal removes artificial adapter sequences introduced during sequencing library construction to avoid interference with subsequent alignments. Quality filtering eliminates low-quality reads, retaining reliable reads with high base quality scores. Reference genome alignment maps the quality-filtered reads to their corresponding positions in the human reference genome, determining the chromosomal origin and genomic coordinates of each read. Repetitive sequence removal identifies and removes repetitive reads caused by amplification biases during whole-genome amplification, ensuring that coverage calculations more accurately reflect the abundance of the original cfDNA template. GC content correction corrects for sequencing depth biases caused by differences in GC content across different genomic regions, making coverage comparable across different regions. Sequencing coverage normalization is used to globally normalize the coverage data after the above processing, eliminating systematic bias caused by differences in the total sequencing volume between different samples.

[0094] Normalized sequencing coverage refers to the normalized value of the number of effective aligned sequences per unit length in each chromosome or genome window after data cleaning and normalization. Normalized sequencing coverage reflects the relative copy number level of each chromosome in the embryo being tested and serves as the data basis for subsequent calculations of coverage deviation and abnormality indicators.

[0095] A standard euploid control sample refers to a single control sample confirmed as euploid by PGT-A testing and whose sequencing data quality meets preset standards. The reference coverage of a standard euploid control sample refers to the normalized sequencing coverage of that control sample on each chromosome. For chromosomes 1 to 22, the normalized coverage of the corresponding standard euploid control sample is used as the reference coverage. For the X and Y chromosomes, the candidate sex chromosome composition is determined based on the X and Y chromosome coverage ratio, and the reference coverage is calculated using a standard euploid control sample that matches the candidate sex chromosome composition to avoid instability in reference coverage calculations due to mismatches in sex chromosome composition.

[0096] Coverage deviation index is used to characterize the relative deviation of the normalized sequencing coverage of each chromosome in the tested embryo from the reference coverage of the corresponding chromosome in the standard euploid control sample. In one specific implementation, for the first chromosome... The coverage of the chromosomes deviates from the target. It can be calculated using the following formula: ,in, Indicates the embryo to be tested, number 1 Normalized sequencing coverage of chromosomes, This indicates the reference coverage of the chromosome corresponding to the standard euploid control sample. The coverage deviation index can intuitively reflect the direction and magnitude of the deviation of the copy number of each chromosome of the tested embryo from a single euploid reference. For example, when the normalized coverage deviation value of chromosome 16 of a tested embryo is 0.46, it indicates that the sequencing coverage of this chromosome has a large positive deviation from the standard euploid control sample, which may suggest the presence of trisomy in this chromosome.

[0097] The normal euploidy control sample set refers to a collection of control samples that have been confirmed as euploid by PGT-A testing and whose sequencing data quality meets preset standards. The normal euploidy control sample set includes multiple normal euploidy control samples. By including multiple normal euploidy control samples, a statistical distribution model of the normalized coverage of each chromosome can be established, providing a reference baseline for calculating statistically significant coverage abnormality indicators.

[0098] Coverage anomaly indices are used to characterize the degree of deviation of the normalized sequencing coverage of each chromosome in the tested embryo from the statistical distribution of a normal euploid control population. In one specific implementation, for the first chromosome... Chromosomes, with abnormal coverage indicators It can be calculated using the following formula: ,in, This indicates the first normal euploid control sample set. The mean of normalized coverage of chromosomes. This indicates the first normal euploid control sample set. The standard deviation of normalized chromosome coverage. Coverage abnormality index, measured in standard deviation, measures the degree of deviation and provides a statistically significant reference for judging chromosome copy number abnormalities. For example, when the normalized coverage abnormality score of chromosome 16 of a tested embryo is 3.2, it indicates that the coverage of this chromosome deviates from the mean of the normal euploid population by 3.2 standard deviations, suggesting the existence of a significant chromosome copy number abnormality.

[0099] Quality metrics for cfDNA sequencing data are a set of parameters used to evaluate the overall quality and reliability of cfDNA sequencing data. Quality metrics include at least one of the following: effective alignment rate, repetitive sequence ratio, genome coverage, coverage uniformity, average fragment length, and amplification quality score. Effective alignment rate refers to the proportion of reads successfully aligned to the reference genome out of all reads, reflecting the usability of the sequencing data. Repetitive sequence ratio refers to the proportion of reads marked as repetitive out of all reads, reflecting the degree of amplification bias introduced during whole-genome amplification. Genome coverage refers to the proportion of the reference genome region covered by at least one read out of the entire reference genome, reflecting the comprehensiveness of the sequencing. Coverage uniformity measures the consistency of coverage between different genomic windows. Average fragment length refers to the average length of the fragmented cfDNA. Amplification quality score is used to comprehensively evaluate the quality of whole-genome amplification.

[0100] Structural genetic feature vectors are formed by combining coverage deviation indicators, coverage anomaly indicators, and quality indicators of cfDNA sequencing data. In one specific implementation, the structural genetic feature vector... It can be represented as: Of these, 24 chromosomes include chromosomes 1 to 22, as well as the X and Y chromosomes. to These represent the effective alignment rate, the proportion of repetitive sequences, genome coverage, coverage uniformity, average fragment length, and amplification quality score, respectively. This yields a 54-dimensional structural genetic feature vector.

[0101] Through the above embodiments, the structural genetic feature vector not only includes a two-dimensional deviation index reflecting the copy number deviation of each chromosome, but also incorporates quality information characterizing the reliability of sequencing data. This enables the classification prediction model to simultaneously consider the magnitude of chromosomal abnormalities, statistical significance, and data quality reliability when using structural genetic features to determine euploidy, thereby improving the accuracy and robustness of the prediction.

[0102] In some embodiments, step S122, extracting the functional molecular feature vector of the embryo to be tested based on the cfRNA expression data, may include the following steps.

[0103] Step S1221: The cfRNA expression data are sequentially subjected to low-quality sequence filtering, reference transcriptome alignment, transcript quantification, batch effect correction, and expression level standardization to obtain a gene expression data set.

[0104] Step S1222: Perform a logarithmic transformation on the expression levels of the target genes in the gene expression dataset to obtain the target expression levels of each target gene.

[0105] Step S1223: The target expression levels of target genes belonging to the same biological pathway are weighted and summed according to preset weights to obtain the pathway activity index of the corresponding biological pathway.

[0106] Step S1224: Combine the gene expression levels of each target gene and the pathway activity indicators of each biological pathway to obtain the functional molecular feature vector of the embryo to be tested.

[0107] In this embodiment, by systematically preprocessing cfRNA expression data, extracting target gene expression levels, aggregating pathway activity, and combining features, the raw transcriptome sequencing data can be transformed into feature vectors that combine single-gene resolution and pathway functional integration, thereby providing a multi-level discriminative basis for classification and prediction models that reflects the molecular functional status of the embryo.

[0108] The gene expression dataset is a corrected and normalized set of gene expression data obtained by sequentially processing cfRNA expression data through low-quality sequence filtering, reference transcriptome alignment, transcript quantification, batch effect correction, and expression level normalization. Low-quality sequence filtering removes low-quality reads and retains reliable reads with high base quality scores. Reference transcriptome alignment maps the quality-filtered reads to their corresponding transcript positions in the reference transcriptome. Transcript quantification counts the number of valid aligned sequences for each transcript and characterizes the expression level of each gene using metrics such as TPM. Batch effect correction eliminates systematic expression biases caused by experimental factors such as different sequencing batches, different operators, or different reagent batches, ensuring comparability of expression data from different batches. Expression level normalization normalizes the sequencing depth of different samples, eliminating differences in expression level scales due to variations in total sequencing volume. Through this process, a gene expression dataset that accurately reflects the functional state of the transcriptome in the tested embryo can be obtained.

[0109] Target genes are a set of genes pre-determined based on cfRNA expression data of training samples during the model training phase, which are significantly associated with euploidy status determination. Specifically, the target gene set is determined by using only embryo samples with clear euploidy supervision tags from the training data. Target genes are screened from all candidate genes based on gene expression differences between euploid and aneuploid embryos, expression stability, and the degree of association between target genes and cell cycle regulation, energy metabolism, and apoptosis pathways. In one specific implementation, differential expression analysis and expression quality screening are first performed on each candidate gene in the training data. Then, target genes are selected based on preset statistical significance thresholds, fold change thresholds, minimum detection ratios, or feature importance scores, forming a pre-determined set of molecular markers. Once this set of molecular markers is determined, it remains unchanged during the model validation and actual prediction phases. Target genes are not re-screened using euploid tags from validation samples, test samples, or embryos to be tested, to avoid introducing prediction bias due to data leakage.

