Assisted reproduction IVF outcome prediction model based on LSTM and LSTM-CNN fusion architecture and construction method

Abstract

The invention provides an assisted reproduction IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture and a construction method, and is applied to the field of data processing. The method comprises the following steps: acquiring assisted reproduction clinical layering core data (including IVF period original data and divided into conventional fertilization IVF and ICSI fertilization IVF according to a fertilization mode), key parameters (including sperm motility and the like in conventional IVF and additional sperm morphological images in ICSI) and target patient information; preprocessing the data, carrying out feature engineering, and generating multi-modal feature vectors adaptive to the two fertilization modes; then, an optimization prediction model is constructed, and an initial model and performance evaluation information are obtained; iteratively optimizing the model, introducing a new module to improve the multi-modal learning ability, and generating an SOTA level model through multi-index evaluation; and finally, processing target patient information by using the model, and outputting a personalized prediction reference scheme and clinical application value analysis information.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to an assisted reproductive IVF outcome prediction model and its construction method based on an LSTM and LSTM-CNN fusion architecture. Background Technology

[0002] Currently, the mainstream international ART outcome prediction tools (such as Vitrolife, ThermoFisher, and Cooper) cannot meet the clinical needs for assessing the overall treatment outcome.

[0003] Insufficient Model Performance During the Research Phase: Current IVF prediction models in the international research phase have the following shortcomings: Predictive dimensions are limited: focusing on a single influencing factor (such as embryonic developmental potential, ovarian reserve, and dosage of ovulation-inducing drugs) or using simple statistical analysis (such as regression analysis) without integrating multi-dimensional data such as physiological parameters of both partners, reproductive hormones, and sperm quality, makes it impossible to fully capture the complex influencing mechanisms of IVF outcomes.

[0004] Key parameter missing: "Sperm morphology" was not included in the predictive variables. However, abnormal sperm morphology is a direct cause of ICSI (intracytoplasmic sperm injection) fertilization failure and has a key impact on ICSI-IVF outcomes. The missing parameter significantly limited the model's predictive accuracy for ICSI outcomes.

[0005] Low accuracy and generalization: Due to insufficient data (mostly based on thousands of samples) and lack of customized models for different fertilization methods (conventional IVF and ICSI IVF), the prediction results fluctuate greatly, making it difficult to implement in clinical practice.

[0006] Furthermore, existing models often employ single-modal data (such as numerical clinical parameters only) without integrating image data (such as sperm morphology images); and the data preprocessing process is simple (such as the lack of standardized processing schemes for missing and outlier values), resulting in low data quality and affecting model predictive performance. Existing models mostly use traditional machine learning algorithms (such as random forests and support vector machines) or simple neural network architectures, which cannot effectively handle the temporal features (such as changes in reproductive hormones throughout the treatment cycle) and multimodal features (such as the correlation between numerical and image parameters) in IVF data; and they lack differentiated architectures for different fertilization methods, making it difficult to balance model versatility and accuracy.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore includes information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] According to one aspect of this application, a method for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture is provided, comprising: acquiring core data and key parameter information of assisted reproductive clinical stratification, and target patient information, wherein the core data of assisted reproductive clinical stratification includes raw IVF cycle data, divided into two datasets according to fertilization method: conventional fertilization IVF and ICSI fertilization IVF; the key parameter information includes conventional fertilization IVF parameters, which include the proportion of grade A sperm, the proportion of grade A+B sperm, the age / BMI of both partners, female reproductive hormones, number of antral follicles, endometrial thickness, cause of female infertility, and duration of infertility; and ICSI fertilization IVF parameters, which incorporate sperm morphology image data in addition to the conventional parameters; for assisted reproductive... Clinical stratification core data and key parameter information are preprocessed and feature-engineered to generate multimodal feature vectors adapted to two types of fertilization methods. Based on the multimodal feature vectors, in vitro fertilization prediction models for the two types of fertilization methods are constructed and optimized to generate initial outcome prediction models and performance evaluation reference information. The initial outcome prediction models are iteratively tested to increase model depth, introduce new modules, improve the joint learning capability of multimodal data, and repeatedly train and validate them. Evaluation is based on confusion matrix, PR curve, AUC value, and F1 value to generate a state-of-the-art (SOTA) IVF outcome prediction model. Based on the SOTA-level IVF outcome prediction model, the target patient information is processed to generate personalized prediction reference schemes for assisted reproductive IVF cycles and clinical application value analysis information.