[0110] The target expression level is the numerical value obtained by logarithmically transforming the expression level of the target gene. In one specific implementation, for the first... The target gene, its target expression level after logarithmic transformation. It can be represented as: ,in, Indicates the first The expression levels of the target genes were corrected for sequencing depth and transcript length. Logarithmic transformation can compress the dynamic range of expression levels, reduce the dominance of highly expressed genes on the feature space, and make the expression distribution closer to a normal distribution, which is beneficial for subsequent model training and feature fusion. The increment operation is used to avoid... The problem of undefined logarithmic calculation when the logarithm is 0.

[0111] Biological pathways are biological processes involving the coordinated participation of multiple functionally related genes. For example, the cell cycle pathway involves multiple genes regulating cell division, the energy metabolism pathway involves multiple genes involved in oxidative phosphorylation and glycolysis, and the apoptosis pathway involves multiple genes regulating programmed cell death. Due to gene dosage effects caused by abnormal chromosome copy number, aneuploid embryos may exhibit differences in the overall activity of pathways such as cell cycle regulation, energy metabolism, and apoptosis signaling compared to euploid embryos. By integrating the expression levels of multiple target genes within a pathway into a pathway-level activity indicator, these differences can be captured at a higher functional level, while simultaneously reducing the impact of individual gene expression noise on prediction results.

[0112] Preset weights are used to characterize the contribution of each target gene to pathway activity within the same biological pathway. Preset weights can be set based on the known functional importance of the target gene in the corresponding biological pathway, gene expression stability, or weight parameters learned during model training. After determining the weights for each target gene during model training, they remain fixed during model validation and actual prediction.

[0113] Pathway activity indicators are numerical values ​​used to characterize the overall activity status of a corresponding biological pathway, obtained by weighting and summing the target expression levels of target genes belonging to the same biological pathway according to preset weights. In one specific implementation, for the first... A biological pathway, and its pathway activity indicators It can be calculated using the following formula: ,in, Indicates the first The set of target genes corresponding to each biological pathway Indicates target gene The preset weights in this pathway or the weights obtained from model training.

[0114] Functional molecular feature vectors are formed by combining the target expression levels of each target gene with pathway activity indicators of each biological pathway. In one specific implementation, 20 cell cycle-related target genes, 15 energy metabolism-related target genes, and 10 apoptosis-related target genes are pre-screened based on model training data. These 45 target genes are then consistently used during the model validation and actual prediction phases to calculate pathway activity indicators for the cell cycle, energy metabolism, and apoptosis pathways, respectively. , and This results in the following functional molecular feature vectors: This yields a 48-dimensional functional molecular feature vector.

[0115] Through the above embodiments, the functional molecular feature vector not only retains the transcriptional expression information of a single target gene, but also introduces a higher level of biological functional state representation through pathway activity aggregation. This enables the classification prediction model to capture the functional molecular features of the embryo under test at both the gene and pathway levels, thereby improving the ability to distinguish between euploid and aneuploid embryos and enhancing the robustness of the model in the face of fluctuations in the expression of a single gene.

[0116] In some embodiments, step S123, extracting the imaging representation vector of the embryo under test based on the time-series image data, may include the following steps.

[0117] Step S1231: Preprocess the region of interest of each embryo image in the time series image data to obtain a target image sequence; wherein, the target image sequence includes target images corresponding to each embryo image; the preprocessing includes at least one of the following: background removal, noise suppression, position alignment, size unification and pixel value normalization.

[0118] Step S1232: Extract features from each target image to obtain the spatial morphological features of each target image.

[0119] Step S1233: Perform spatiotemporal feature fusion on the spatial morphological features of each target image and its corresponding temporal location code to obtain a spatiotemporal representation vector.

[0120] Step S1234: Based on the time-series image data, extract the developmental dynamics parameter vector of the embryo to be tested.

[0121] Step S1235: Concatenate the spatiotemporal representation vector with the developmental dynamics parameter vector to obtain the imaging representation vector of the embryo to be tested.

[0122] In this embodiment, by employing a dual-branch extraction strategy that combines deep learning spatiotemporal features with explicit developmental dynamics parameters, it is possible to simultaneously capture implicit, high-dimensional spatiotemporal representation information and explicit developmental dynamics indicators with clear biological meaning from time-series image data. This provides classification and prediction models with multi-level imaging features that comprehensively reflect the embryonic developmental state.

[0123] The region of interest (ROI) refers to the image region containing the main embryo in each embryo image. Since the raw images captured by the time-lapse imaging system may contain large background areas, culture dish edges, or non-target debris, identifying the ROI containing the main embryo allows subsequent processing to focus on the morphological changes of the embryo itself, reducing background interference and computational redundancy. Preprocessing includes at least one of the following: background removal, noise suppression, position alignment, size unification, and pixel value normalization. Background removal eliminates interference from the culture medium background or non-embryo areas on feature extraction. Noise suppression reduces random noise during imaging, improving the image signal-to-noise ratio. Position alignment adjusts the main embryo in images acquired at different time points to a relatively fixed position in the image, eliminating spatial displacement caused by embryo movement or rotation in the culture medium. Size unification adjusts each frame of the image to the same pixel size to meet the consistency requirements of the input size for subsequent deep learning models.

[0124] Pixel value normalization is used to adjust the pixel values ​​of each frame of the image to a uniform range, thereby reducing the impact of image brightness, contrast, and device differences on the model. In one specific implementation, pixel value normalization can be performed in the following way: ,in, Represents the original pixel value. and These represent the mean and standard deviation of the image pixel values ​​in the model training dataset, respectively. Once determined during model training, these mean and standard deviation remain fixed during model validation, testing, and actual prediction to ensure consistency across different batches of data input.

[0125] The target image sequence refers to a standardized sequence of embryo images obtained by performing the aforementioned preprocessing on the region of interest of each embryo image in the time-series image data. Each target image in the target image sequence is an image that contains only the main embryo region and has a uniform format and numerical range after background removal, noise suppression, position alignment, size unification, and pixel value normalization.

[0126] Spatial morphological features refer to feature vectors extracted from a single target image to characterize the spatial morphological attributes of the embryo at that specific time point. Specifically, each target image can be input into a convolutional neural network, and the hierarchical visual features in the image can be automatically extracted using the multi-layer convolution and pooling operations of the convolutional neural network. In one specific implementation, for the first... Target images at each time point Its corresponding spatial morphological features It can be represented as: ,in, It is a 256-dimensional spatial morphological feature vector. Spatial morphological features can capture spatial visual information of the embryo at a single time point, such as cell outlines, inner cell mass morphology, trophectoderm structure, and blastocoel size.

[0127] Temporal position encoding is used to encode the positional information of each target image on the developmental timeline into a vector representation with the same dimension as the spatial morphological features. Since spatial morphological features themselves do not contain temporal sequence information, simply relying on combinations of spatial morphological features cannot effectively distinguish morphological differences between different developmental stages. By adding corresponding temporal position encoding to the spatial morphological features of each target image, the model can perceive the temporal relationship between images. In one specific implementation, temporal position encoding... In order to match the spatial morphological characteristics Vectors of the same dimension, such as a 256-dimensional vector. Temporal location encoding can be generated based on the actual acquisition timestamps of each frame of imagery to accurately reflect the time intervals between images.

[0128] Spatiotemporal feature fusion refers to the process of fusing the spatial morphological features of each time point with their corresponding temporal location codes to obtain a unified feature representation that simultaneously contains spatial morphological information and temporal developmental information. In one specific implementation, inputting the spatial morphological features and their temporal location codes of each time point into the Transformer model for spatiotemporal feature fusion can be represented as follows: ,in, Indicates the first Spatial morphological features of the target image frame. Indicates the first Temporal location encoding of the target image in a frame. Indicates a time validity mask. This indicates the length of the image sequence after uniform length processing. This represents the time-series feature matrix. By adding the spatial morphological features to the temporal location encoding and inputting the result into the Transformer, the model can automatically learn the long-range dependencies between images at different time points using the Transformer's self-attention mechanism, thereby effectively fusing the temporal dynamic information throughout the entire developmental process.

[0129] To address the issue of inconsistent effective image frame counts among different embryos due to differences in culture start time, occasional missed shots by the imaging system, or substandard image quality, methods such as resampling at equal time intervals, truncation, or zero-padding can be used to unify the image sequence to a preset length. Simultaneously, a time validity mask vector is constructed. This is used to mark whether the image frame corresponding to each time position is a real and valid frame. When the... When each position is a real and valid image frame When the first When a position is a padding frame or an invalid image frame During the self-attention calculation process, the Transformer uses this temporal validity mask to mask padded or invalid image frames, making their corresponding attention weights close to 0, thereby reducing the impact of padded frames on the time series feature extraction results.

[0130] The spatiotemporal representation vector is a fixed-dimensional vector extracted from the time-series feature matrix through pooling or classification labeling, used to comprehensively represent the spatiotemporal features of the entire time-series image. In one specific implementation, the time-series feature matrix is ​​subjected to masked average pooling based on a time validity mask, resulting in the spatiotemporal representation vector. It can be represented as: ,in, Represents the first element in the time series feature matrix. The features corresponding to the rows, This represents a pre-defined positive stable term to prevent the denominator from being zero. Therefore, padding frames or invalid image frames do not participate in the effective calculation of the spatiotemporal representation vector, resulting in... It is a 256-dimensional spatiotemporal representation vector.

[0131] The developmental kinetic parameter vector is an explicit developmental parameter vector with clear biological meaning, directly calculated from time-series image data. Specifically, several explicit radiological and developmental kinetic parameters can be calculated by analyzing the time points and magnitudes of morphological changes in embryos within consecutive images. In one specific implementation, the developmental kinetic parameter vector... The following parameters can be included: first cleavage time (i.e., time to reach the 2-cell stage), second cleavage time (i.e., time to reach the 4-cell stage), morula formation time, blastocyst initiation time, complete blastocyst formation time, blastocyst expansion rate, maximum collapse amplitude, post-collapse expansion rate, inner cell mass area change rate, and trophectoderm integrity score. The developmental kinetic parameter vector can be represented as: This yields a 10-dimensional vector of developmental kinetic parameters. These parameters directly reflect the time points and dynamic changes of key developmental events from cleavage to blastocyst formation, providing a biologically interpretable basis for developmental kinetic discrimination in classification and prediction models.

[0132] The imaging representation vector is formed by concatenating the spatiotemporal representation vector with the developmental dynamics parameter vector. In one specific implementation, It is a 256-dimensional spatiotemporal representation vector. The 10-dimensional developmental dynamics parameter vector, after being concatenated, yields a 266-dimensional imaging representation vector: .

[0133] Through the dual-branch parallel extraction strategy described in the above embodiments, the radiographic representation vector simultaneously integrates high-dimensional spatiotemporal latent features automatically learned by the deep learning model, as well as explicit developmental dynamic parameters with clear biological meaning. The spatiotemporal representation vector captures complex developmental patterns that may be difficult for human experts to quantify directly, while the developmental dynamic parameter vector provides indicators of key developmental events with clinical interpretability. The two complement each other, jointly providing a rich and robust radiographic information foundation for subsequent multimodal feature fusion and euploidy prediction.

[0134] In some embodiments, step S130 involves cross-modal feature fusion of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector, which may include the following steps.

[0135] Step S131: Map the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to the latent space respectively to obtain the target structural feature, the target molecular feature, and the target imaging feature.

[0136] Step S132: Add structural feature encoding to the target structural features, add molecular feature encoding to the target molecular features, add image feature encoding to the target imaging features, and construct an initial multimodal feature matrix.

[0137] Step S133: Based on the initial multimodal feature matrix, perform cross-modal feature fusion to obtain an enhanced feature matrix; wherein the enhanced feature matrix includes enhanced structural feature vectors, enhanced molecular feature vectors, and enhanced imaging feature vectors.

[0138] Step S134: Generate a fused feature vector based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector.

[0139] In this embodiment, by first mapping three modal features with different dimensions and distributions to a unified latent space, and then constructing an initial multimodal feature matrix that can distinguish modal origins by adding modality type encoding, and then obtaining enhanced features by cross-modal information interaction based on this matrix, and finally generating a fused feature vector based on the enhanced features, it is possible to achieve deep integration of the three modal features at the feature level, so that the fused features simultaneously contain the discrimination information of each modality itself and the complementary information generated by the interaction between modalities.

[0140] Latent space refers to a feature representation space with a unified dimension. Since structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors originate from different detection platforms and analysis processes, their feature dimensions, numerical ranges, and statistical distributions differ. For example, in one specific implementation, the structural genetic feature vector is 54-dimensional, the functional molecular feature vector is 48-dimensional, and the imaging representation vector is 266-dimensional. If a fusion operation is directly performed on these feature vectors with different dimensions, the features of different modalities will occupy unequal representation spaces during the fusion process, affecting the fusion effect.

[0141] Mapping structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors to latent space refers to the process of mapping three different-dimensional input features to a latent space representation of the same dimension using their respective modal encoders. In one specific implementation, this mapping process can be represented as: , , ,in, , and These represent the input structural genetic feature vector, functional molecular feature vector, and imaging representation vector, respectively. , and These represent the mapping matrices for genetic modalities, molecular modalities, and imaging modalities, respectively. , and Indicates the corresponding bias term; represents a nonlinear activation function, used to introduce nonlinear transformation capabilities; LN represents a layer normalization operation, used to normalize the mapped features, accelerating model training and improving stability. Through the above mapping, 54-dimensional structural genetic features, 48-dimensional functional molecular features, and 266-dimensional imaging features are each mapped to 128-dimensional latent features, i.e. The result obtained after mapping to the latent space , and These are respectively referred to as target structural features, target molecular features, and target imaging features.

[0142] Modality type encoding is a vector used to add learnable modality identifiers to features of different modalities. Since the target features of the three modalities are mapped to a latent space of the same dimension, the feature dimensions are the same but their semantic origins are different. If the model does not distinguish between them, it will be difficult to effectively identify the modality affiliation of each feature in the subsequent fusion process. By setting a trainable modality type encoding for each modality... , and Adding this to the corresponding target features allows the model to perceive the modality source corresponding to each feature row during subsequent self-attention interactions.

[0143] The initial multimodal feature matrix is ​​formed by stacking three target features with modality type encoding. In one specific implementation, the initial multimodal feature matrix can be represented as:

[0144] Among them, matrix The first row corresponds to structural target features encoded in the genetic modality, the second row corresponds to functional molecular target features encoded in the molecular modality, and the third row corresponds to imaging target features encoded in the imaging modality. This initial multimodal feature matrix provides a structured input format for subsequent multimodal self-attention interactions.

[0145] Cross-modal feature fusion refers to the process of enabling information interaction and complementary enhancement among modal features based on an initial multimodal feature matrix using an attention mechanism. During cross-modal feature fusion, features of any effective modality can update themselves based on feature information from other effective modalities, thereby obtaining an enhanced representation that surpasses the information of a single modality. Cross-modal feature fusion can be achieved through a multimodal self-attention layer. This layer calculates attention weights between features of different modalities, allowing features of any modality to aggregate information from other modalities, achieving information flow and complementarity between modalities. After cross-modal feature fusion, an enhanced feature matrix is ​​obtained.

[0146] The enhanced feature matrix refers to the feature matrix obtained after cross-modal information interaction from the initial multimodal feature matrix, containing enhanced representations of each modality. The enhanced feature matrix includes enhanced structural feature vectors. Enhance molecular eigenvectors and enhanced image feature vectors In one specific implementation, the enhanced feature matrix can be represented as: .

[0147] Compared to the initial multimodal feature matrix, each row of the enhanced feature matrix no longer relies solely on the original information of the corresponding modality itself, but rather incorporates complementary information from other effective modalities. For example, enhanced structural feature vectors... It not only contains chromosomal structural information reflected by cfDNA, but may also incorporate discriminative information related to chromosomal abnormalities in the functional state of cfRNA molecules and developmental dynamics features observed in imaging.

[0148] Generating a fused feature vector based on enhanced structural feature vectors, enhanced molecular feature vectors, and enhanced imaging feature vectors can include weighted aggregation and / or concatenation of the three enhanced feature vectors. Weighted aggregation adaptively assigns contribution weights to the three enhanced feature vectors based on the importance of each modality in the current embryo prediction task, and obtains a comprehensive weighted aggregated feature through weighted summation. The concatenation operation combines the three enhanced feature vectors with the weighted aggregated feature to form a fused feature vector containing independent enhancement information from each modality and aggregated comprehensive information. The final generated fused feature vector integrates complementary discriminative information from the three modalities and serves as input to subsequent classification and prediction models.

[0149] Through the above embodiments, cross-modal feature fusion not only achieves the unification of different modal features in terms of dimensions and the alignment of semantics, but also realizes information interaction and enhancement between modalities through the self-attention mechanism, and aggregates the enhanced features based on adaptive weights, effectively overcoming the problems of insufficient robustness and limited prediction accuracy of single information sources, and providing a more information-rich and discriminative fusion feature foundation for euploid prediction.

[0150] In some embodiments, step S133, based on the initial multimodal feature matrix, performs cross-modal feature fusion to obtain an enhanced feature matrix, which may include the following steps.

[0151] Step S1331: Generate an effectiveness matrix based on the quality index of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector; wherein the effectiveness matrix is ​​a binarized diagonal matrix.