[0009] Another aspect of this application discloses a device for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, comprising: an acquisition module for acquiring core data and key parameter information of assisted reproductive clinical stratification, and target patient information, wherein the core data of assisted reproductive clinical stratification includes raw IVF cycle data, divided into two datasets according to fertilization method: conventional IVF and ICSI IVF; key parameter information includes conventional IVF parameters, which include the proportion of grade A sperm, the proportion of grade A+B sperm, the age / BMI of both partners, female reproductive hormones, number of antral follicles, endometrial thickness, cause of female infertility, and duration of infertility; and ICSI IVF parameters, which incorporate sperm morphology image data in addition to the conventional parameters; and a processing module, using... The core data and key parameters of assisted reproductive clinical stratification are preprocessed and feature-engineered to generate multimodal feature vectors adapted to two fertilization methods. Based on the multimodal feature vectors, in vitro fertilization prediction models for the two fertilization methods are constructed and optimized to generate initial outcome prediction models and performance evaluation reference information. The initial outcome prediction models are iteratively processed to increase model depth, introduce new modules, and improve the joint learning ability of multimodal data. The models are repeatedly trained and validated, and evaluated based on confusion matrix, PR curve, AUC value, and F1 value to generate a state-of-the-art (SOTA) IVF outcome prediction model. Based on the SOTA-level IVF outcome prediction model, the target patient information is processed to generate personalized prediction reference schemes for assisted reproductive IVF cycles and clinical application value analysis information.

[0010] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for constructing an assisted reproductive IVF outcome prediction model based on a fusion architecture of LSTM and LSTM-CNN.

[0011] The assisted reproductive IVF outcome prediction model and construction method based on the LSTM and LSTM-CNN fusion architecture provided in this application have been approved.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0013] Figure 1 The flowchart illustrates a method for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, according to an embodiment of this application. Figure 2 The diagram shows a schematic of a device for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, according to an embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] The following is combined Figure 1 This paper describes a method for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture according to an exemplary embodiment of this application.

[0016] S101: Obtain core data and key parameter information for clinical stratification of assisted reproduction, as well as information on target patients.

[0017] In one implementation, the core data for assisted reproductive clinical stratification is based on the original data of approximately 80,000 assisted reproductive in vitro fertilization (IVF) cycles from a central hospital. During the acquisition process, invalid cycle data must first be screened out. Specifically, IVF cycles in which frozen embryos have been preserved but have not been transferred (such as cycles in which patients canceled the transfer for personal reasons), cycles in which the fertilization step was not carried out because no usable eggs were obtained during the egg retrieval operation or no usable sperm were obtained during the sperm retrieval operation, and cycles in which the patient's subsequent follow-up was lost and the complete outcome could not be tracked, are excluded. Only valid cycle data with complete procedures and traceable outcomes are retained.

[0018] Based on valid cycle data, the datasets are divided into two categories according to the fertilization method: conventional fertilization IVF and ICSI fertilization IVF. The conventional fertilization IVF dataset includes IVF cycle data using natural fertilization (sperm and egg combining naturally). For example, a 30-year-old female patient, with a BMI of 22 and a BMI of 24 for both partners, a female AMH of 2.5 ng / mL and a FSH of 6.2 mIU / mL, 12 antral follicles, and an endometrial thickness of 8 mm, has been infertile for 2 years due to fallopian tube blockage. This dataset contains cycle data using the conventional fertilization method. The ICSI IVF dataset contains IVF cycle data using intracytoplasmic sperm injection (ICSI), where a single sperm is artificially injected into an egg. For example, a 35-year-old female patient, with a BMI of 23 and a BMI of 25, has a female AMH of 1.8 ng / mL, an FSH of 7.5 mIU / mL, 10 antral follicles, and an endometrial thickness of 7.5 mm. She has been infertile for 3 years due to low sperm motility. This is a cycle data that includes sperm morphology images and was conceived using ICSI.

[0019] The parameters for routine IVF (Intracytoplasmic Sperm Fertilization) are collected based on core indicators affecting the outcome of routine IVF. Examples of parameter acquisition are as follows: For sperm motility, obtained through sperm quality testing equipment, in a typical IVF cycle, the male partner's sperm count was 25% grade A and 50% grade A+B, meeting the basic requirements for sperm motility in routine IVF. For the couple's age / BMI, obtained through medical records and physical examinations, the couple's ages were 29 and 31 years old, and their BMIs were 21.5 kg / m² and 23.8 kg / m², respectively, both within the clinically appropriate range. Regarding female reproductive hormones, blood tests were conducted. On day 3 of the woman's menstrual cycle, the levels were: AMH 2.2 ng / mL (reflecting ovarian reserve), FSH 5.8 mIU / mL (reflecting follicle-stimulating hormone level), LH 4.5 mIU / mL (reflecting luteinizing hormone level), E2 35 pg / mL (reflecting estradiol level), P 0.3 ng / mL (reflecting progesterone level), and PRL 18 ng / mL (reflecting prolactin level). All hormone levels were within the normal reference range.

[0020] Regarding the number of antral follicles in the woman, obtained through transvaginal ultrasound, a total of 11 antral follicles were found in both ovaries, indicating good ovarian reserve. Regarding endometrial thickness, obtained through ultrasound before embryo transfer, the woman's endometrial thickness was 9mm one day prior to transfer, meeting the thickness requirement for embryo implantation. Regarding the cause of female infertility, clinical examination and medical history analysis determined that the woman's infertility was due to endometriosis, clearly classifying it as organic infertility. Regarding the duration of infertility, the couple had been engaging in regular sexual intercourse without using contraception for 3 years without conceiving; the duration of infertility was recorded as 3 years.

[0021] ICSI IVF parameters, in addition to conventional IVF parameters, incorporate sperm morphology image data. The acquisition method and examples for conventional parameters are the same as for conventional IVF. An example of acquiring sperm morphology image data is as follows: images of the male's sperm are captured using professional microscopic imaging equipment. The images must clearly show the overall structure of the sperm (including the head, neck, and tail). The acquired sperm images include morphologically normal sperm (oval head, no neck curvature, and complete tail) and morphologically abnormal sperm (malformed head, curved neck, and incomplete tail). Subsequently, image analysis techniques can be used to further extract sperm morphological features, such as the proportion of normal sperm and the distribution of abnormal sperm types, providing image-dimensional data support for predicting ICSI fertilization outcomes.

[0022] The acquisition of target patient information must be consistent with the collection dimensions of the core data and key parameters of the clinical stratification of assisted reproduction mentioned above to ensure data correspondence and completeness. For example, if a target patient is a 32-year-old woman who plans to undergo assisted reproductive treatment, first determine that the fertilization method she intends to use is ICSI fertilization, and then collect her key information: the ages of the couple are 32 and 34 years old, respectively, and their BMIs are 22.3 kg / m² and 24.1 kg / m², respectively; the woman's reproductive hormone tests show that AMH is 2.0 ng / mL, FSH is 6.5 mIU / mL, LH is 4.8 mIU / mL, the number of antral follicles is 9, the endometrial thickness is 8.2 mm, and the man has been infertile for 2 years due to severe oligoasthenospermia; at the same time, sperm morphology images of the man are collected to obtain clear sperm morphology visualization data. Complete target patient information can be directly used as input for subsequent state-of-the-art (SOTA) level IVF outcome prediction models to ensure the personalization and accuracy of prediction results.

[0023] S102 preprocesses and performs feature engineering on the core data and key parameters of assisted reproductive clinical stratification to generate multimodal feature vectors that are compatible with two types of fertilization methods.

[0024] In one implementation, for numerical parameters in conventional IVF data, such as the proportion of grade A sperm, the proportion of grade A+B sperm, the couple's age / BMI, female reproductive hormones, antral follicle count, endometrial thickness, and duration of infertility, data cleaning is first performed by deduplication, missing value completion, and outlier removal using box plots. Next, continuous variables are standardized to eliminate the influence of dimensions, and categorical variables representing the causes of female infertility are coded and transformed. Finally, correlation analysis is used to identify high-contribution features, construct composite features, and generate a numerical feature vector suitable for conventional IVF. Data cleaning is first performed by deduplication, missing value completion, and outlier removal using box plots. In a batch of routine IVF data, three completely duplicate records (the same parameters from the same patient in the same cycle) were found, and these duplicate entries were directly deleted. For 20 records missing the "Grade A sperm percentage", the data was supplemented using the average sperm motility from other tests conducted on the same patient during the same period or the median sperm motility of men of the same age. Through box plot analysis, two abnormal values ​​were found in the "female age" (18 years old and 55 years old, respectively, exceeding the routine age range of 20-45 years for assisted reproduction). These were determined to be data entry errors and were removed to ensure data accuracy.