[0152] Step S1332: Generate attention masks for each modality based on the validity matrix.

[0153] Step S1333: Generate an initial attention feature matrix based on the attention mask and the initial multimodal feature matrix.

[0154] Step S1334: Multiply the effectiveness matrix with the initial attention feature matrix to obtain the target attention feature matrix.

[0155] Step S1335: Using the effectiveness matrix, modify the residual connection and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as the output result of the feedforward neural network, to obtain the enhanced feature matrix.

[0156] In this embodiment, by introducing a validity matrix and attention mask mechanism based on data quality during the cross-modal feature fusion process, the interference of the invalid modality on the feature interaction of the valid modality can be effectively shielded when the data of any modality is missing or the data quality does not meet the preset requirements. This ensures that the cross-modal fusion process only exchanges information between the actually valid modalities, thereby improving the accuracy and robustness of feature fusion.

[0157] Quality metrics are a set of parameters used to evaluate the quality and reliability of input data for each modality. For genetic modalities corresponding to structural genetic feature vectors, quality metrics may include at least one of the following: effective alignment rate, proportion of repetitive sequences, genome coverage, coverage uniformity, average fragment length, and amplification quality score. For molecular modalities corresponding to functional molecular feature vectors, quality metrics may include at least one of the following: number of effective sequences, reference transcriptome alignment rate, target gene detection rate, library complexity, expression coverage, and background noise level. For imaging modalities corresponding to imaging representation vectors, quality metrics may include at least one of the following: proportion of missing frames, image sharpness, exposure stability, integrity of embryonic region of interest, and timestamp integrity. When data for the corresponding modality exists, and all preset necessary quality metrics meet their respective thresholds, or when the comprehensive quality score calculated from multiple quality metrics reaches the preset quality threshold, the modality is determined to be a valid modality. When data for the corresponding modality is missing, the necessary quality metrics do not meet the requirements, or the comprehensive quality score is lower than the preset quality threshold, the modality is determined to be an invalid modality.

[0158] The validity matrix is ​​a binary diagonal matrix used to indicate the validity of each modality in the cross-modal feature fusion process. The validity indicators for genetic modality, molecular modality, and image modality are defined as follows: and The modality validity vector is defined as follows: its value is 1 when the corresponding modality is valid; its value is 0 when the corresponding modality is missing or invalid. And construct a diagonal validity matrix based on the modality validity vector. Validity matrix The elements on the diagonal of the matrix represent the validity indicators of the corresponding modalities, while the off-diagonal elements are all 0. The validity matrix is ​​used to set the feature rows corresponding to invalid modalities to zero during subsequent feature interactions, thereby preventing invalid modalities from interfering with the information of valid modalities.

[0159] When genetic, molecular, and imaging modalities are all missing or do not meet the preset data quality requirements, i.e. If the condition is not met, the modality weight calculation, feature fusion, and subsequent classification prediction will be stopped, and a message indicating insufficient input data or invalid modalities will be output.

[0160] An attention mask, generated based on the validity matrix, is used to mask invalid modalities during multimodal self-attention computation. The purpose of the attention mask is to significantly attenuate the attention scores corresponding to invalid modalities before Softmax computation, causing their attention weights to approach 0 after Softmax normalization. This prevents valid modalities from incorrectly aggregating information from invalid modalities when generating their own augmented features. In one specific implementation, an attention mask is defined. Its value selection rule is: when the first... One modality is valid, that is hour, That is, no attenuation is performed; when the first One modality is missing or invalid, i.e. hour, ,in The preset large positive number causes the corresponding attention score to be subtracted by a large value, resulting in the weights approaching 0 after the Softmax operation. Here, based on... Setting a mask is because of the... The column corresponds to the first key and value. There are several modalities. When a modality is missing, the corresponding attention score is reduced by a larger value so that its weight after Softmax normalization is close to 0.

[0161] The initial attention feature matrix is ​​the attention feature matrix obtained after multimodal self-attention computation from the initial multimodal feature matrix, but before output row masking. In one specific implementation, the initial multimodal feature matrix is ​​first used as the basis for... Calculate the query matrix Key matrix Sum matrix : , , ,in, , which is a trainable mapping weight matrix. Then, the initial attention feature matrix is ​​calculated using the attention mask: ,in, This represents the initial attention feature matrix before the missing modality output rows are masked. This represents the dimension of the key vector. This represents an attention mask. By using an attention mask... Adding the attention score to the scaled attention score matrix significantly reduces the attention score when invalid modes are used as keys and values, ensuring that valid modes are not affected by invalid modes during feature aggregation. In one specific implementation, and All 128-dimensional modal features are mapped to 128-dimensional features, making K and Both are 3×128 matrices, thus enabling the... and Add the residuals together.

[0162] The target attention feature matrix is ​​obtained by multiplying the validity matrix by the initial attention feature matrix and then forcibly setting the output rows corresponding to invalid modes to zero. In one specific implementation, the target attention feature matrix can be expressed as: ,in, For the validity matrix, The initial attention feature matrix, This is the target attention feature matrix after being masked by the validity matrix. When a certain modality is invalid, its value in the validity matrix is... The corresponding diagonal element is 0. After multiplication, this modality appears in the target attention feature matrix. The corresponding entire row of features is set to a zero vector. In this way, not only are invalid modalities, as information sources, blocked during the attention aggregation process, but their corresponding output positions are also forcibly cleared to zero, ensuring that the features of invalid modalities do not flow to subsequent network layers in any form.

[0163] Residual connections are used to add the initial multimodal feature matrix to the target attention feature matrix to preserve the original modal information and prevent gradient vanishing in deep networks. However, if masking is not performed in the residual connections, the all-zero attention outputs corresponding to invalid modalities will be added to the placeholder features in the initial features, and after layer normalization, non-zero features will be generated again, thus destroying the masking effect. Therefore, the residual connection and layer normalization results are modified using an effectiveness matrix. In one specific implementation, the modified residual connection and layer normalization results are as follows: Where LN represents the layer normalization operation, This is the validity matrix. The layer normalization result is compared with the validity matrix. After multiplication, the output row corresponding to the invalid mode is set to zero again.

[0164] The feedforward neural network is used to perform further row-by-row nonlinear transformations on the output of the self-attention layer to enhance the representational power of the features. Similarly, to prevent invalid modes from producing non-zero outputs after processing by the feedforward neural network, a validity matrix is ​​used to modify the output of the feedforward neural network. In one specific implementation, the final enhanced feature matrix can be expressed as: Where FFN represents a feedforward neural network. Enhanced feature matrix. The three rows in the image represent the enhanced structural feature vectors after being augmented with other effective modal information. Enhance molecular eigenvectors and enhanced image feature vectors .

[0165] Through the above embodiments, the effectiveness matrix and attention mask mechanism are integrated throughout the entire process of attention calculation, residual connection, layer normalization, and feedforward neural network in cross-modal feature fusion. This ensures that invalid modalities remain as zero features throughout the entire enhancement process. They neither interfere with the feature updates of effective modalities as information sources nor generate false feature outputs themselves. This effectively guarantees the accuracy and robustness of cross-modal feature fusion in complex clinical scenarios where the quality of multimodal data varies or some modalities are missing.

[0166] In some embodiments, in step S1335, invalid data in the residual connections and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as the output results of the feedforward neural network, are masked using the validity matrix to obtain an enhanced feature matrix, which may include the following steps.

[0167] Step S13351: Multiply the effectiveness matrix by the residual connection and layer normalization result of the initial attention feature matrix and the target attention feature matrix to obtain the initial enhancement feature matrix.

[0168] Step S13352: Multiply the validity matrix by the result of passing the initial enhanced feature matrix through the feedforward neural network to obtain the enhanced feature matrix.

[0169] In this embodiment, by forcibly setting the feature rows corresponding to invalid modes to zero twice, after residual connection and layer normalization, and after the output of the feedforward neural network, the shielding effect of invalid modes can be maintained continuously at multiple key nodes of cross-modal feature fusion, preventing the feature information of invalid modes from being accidentally activated or propagated during deep network computation.

[0170] The initial enhanced feature matrix is ​​obtained by performing residual concatenation and layer normalization on the initial multimodal feature matrix and the target attention feature matrix, and then masking the invalid modality output rows through the effectiveness matrix. In one specific implementation, the initial enhanced feature matrix... It can be obtained through the following formula: ,in, This represents the initial multimodal feature matrix. Let LN denote the target attention feature matrix, and LN denote the layer normalization operation. This represents the validity matrix.