[0025] Continuous variables are then standardized to eliminate the influence of dimensions, and categorical variables of female infertility causes are coded and transformed. Continuous variables such as "BMI of both partners (unit: kg / m²)", "female AMH (unit: ng / mL)", and "number of antral follicles (unit: number)" are processed using the Z-score standardization formula (standardized value = (original value - mean) / standard deviation). For example, if a patient's original BMI is 24 kg / m², the mean BMI of this batch of data is 22 kg / m², and the standard deviation is 2.5 kg / m², the standardized value is (24-22) / 2.5 = 0.8. Categorical variables such as "causes of female infertility" (e.g., fallopian tube blockage, endometriosis, ovulation disorders) are converted into numerical formats recognizable by the model using unique hot coding, such as "fallopian tube blockage" coded as [1,0,0] and "endometriosis" coded as [0,1,0].

[0026] Finally, correlation analysis was used to identify high-contribution features, construct composite features, and generate a numerical feature vector suitable for conventional IVF. Pearson correlation analysis revealed that the correlation coefficients between "Grade A sperm ratio," "AMH," "antral follicle count," and "fertilization rate" were 0.68, 0.55, and 0.42, respectively (all > 0.4, indicating high-contribution features), and these three features were retained. Simultaneously, a "age-antral follicle count composite feature" (calculating the product of female age and antral follicle count) and a "hormonal balance feature" (calculating the ratio of FSH to LH) were constructed. Finally, the selected high-contribution features and the constructed composite features were integrated to form an 8-dimensional numerical feature vector (e.g., [standardized value of Grade A sperm ratio, standardized value of Grade A+B sperm ratio, standardized value of female age, standardized value of female BMI, standardized value of AMH, standardized value of antral follicle count, age-antral follicle count composite feature value, hormone balance feature value]), which is the numerical feature vector suitable for conventional IVF.

[0027] For numerical parameters in ICSI-induced IVF data that are common to conventional IVF, such as couple's age / BMI, female reproductive hormones, and antral follicle count, the same cleaning and standardization procedures as conventional IVF were used. Correlation analysis and feature importance assessment were conducted to screen numerical features influencing ICSI outcomes, generating a basic numerical feature vector for ICSI-induced IVF. Using the same cleaning and standardization procedures as conventional IVF, for the "couple's age" data, two duplicate records were deleted, and five missing "male age" records were added (using the average female age from the same cycle or the average male age from the same batch of patients). One outlier, "female age 48 years," was removed using a box plot. Continuous variables such as "female AMH" and "antral follicle count" were then Z-score standardized, and the categorical variable "cause of female infertility" was one-hot coded. The processing methods were completely consistent with conventional IVF to ensure uniform data processing standards.

[0028] Through correlation analysis and feature importance assessment, numerical features influencing ICSI outcomes were screened, generating a basic numerical feature vector for ICSI-induced IVF. Correlation analysis revealed correlation coefficients of 0.58, 0.45, and 0.38 between "female AMH," "antral follicle count," "male BMI," and ICSI "live birth rate," respectively. Feature importance assessment using the random forest algorithm showed that "female AMH," "antral follicle count," and "couple's age" had importance scores of 0.25, 0.21, and 0.18 (all ranking in the top 5). These five key numerical features were ultimately selected and integrated into a 5-dimensional basic numerical feature vector (e.g., [standardized values ​​of female age, male BMI, AMH, FSH, and antral follicle count]), which constitutes the basic numerical feature vector for ICSI-induced IVF.