[0171] The execution logic of this operation lies in: in multimodal self-attention computation, the target attention feature matrix... The row corresponding to the invalid mode in the validity matrix has been... Set it to the zero vector. However, when performing residual join operations... At that time, the row corresponding to the invalid mode is in the initial multimodal feature matrix. The values ​​in the matrix may not be all zeros, but rather exist as placeholder features or their modality type encoding. If layer normalization is directly applied to the residual connection results, the normalization operation will rescale and translate each modality row along the feature dimension. Since the rows corresponding to invalid modalities are in the target attention feature matrix... The value is zero, but... There are still placeholder values ​​in the data, so the row after the residual join is not all zeros. Layer normalization will renormalize this non-all-zero row into a non-zero vector with certain distribution characteristics, so that the feature row of this invalid mode will regain a non-zero numerical representation after normalization, thereby destroying the effect of the aforementioned masking operation. This is achieved by multiplying by the validity matrix again after layer normalization. This can force the row corresponding to the invalid mode to be reset to the zero vector, ensuring that the masking effect is maintained.

[0172] A feedforward neural network is a network structure used to perform row-by-row nonlinear transformations on the output of a self-attention layer. It typically consists of two fully connected layers and a nonlinear activation function. The feedforward neural network uses an initial augmentation feature matrix... Each row in the matrix is ​​used as an independent input, and the same nonlinear transformation is performed on each row to enhance the representational power of the features. However, the feedforward neural network does not change the shape of the input matrix when performing the row-by-row transformation; that is, if the input is a 3×128 matrix, the output is also a 3×128 matrix. This is because the initial enhanced feature matrix... The row corresponding to the invalid mode is a zero vector. After passing through the weight matrix multiplication and bias addition of the fully connected layer in the feedforward neural network, the zero row will become the value of the bias vector, thus becoming a non-zero vector, which will also destroy the masking effect.

[0173] The enhanced feature matrix is ​​the final feature matrix obtained by passing the initial enhanced feature matrix through a feedforward neural network, and then masking the invalid mode output rows with an effectiveness matrix. In one specific implementation, the enhanced feature matrix... It can be obtained through the following formula: ,in, denoted as the initial enhanced feature matrix, FFN represents the feedforward neural network, LN represents the layer normalization operation, and D represents the effectiveness matrix.

[0174] The execution logic of this operation lies in: initializing the enhanced feature matrix. Although the row corresponding to the invalid mode is a zero vector, after the linear transformation and bias addition of the feedforward neural network FFN, the output value of that row will no longer be a zero vector. If this non-zero output is directly compared with... Performing a residual join will cause the rows corresponding to invalid modes in the residual join result to reappear with non-zero features. Subsequent layer normalization will further normalize these non-zero feature rows to have a specific numerical distribution. Therefore, it is necessary to multiply by the validity matrix again after layer normalization. This forces the rows corresponding to invalid modes to be set to zero vectors. Thus, the feature matrix is ​​enhanced. In this model, the feature rows corresponding to invalid modes remain zero vectors after all computation steps, while the feature rows corresponding to valid modes contain enhanced feature information after multimodal attention interaction, residual connection, layer normalization, and nonlinear transformation of the feedforward neural network.

[0175] Through the above embodiments, the validity matrix At three key nodes in cross-modal feature fusion—attention output masking, residual connection post-masking, and feedforward neural network output post-masking—the feature rows corresponding to invalid modalities are forcibly set to zero. This multi-node masking mechanism, which runs throughout the entire cross-modal feature fusion process, ensures that invalid modalities do not interfere with the feature interactions of valid modalities during the feature enhancement process, nor are they accidentally activated in deep computation to produce spurious feature representations, thus effectively guaranteeing the purity and robustness of cross-modal feature fusion.

[0176] In some embodiments, in step S134, generating a fused feature vector based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector may include the following steps.

[0177] Step S1341: Multiply the enhanced feature matrix with the preset modality mapping weight vector to obtain the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the image modality contribution weight corresponding to the enhanced imaging feature vector.

[0178] Step S1342: According to the genetic modality contribution weight, the molecular modality contribution weight, and the imaging modality contribution weight, the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector are weighted and aggregated to obtain a weighted aggregated feature vector.

[0179] Step S1343: Concatenate the enhanced structural feature vector, the enhanced molecular feature vector, the enhanced imaging feature vector, and the weighted aggregated feature vector to obtain a fused feature vector.

[0180] In this embodiment, by adaptively learning the contribution weights of each modality in the current embryo prediction task from the enhanced feature matrix, and weighting and aggregating the three enhanced features based on these weights, and then concatenating each enhanced feature with the weighted aggregated features, the fused feature vector can simultaneously contain the independent enhanced discriminative information of each modality and the comprehensive discriminative information of the adaptive aggregation between modalities, thereby providing the classification prediction model with input features of richer information hierarchy.

[0181] The preset modality mapping weight vector is a trainable parameter vector used to map each modality feature in the augmented feature matrix to its corresponding modality importance score. In one specific implementation, the augmented feature matrix... It can be represented as: .in, , and These represent the enhanced structural feature vector, enhanced molecular feature vector, and enhanced imaging feature vector, respectively, after multimodal information interaction enhancement. All three belong to... By enhancing the feature matrix Mapping weight vector to preset mode Multiply and add the corresponding bias term. The initial importance scores for the three modalities can be calculated: .in, , , and These represent the initial importance scores for the genetic modality, molecular modality, and imaging modality, respectively.

[0182] The contribution weights of the genetic modality, molecular modality, and imaging modality are weight coefficients obtained after normalization based on the initial importance scores mentioned above. These weights characterize the relative importance of each modality in the current embryo prediction task. In one specific implementation, Softmax normalization of the initial importance scores yields the contribution weights of the three modalities: ,in, , and These represent the contribution weights of the genetic modality, molecular modality, and imaging modality, respectively. Each contribution weight is non-negative, and the sum of the three is 1. , When an enhanced feature of a certain modality obtains a high initial importance score after modality weight mapping, its corresponding contribution weight increases accordingly, and vice versa. Thus, the model can adaptively adjust the participation of each modality in the fusion process based on the actual quality and discriminative information of the data from different embryos.

[0183] The weighted aggregated feature vector is a comprehensive feature vector obtained by weighting and summing the three enhanced feature vectors according to the contribution weights of each modality. In one specific implementation, the weighted aggregated feature vector... It can be calculated using the following formula: ,in, The weighted aggregated feature vector comprehensively represents the overall contribution of the three effective modalities in the current embryo prediction task, and is an adaptive fusion representation of the enhancement features of each modality. Compared with single-modality features, the weighted aggregated feature vector can automatically reduce the contribution of a certain modality when the data quality is poor through adaptive adjustment of the fusion weights, and rely more on the information of other higher-quality modalities, thereby improving the overall robustness of the fused features.

[0184] The fused feature vector is a feature vector obtained by concatenating three enhanced feature vectors with a weighted aggregated feature vector. In one specific implementation, the fused feature vector... It can be represented as: ,in, , and Each is a 128-dimensional enhanced feature vector. The 128-dimensional weighted aggregated feature vectors are concatenated to obtain a 512-dimensional fused feature vector. .

[0185] The above embodiments, by jointly concatenating enhanced features and weighted aggregated features to generate a fused feature vector, not only fully preserve the independent feature representations of each modality after cross-modal interactive enhancement, enabling the classification prediction model to continue utilizing the specific discriminative information of each modality, but also introduce a comprehensive feature representation obtained through adaptive weight aggregation, allowing the classification prediction model to simultaneously utilize the global discriminative information after complementary fusion between modalities. This multi-level fused feature construction method can provide subsequent classification prediction models with inputs of richer information granularity, thereby helping to improve the accuracy and robustness of euploidy prediction.

[0186] In some embodiments, step S1343, multiplying the enhanced feature matrix with a preset modality mapping weight vector to obtain the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the image modality contribution weight corresponding to the enhanced imaging feature vector, may include the following steps.

[0187] Step S13431: Generate a modality mapping vector based on trainable modality weight mapping parameters; wherein the modality mapping vector includes initial weights for genetic modalities, initial weights for molecular modalities, and initial weights for image modalities.

[0188] Step S13432: Generate a binary mask value vector based on the quality index of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector; wherein the mask value vector is used to characterize the effectiveness of each modality.

[0189] Step S13433: Normalize the result of adding the modality mapping vector and the mask value vector to obtain the genetic modality contribution weight, molecular modality contribution weight and image modality contribution weight; wherein the sum of the genetic modality contribution weight, the molecular modality contribution weight and the image modality contribution weight is 1.

[0190] In this embodiment, by introducing a mask value vector based on data quality indicators during the modality weight calculation process, and adding the mask value vector to the modality mapping vector learned from the enhanced feature matrix and then normalizing it, the weights of invalid modalities can be synchronously reduced to zero during the adaptive learning of the contribution weights of each modality. This ensures that the final modality contribution weights are only allocated among the actually valid modalities, thereby avoiding interference from invalid modalities in the calculation of weighted aggregate features.