[0029] For sperm morphology image data specific to ICSI fertilization IVF, noise reduction was first achieved by removing blurry or incomplete sperm images. Then, image enhancement was achieved through random flipping, ±15° rotation, and ±20% brightness adjustment. The processed images were then uniformly scaled to a fixed pixel size, and pixel values ​​were normalized. Finally, a convolutional neural network was used to extract high-dimensional visual features of sperm morphology edge contours and structural integrity, generating a high-dimensional visual feature vector of sperm morphology. Noise reduction was first achieved by removing blurry or incomplete sperm images. Of the 1000 sperm morphology images collected, 80 images blurred due to lens smudges and 50 images that did not capture complete sperm (only showing sperm fragments) were selected. These 130 invalid images were directly discarded, retaining 870 clear images containing complete sperm, ensuring image data quality.

[0030] Image enhancement is achieved through random flipping, ±15° rotation, and ±20% brightness adjustment. The processed images are then uniformly scaled to a fixed pixel size, and pixel values ​​are normalized. For each valid image, it is randomly flipped horizontally or vertically (enhancing image data diversity), and rotated at random angles of -15°, -10°, 5°, and 15° (simulating sperm morphology from different shooting angles). The brightness is randomly increased or decreased by 20% (adapting to image characteristics under different lighting conditions). All enhanced images are uniformly scaled to a 224×224 pixel size, and the Min-Max normalization formula (normalized value = (original pixel value - minimum pixel value) / (maximum pixel value - minimum pixel value)) is used to map pixel values ​​from the [0,255] range to the [0,1] range, eliminating the influence of differences in image size and pixel scale.

[0031] Finally, a convolutional neural network (CNN) is used to extract high-dimensional visual features of sperm morphology, including edge contours and structural integrity, generating a high-dimensional visual feature vector of sperm morphology. A pre-trained ResNet50 network is used as the CNN architecture, with fully connected layers removed and convolutional and pooling layers retained. The normalized 224×224×3 (RGB channels) sperm image is input into the CNN. Visual features such as sperm head edges, neck curvature, and tail integrity are extracted through the convolutional layers. The features are then mapped to a fixed-dimensional (e.g., 2048-dimensional) vector through a global average pooling layer. This vector contains the high-dimensional visual features of sperm morphology and can be directly used for subsequent multimodal feature fusion.

[0032] This study classifies, integrates, and fuses the numerical feature vectors of conventional IVF, the basic numerical feature vectors of ICSI IVF, and the visual feature vectors of sperm morphology to generate multimodal feature vectors suitable for both fertilization methods. The numerical feature vectors of conventional IVF are integrated separately to form numerical feature vectors adapted to conventional IVF. The classification, integration, and modal fusion of feature vectors are based on the core principle of "differences in data modal requirements for fertilization methods." The outcome of conventional IVF is mainly influenced by numerical parameters such as sperm motility (e.g., the proportion of grade A sperm), physiological parameters of both partners (age / BMI), and female reproductive health indicators (AMH, number of antral follicles, etc.), and accurate prediction can be achieved without relying on sperm morphology images. However, for ICSI IVF, which addresses cases of poor male sperm quality (e.g., severe oligoasthenospermia), sperm morphology (e.g., head integrity, degree of tail malformation) significantly affects fertilization and subsequent pregnancy outcomes, requiring a combination of numerical and image-based dual-modal data. Based on this logic, the feature vectors of the two fertilization methods are classified and integrated to ensure that the features are highly compatible with the model architecture (LSTM model of conventional fertilization IVF and LSTM-CNN fusion model of ICSI fertilization IVF), while avoiding the interference of redundant modal data on model training efficiency and prediction accuracy.

[0033] The numerical feature vector for conventional IVF fertilization is an 8-dimensional vector generated after data cleaning, standardization, feature selection, and composite feature construction. The specific dimensional composition and integration logic are as follows: For the core basic features (6 dimensions), there are standardized values ​​for the proportion of grade A sperm (reflecting the core indicator of sperm motility, such as 0.6 after standardization for a certain sample), standardized values ​​for the proportion of grade A+B sperm (supplementing the overall sperm motility level, such as 0.7 after standardization for a certain sample), standardized values ​​for the ages of both partners (standardized value of -0.2 for a 30-year-old woman and -0.1 for a 32-year-old man), standardized values ​​for the BMI of both partners (standardized value of -0.3 for a 22 kg / m² woman and 0.1 for a 24 kg / m² man), standardized value for the female AMH (standardized value of 0.4 after standardization for 2.5 ng / mL), and standardized value for the number of antral follicles (standardized value of 0.5 after standardization for 12 follicles). These features directly reflect the key physiological and reproductive parameters that affect the outcome of conventional fertilization and are the core basis for model learning.