[0191] Trainable modality weight mapping parameters are learnable parameters used to map the feature representations of each modality in the enhanced feature matrix to initial importance scores. In one specific implementation, the trainable modality weight mapping parameters include a modality weight mapping vector. and the corresponding bias term Modal weight mapping vector With enhanced feature matrix Multiplication allows the 128-dimensional enhancement features of each modality to be projected into a scalar score value.

[0192] The modality mapping vector is a vector obtained by linearly mapping the enhanced feature matrix based on trainable modality weight mapping parameters. It characterizes the initial importance of each modality in the current embryo prediction task. In one specific implementation, the modality mapping vector... It can be calculated using the following formula: .in, , Represents the initial weights of the genetic modality. Indicates the initial weights of the molecular modes. This represents the initial weights of the image modalities. The values ​​of each element in the modality mapping vector reflect the importance of that modality to the current prediction task, as automatically learned by the model based on the enhancement features of each modality. When the enhancement features of a certain modality contain more information related to euploidy discrimination, the initial weight value corresponding to that modality is relatively high.

[0193] The binarized mask value vector is a vector generated based on the quality indicators of each modality's data, used to mark and mask invalid modalities during weight calculation. In one specific implementation, the mask value vector... It can be represented as: ,in, , and These represent the mask values ​​for genetic, molecular, and imaging modalities, respectively. The value of each element in the mask value vector is directly related to the validity identifier of the corresponding modality. When the... Each modality is valid, i.e., its validity indicator. When, the corresponding mask value That is, no decay occurs; when the first One modality is missing or the data quality does not meet the preset requirements, i.e. When, the corresponding mask value ,in This is a pre-set, relatively large positive number. By setting the mask value of the invalid mode to a relatively large negative number, the weight obtained by that mode in the subsequent Softmax normalization process can be made close to 0.

[0194] The mask value vector is used to characterize the effectiveness of each mode. (This is related to the effectiveness matrix.) Unlike the row masking function in multimodal self-attention interaction, the mask value vector is specifically used in the calculation of modality contribution weights. By significantly attenuating the initial score of invalid modalities before Softmax normalization, the contribution weights of invalid modalities approach 0 after normalization.

[0195] Normalization is the process of converting the result of adding the modality mapping vector and the mask value vector into modality contribution weights that satisfy the constraints of non-negativity and a sum of 1. In one specific implementation, normalization is achieved using the Softmax function: Modality mapping vector With mask value vector After addition, the initial score corresponding to the effective mode remains unchanged (because its mask value is 0), while the initial score corresponding to the invalid mode is subtracted by a large positive number. This becomes a very large negative value. After Softmax exponentiation and normalization, the contribution weights corresponding to the invalid modes approach 0, while the contribution weights of the valid modes are renormalized within the valid mode range, making the sum of the contribution weights of the genetic mode, molecular mode, and image mode equal to 1.

[0196] For example, when all three modes are valid, the mask value vector The final contribution weights are entirely determined by the modality mapping vector. Determined after Softmax normalization. When image modalities are missing or data quality does not meet preset requirements, the mask value vector... ,through After calculation, the image modal contribution weights Approaching 0, the contribution weight of genetic modality Molecular mode contribution weights Renormalize within the effective modal range to satisfy... .

[0197] In the above embodiments, the calculation process of modality contribution weights integrates two aspects of information: first, the discriminative importance of each modality is automatically learned based on the enhanced feature matrix; and second, the effectiveness status of each modality is determined based on data quality indicators. These two aspects are organically combined through a computational chain of "initial score mapping - mask value addition - Softmax normalization," enabling the final modality contribution weights to adaptively reflect the information contribution degree of each modality in the current prediction task. Furthermore, when some modalities are missing or invalid, their weights are automatically reset to zero, and the weight ratio of effective modalities is redistributed. This ensures that the calculation of the weighted aggregated feature vector relies only on actual, effective, and discriminative modal features, improving the robustness and reliability of feature fusion.

[0198] In some embodiments, step S140, inputting the fused feature vector into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested, may include the following steps.

[0199] Step S141: Perform feature compression on the fused feature vector to obtain a compressed fused feature vector.

[0200] Step S142: Input the compressed fusion feature vector into a preset classification model to obtain the euploidy prediction probability of the embryo to be tested.

[0201] Step S143: If the euploid prediction probability is greater than or equal to a preset threshold, the prediction result of the embryo to be tested is determined to be euploid.

[0202] In this embodiment, by performing feature compression before inputting the fused feature vector into the preset classification model, the dimensionality of the fused feature vector can be reduced, redundant information between features can be removed, and the most discriminative feature representation can be extracted. Furthermore, by performing binarization decision on the predicted probability output by the model through a preset threshold, the continuous probability output can be transformed into a clear euploid or aneuploid classification result, thereby directly providing an operable basis for clinical embryo selection.

[0203] Feature compression is a process used to map high-dimensional fused feature vectors into lower-dimensional, compact feature representations. Fused feature vectors are constructed by concatenating enhanced structural feature vectors, enhanced molecular feature vectors, enhanced imaging feature vectors, and weighted aggregated feature vectors, resulting in high dimensionality.

[0204] In one specific implementation, feature vectors are fused. The feature vector is 512-dimensional. Directly inputting such a high-dimensional feature vector into a pre-defined classification model may lead to training difficulties or overfitting due to feature redundancy. By compressing the fused feature vector, the feature dimensionality can be reduced while retaining the main discriminative information, thus improving the model's generalization ability. In one specific implementation, feature compression can be achieved through a fused feature compression layer: ,in, Represents the compression mapping matrix, This indicates the corresponding bias term. Let LN denote a nonlinear activation function, and LN denote a layer normalization operation. Through the above feature compression, a compressed fused feature vector can be obtained. In one specific implementation, the fused feature vector is 512-dimensional. Compressed into a 256-dimensional fused feature vector .

[0205] Compressed fusion feature vectors are compact feature representations obtained by compressing fusion feature vectors. Compressed fusion feature vectors simultaneously contain coverage characterization information related to chromosome copy number abnormalities, embryonic molecular functional state information, spatial morphological information, and temporal developmental dynamics information, and are a synthesis and condensation of multiple modality discrimination information.

[0206] The pre-defined classification model is a classifier used to output the urlocity prediction probability based on the compressed fused feature vector. In one specific implementation, the pre-defined classification model may include multiple fully connected layers, non-linear activation layers, and random deactivation layers. Specifically, the first fully connected layer maps the 256-dimensional compressed fused feature vector to 128-dimensional intermediate features: The second fully connected layer further maps the 128-dimensional intermediate features into 64-dimensional discriminative features: The classification output layer calculates two classification scores for neuploid and aneuploid based on the 64-dimensional discriminant features: Dropout, representing a random deactivation layer, is used to randomly discard some neurons during training to prevent overfitting. The classification output layer outputs two real scores, corresponding to the euploid and aneuploid categories, respectively.

[0207] The euploidy prediction probability is a probability value used to characterize whether the embryo being tested is euploid, obtained by transforming the classification score output by the classification output layer through a normalized exponential function. In one specific implementation, the euploidy prediction probability... Aneuploid prediction probability It can be calculated using the Softmax function: , ,in, Represented by natural constant It is an exponential function with base 0. After Softmax normalization, the predicted probabilities of euploidy and aneuploidy are both values ​​between 0 and 1, and satisfy the following conditions: The closer the euploidy prediction probability is to 1, the higher the confidence level of the model in determining that the embryo being tested is euploid.

[0208] The preset threshold is a decision boundary used to convert the euploidy prediction probability into a binary classification result. When the euploidy prediction probability is greater than or equal to the preset threshold, the predicted result of the embryo to be tested is determined to be euploid; when the euploidy prediction probability is less than the preset threshold, the predicted result of the embryo to be tested is determined to be aneuploid. In one specific embodiment, the preset threshold... It can be set to 0.5, that is, when When, it is determined to be euploid; when When this occurs, it is determined to be aneuploid. In another specific implementation, the preset threshold can also be determined based on the receiver operating characteristic curve of the validation dataset to achieve the optimal balance between sensitivity and specificity required clinically.

[0209] During the model training phase, the pre-defined classification model uses PGT-A results obtained from trophoblastic ectoderm biopsies as supervisory labels for euploidy classification. Only embryo samples explicitly identified as euploid or aneuploid by PGT-A results are used for binary classification model training; chimeric embryo samples are not included in the calculation of the primary classification loss. The model training uses the classification cross-entropy loss function: .in, Indicates category The corresponding real tags, This represents the predicted probability of the class output by the model. For cases where the number of euploid and aneuploid samples is imbalanced, class weights can be added to the loss function above to balance the contributions of different classes of samples to model training.

[0210] Furthermore, in sharing discriminative features Based on this, an implantation outcome auxiliary prediction branch and a clinical pregnancy outcome auxiliary prediction branch are set up, outputting the implantation probability and clinical pregnancy probability respectively (the corresponding auxiliary task loss is calculated only for samples with corresponding clinical outcome labels; for embryos that were not transferred or lack clinical outcome records, the auxiliary task loss for that sample is not calculated). The model is trained using the following joint loss function:

[0211] in, Indicates euploid classification loss. This indicates the predicted implantation outcome loss. This indicates a loss of prediction for pregnancy outcomes. and This indicates the weight of the corresponding auxiliary task. and These represent the validity indicators of the implantation outcome label and the clinical pregnancy outcome label, respectively, and must meet the following requirements: When a sample has a corresponding clinical outcome label, the corresponding label validity flag is set to 1; when a sample is not transplanted, lacks a corresponding clinical outcome record, or the corresponding label is unavailable, the corresponding label validity flag is set to 0, so that the sample does not generate loss for the corresponding auxiliary task. The clinical outcome auxiliary prediction branch is used for joint optimization during the model training phase, and in the actual euploidy prediction process, it can output only the euploidy classification result.

[0212] In the actual prediction stage after the model training is completed, the preset classification model only uses the compressed fusion feature vector obtained by non-invasively acquiring culture medium nucleic acid data and time series image data after processing by the aforementioned steps as input. It can output the euploidy prediction probability and the corresponding binary classification result without performing trophectoderm biopsy on the embryo to be tested.

[0213] In the above embodiments, feature compression can effectively reduce feature dimensionality and computational complexity while preserving multimodal fusion discrimination information. The preset classification model can automatically learn the classification boundary between euploids and aneuploids from the compressed fusion feature vectors. The preset threshold mechanism provides clear and operable judgment criteria for clinical decision-making, thereby realizing fully automated prediction from multimodal raw data to clear euploid classification results.

[0214] In some embodiments, the number of embryos to be tested is multiple, and after step S143, the embryo euploidy prediction method based on multimodal fusion may further include the following steps.

[0215] Step S151: Sort the euploidy prediction probabilities of multiple embryos in descending order to generate embryo prediction sequences.

[0216] Step S152: Output the embryo prediction sequence and the embryo prediction result corresponding to each prediction probability in the embryo prediction sequence.

[0217] In this embodiment, when there are multiple candidate embryos for the same patient, the euploidy prediction probability of each embryo is uniformly sorted to generate an embryo prediction sequence that intuitively reflects the order of the likelihood of euploidy. This can provide clinicians with a quantitative basis for making optimal choices in embryo transfer decisions, thereby assisting doctors in prioritizing the selection of the embryo with the highest likelihood of euploidy for transfer and improving the success rate of a single transfer.

[0218] Multiple embryos refer to multiple in vitro cultured embryos obtained from the same patient in the same in vitro fertilization (IVF) or intracytoplasmic sperm injection (ICSI) treatment cycle. In clinical assisted reproductive practice, a single treatment cycle typically yields multiple embryos. In traditional embryo selection procedures, embryologists primarily rely on morphological scoring to grade candidate embryos and determine their transfer order. However, because the correlation between morphological parameters and chromosomal neuploidy is limited, relying solely on morphological scoring for embryo selection may result in the priority transfer of morphologically sound but chromosomally aneuploid embryos, thereby reducing the success rate of implantation.

[0219] The euploidy prediction probability is a probability value output by the classification prediction model for each embryo to be tested, used to characterize that embryo as euploid. Since the euploidy prediction probability is a comprehensive judgment result obtained by multimodal fusion and classification prediction based on cfDNA structural genetic information, cfRNA functional molecular information, and developmental dynamics information from imaging, it can reflect the euploidy status of the embryo more comprehensively and accurately than single-modal detection indicators or morphological scores.

[0220] The embryo prediction sequence is an optimized sequence generated by sorting the euploidy prediction probabilities of multiple embryos to be tested in descending order. Specifically, the embryos to be tested are arranged in descending order of their euploidy prediction probabilities, with the embryo with the highest prediction probability at the beginning of the sequence, the second highest at the end, and so on. Each position in the embryo prediction sequence contains the euploidy prediction probability value of that embryo and the corresponding euploidy or aneuploidy classification result. For example, for three candidate embryos from the same patient, the euploidy prediction probabilities output by the classification prediction model are 0.86, 0.63, and 0.21, respectively. The embryo prediction sequence will then place the embryo with a euploidy prediction probability of 0.86 at the beginning, the embryo with a probability of 0.63 at the end, and the embryo with a probability of 0.21 at the end. Furthermore, if the preset classification threshold is set to 0.5, the embryos with prediction probabilities of 0.86 and 0.63 will be classified as euploid, and the embryo with a prediction probability of 0.21 will be classified as aneuploid.

[0221] Outputting embryo prediction sequences and corresponding prediction results means presenting the sorted, optimized sequences and their predicted probabilities and classification results to clinicians in a readable format. The output information may also include the genetic modality contribution weight, molecular modality contribution weight, and imaging modality contribution weight for each embryo, as well as the data quality or modality integrity indicator for each modality. This supplementary information helps clinicians understand the reliability of each prediction result and the quality of the data for each modality, providing a more comprehensive reference for clinical decision-making. For example, if an embryo has a high euploidy prediction probability, but its corresponding modality integrity indicator shows that a certain modality of data is missing or of poor quality, the doctor can combine other clinical information to comprehensively judge the reliability of the prediction result.

[0222] The above embodiments can transform the euploidy prediction results of multiple candidate embryos by the classification prediction model into an intuitive preferred sequence, providing an objective and quantitative ranking basis for clinical embryo transfer decisions without the need for invasive biopsies of the embryos.

[0223] Please see Figure 3 This embodiment provides an embryo euploidy prediction device based on multimodal fusion, which may include: The data acquisition module 210 is used to acquire cfDNA sequencing data, cfRNA expression data and time-series image data of the embryos to be tested.

[0224] The feature extraction module 220 is used to extract the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, respectively.

[0225] The feature fusion module 230 is used to perform cross-modal feature fusion on the structural genetic feature vector, the functional molecular feature vector and the imaging representation vector to obtain a fused feature vector.

[0226] The embryo prediction module 240 is used to input the fused feature vector into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

[0227] The specific functions and effects of the embryo euploidy prediction device based on multimodal fusion can be explained by referring to other embodiments in this specification, and will not be repeated here. Each module in the embryo euploidy prediction device based on multimodal fusion can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0228] Please see Figure 4 One embodiment of the present invention can provide an electronic device, the electronic device comprising: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors, which, when executed by the one or more processors, enable the one or more processors to implement the embryo euploidy prediction method based on multimodal fusion as described in any of the above embodiments.

[0229] One embodiment of the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the embryo euploidy prediction method based on multimodal fusion as described in any of the above embodiments.

[0230] An embodiment of the present invention also provides a computer program product comprising a computer program, which, when executed by a computer, implements the embryo euploidy prediction method based on multimodal fusion as described in any of the above embodiments.

[0231] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments described herein, and are not intended to limit the scope of the invention.

[0232] It is understood that in the various embodiments described in this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments described in this specification.

[0233] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and the implementation methods in this specification are not limited in this respect.

[0234] Unless otherwise stated, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0235] It is understood that the processor in this invention can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method implementation can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this specification. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this specification can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0236] It is understood that the memory in this invention can be volatile memory or non-volatile memory, or may include both. Specifically, the non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0237] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and corresponding operation entry points are provided for the user to choose to authorize or refuse.

[0238] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0239] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0240] In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0241] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0242] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0243] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of this specification, in essence, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of software products. These computer software products are stored in a storage medium and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this specification. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0244] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting embryonic euploidy based on multimodal fusion, characterized in that, The method includes: Collect cfDNA sequencing data, cfRNA expression data, and time-series image data from the embryos to be tested; Based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test are extracted respectively. Cross-modal feature fusion is performed on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector; The fused feature vector is input into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

2. The method according to claim 1, characterized in that, Collect cfDNA sequencing data, cfRNA expression data, and time-series imaging data from the embryos to be tested, including: cfDNA and cfRNA nucleic acid samples were extracted from the culture medium samples of the embryos to be tested collected at preset sampling time points. The cfDNA nucleic acid sample was subjected to whole-genome amplification and sequencing to obtain cfDNA sequencing data; The cfRNA nucleic acid samples were reverse transcribed and amplified, and the transcriptome was sequenced to obtain cfRNA expression data; Using a time-difference imaging system, embryo images of the embryo under test are continuously acquired from a preset imaging start time point at preset time intervals until a preset prediction time point, to obtain time-series image data; wherein, each embryo image corresponds to an acquisition timestamp, and the preset prediction time point is equal to or later than the preset sampling time point.

3. The method according to claim 1, characterized in that, Based on the cfDNA sequencing data, the cfRNA expression data, and the time-series imaging data, respectively, structural genetic feature vectors, functional molecular feature vectors, and imaging representation vectors of the embryo under test are extracted, including: Based on the cfDNA sequencing data, a structural genetic feature vector is extracted from the embryo to be tested; wherein, the structural genetic feature vector is used to characterize the difference between euploidy and aneuploidy; Based on the cfRNA expression data, the functional molecular feature vector of the embryo under test is extracted; wherein, the functional molecular feature vector is used to characterize the functional state of the embryo under test. Based on the time-series image data, the imaging representation vector of the embryo under test is extracted; wherein, the imaging representation vector is used to characterize the dynamic developmental features and spatial morphological features of the blastocyst of the embryo under test.

4. The method according to claim 3, characterized in that, Based on the cfDNA sequencing data, the structural genetic feature vector of the embryo to be tested is extracted, including: The cfDNA sequencing data is cleaned to obtain the normalized sequencing coverage of each chromosome in the embryo to be tested; wherein, the data cleaning includes sequentially performing adapter removal, quality filtering, reference genome alignment, repetitive sequence removal, GC content correction and sequencing coverage normalization. The deviation between the normalized sequencing coverage of each chromosome and the reference coverage of the corresponding chromosome in the standard euploid control sample was calculated to obtain the coverage deviation index for each chromosome. The deviation between the normalized sequencing coverage of each chromosome and the mean normalized coverage of the corresponding chromosome in the normal euploid control sample set is calculated to obtain the coverage abnormality index of each chromosome; wherein, the normal euploid control sample set includes multiple normal euploid control samples. The structural genetic feature vector of the embryo under test is obtained by combining the coverage deviation index, the coverage abnormality index, and the quality index of the cfDNA sequencing data of each chromosome; wherein the quality index includes at least one of the following: effective alignment rate, repetitive sequence ratio, genome coverage, coverage uniformity, average fragment length, and amplification quality score.

5. The method according to claim 3, characterized in that, Based on the cfRNA expression data, the functional molecular feature vector of the embryo under test is extracted, including: The cfRNA expression data were sequentially subjected to low-quality sequence filtering, reference transcriptome alignment, transcript quantification, batch effect correction, and expression level standardization to obtain a gene expression dataset. Logarithmically transform the expression levels of target genes in the gene expression dataset to obtain the target expression levels of each target gene. The target expression levels of target genes belonging to the same biological pathway are weighted and summed according to preset weights to obtain the pathway activity index of the corresponding biological pathway. The gene expression levels of each target gene and the pathway activity indicators of each biological pathway are combined to obtain the functional molecular feature vector of the embryo under test.

6. The method according to claim 3, characterized in that, Based on the time-series image data, the imaging representation vector of the embryo to be tested is extracted, including: The region of interest of each embryo image in the time series image data is preprocessed to obtain a target image sequence; wherein, the target image sequence includes target images corresponding to each embryo image; the preprocessing includes at least one of the following: background removal, noise suppression, position alignment, size unification and pixel value normalization. Feature extraction is performed on each target image to obtain the spatial morphological features of each target image; The spatial morphological features of each target image and its corresponding temporal location code are fused to obtain a spatiotemporal representation vector. Based on the time-series image data, the developmental dynamics parameter vector of the embryo under test is extracted; The spatiotemporal representation vector is concatenated with the developmental dynamics parameter vector to obtain the imaging representation vector of the embryo under test.

7. The method according to any one of claims 1 to 6, characterized in that, Cross-modal feature fusion is performed on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector, including: The structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector are mapped to the latent space to obtain the target structural feature, the target molecular feature, and the target imaging feature; Structural feature encoding is added to the target structural features, molecular feature encoding is added to the target molecular features, and image feature encoding is added to the target imaging features, and an initial multimodal feature matrix is ​​constructed; Based on the initial multimodal feature matrix, cross-modal feature fusion is performed to obtain an enhanced feature matrix; wherein, the enhanced feature matrix includes enhanced structural feature vectors, enhanced molecular feature vectors, and enhanced imaging feature vectors; A fused feature vector is generated based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector.

8. The method according to claim 7, characterized in that, Based on the initial multimodal feature matrix, cross-modal feature fusion is performed to obtain an enhanced feature matrix, including: An effectiveness matrix is ​​generated based on the quality indices of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector; wherein the effectiveness matrix is ​​a binarized diagonal matrix. Based on the validity matrix, generate attention masks for each modality; Based on the attention mask and the initial multimodal feature matrix, an initial attention feature matrix is ​​generated; Multiply the effectiveness matrix by the initial attention feature matrix to obtain the target attention feature matrix; Using the effectiveness matrix, the residual connections and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as the output of the feedforward neural network, are modified to obtain the enhanced feature matrix.

9. The method according to claim 8, characterized in that, Using the validity matrix, invalid data in the residual connections and layer normalization results of the initial attention feature matrix and the target attention feature matrix, as well as in the output of the feedforward neural network, are masked to obtain an enhanced feature matrix, including: The initial enhancement feature matrix is ​​obtained by multiplying the effectiveness matrix by the residual concatenation and layer normalization results of the initial attention feature matrix and the target attention feature matrix; The enhanced feature matrix is ​​obtained by multiplying the validity matrix by the initial enhanced feature matrix and passing the result through the feedforward neural network.

10. The method according to claim 7, characterized in that, Based on the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector, a fused feature vector is generated, including: Multiplying the enhanced feature matrix with the preset modality mapping weight vector yields the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the imaging modality contribution weight corresponding to the enhanced imaging feature vector. According to the genetic modality contribution weight, the molecular modality contribution weight, and the imaging modality contribution weight, the enhanced structural feature vector, the enhanced molecular feature vector, and the enhanced imaging feature vector are weighted and aggregated to obtain a weighted aggregated feature vector; The enhanced structural feature vector, the enhanced molecular feature vector, the enhanced imaging feature vector, and the weighted aggregated feature vector are concatenated to obtain a fused feature vector.

11. The method according to claim 10, characterized in that, Multiplying the enhanced feature matrix by a preset modality mapping weight vector yields the genetic modality contribution weight corresponding to the enhanced structural feature vector, the molecular modality contribution weight corresponding to the enhanced molecular feature vector, and the imaging modality contribution weight corresponding to the enhanced imaging feature vector, including: Based on trainable modality weight mapping parameters, a modality mapping vector is generated; wherein, the modality mapping vector includes initial weights for genetic modalities, initial weights for molecular modalities, and initial weights for image modalities; Based on the quality indices of the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector, a binary mask value vector is generated; wherein, the mask value vector is used to characterize the effectiveness of each modality; The result of adding the modality mapping vector and the mask value vector is normalized to obtain the genetic modality contribution weight, the molecular modality contribution weight, and the image modality contribution weight; wherein the sum of the genetic modality contribution weight, the molecular modality contribution weight, and the image modality contribution weight is 1.

12. The method according to claim 1, characterized in that, The fused feature vector is input into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo under test, including: The fused feature vector is compressed to obtain a compressed fused feature vector; The compressed and fused feature vector is input into a preset classification model to obtain the euploidy prediction probability of the embryo under test; If the euploid prediction probability is greater than or equal to a preset threshold, the prediction result of the embryo to be tested is determined to be euploid.

13. The method according to claim 12, characterized in that, The number of embryos to be tested is multiple, and the method further includes: The euploidy prediction probabilities of multiple embryos to be tested are sorted in descending order to generate an embryo prediction sequence; Output the embryo prediction sequence, and the embryo prediction result corresponding to each prediction probability in the embryo prediction sequence.

14. A device for predicting embryonic euploidy based on multimodal fusion, characterized in that, The device includes: The data acquisition module is used to acquire cfDNA sequencing data, cfRNA expression data, and time-series image data of the embryos to be tested; The feature extraction module is used to extract the structural genetic feature vector, functional molecular feature vector, and imaging representation vector of the embryo under test based on the cfDNA sequencing data, the cfRNA expression data, and the time-series image data, respectively. The feature fusion module is used to perform cross-modal feature fusion on the structural genetic feature vector, the functional molecular feature vector, and the imaging representation vector to obtain a fused feature vector; The embryo prediction module is used to input the fused feature vector into a pre-trained classification prediction model to obtain the euploidy prediction result of the embryo to be tested.

15. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the embryo euploidy prediction method based on multimodal fusion as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the embryonic euploidy prediction method based on multimodal fusion as described in any one of claims 1 to 13.

Images