[0034] For constructing composite features (2D), including "age-antral follicle count composite feature" (the product of the woman's age and the number of antral follicles, such as 30 years old × 12 follicles = 360, with a standardized value of 0.3) and "hormonal balance feature" (the ratio of FSH to LH, such as FSH 6.2 mIU / mL, LH 4.5 mIU / mL, with a ratio of 1.38, and a standardized value of 0.2). These features further enhance the model's ability to capture numerical time-series correlation features by mining the correlation between basic parameters (such as the negative impact of age on the number of antral follicles, and the reflection of ovarian function by the FSH / LH ratio).

[0035] During integration, the above 8-dimensional features are directly concatenated in the order of "basic features + composite features" to form a numerical feature vector with fixed dimensions (e.g., [0.6, 0.7, -0.2, -0.1, -0.3, 0.1, 0.4, 0.5, 0.3, 0.2]). No image modality data is introduced. On the one hand, in the clinical practice of conventional IVF, doctors can determine whether conventional IVF is suitable only by detecting sperm motility (rather than morphology). On the other hand, according to previous correlation analysis, the correlation coefficient between sperm morphology and the outcome of conventional IVF is only 0.12 (far lower than 0.68 for the proportion of grade A sperm and 0.55 for AMH). Including image data will increase the computational load of the model (e.g., processing an additional 224×224 pixel image requires more than 3 times the computational resources) and may introduce irrelevant noise (e.g., fertilization failure cases caused by normal sperm morphology but insufficient motility), which may reduce the prediction accuracy.

[0036] The conventional IVF prediction model employs an LSTM (Long Short-Term Memory) architecture, whose core advantage lies in capturing the temporal correlation features of numerical parameters (such as the long-term impact of AMH fluctuations on live birth rate and the dynamic correlation between sperm motility and fertilization rate). The separately integrated numerical feature vector perfectly matches the input requirements of LSTM—the input layer dimension is set to 8 (consistent with the feature vector dimension), allowing direct learning of the temporal evolution of different numerical features during the IVF cycle through LSTM's gating mechanisms (input gate, forget gate, output gate) (such as the impact of FSH changes during ovulation induction on the number of oocytes retrieved, and thus the correlation with fertilization rate). For example, after inputting the numerical feature vector of a conventional IVF sample into the LSTM model, the model can filter out short-term fluctuations in less influential features such as male BMI through the forget gate, and strengthen the temporal weights of core features such as the proportion of grade A sperm and AMH through the input gate, ultimately outputting accurate prediction probabilities for six outcomes. This meets clinical prediction needs without relying on image data (the PR values ​​for all six outcomes in the test set are ≥0.91, demonstrating the effectiveness of the standalone numerical modality).

[0037] The feature fusion module performs dimensionality matching and weight allocation between the basic numerical feature vector of ICSI fertilization IVF and the high-dimensional visual feature vector of sperm morphology, constructing a dual-modal feature vector of numerical and image data. The feature fusion of ICSI fertilization IVF, centered on "basic numerical features + visual image features," is clearly different from the separate numerical integration of conventional fertilization IVF. Regarding the preprocessing adaptation of the basic numerical feature vector, the cleaning and standardization process of the ICSI fertilization IVF basic numerical feature vector (5 dimensions, including standardized values ​​of female age, male BMI, AMH, FSH, and antral follicle count) is completely consistent with that of conventional fertilization IVF (e.g., a standardized value of 0.8 for a 35-year-old female and -0.2 for an AMH level of 1.8 ng / mL), ensuring uniform numerical feature processing standards for both fertilization methods.

[0038] Regarding the fusion logic of image visual feature vectors, the high-dimensional visual feature vector of sperm morphology (2048 dimensions, with edge contour and structural integrity features extracted by CNN) is matched with the basic numerical feature vector through a feature fusion module to achieve dimensionality matching (upgrading the 5-dimensional numerical vector to 2048 dimensions through a fully connected layer) and dynamic weight allocation (e.g., the weight of sperm head integrity feature is 0.6, and the weight of AMH in the numerical features is 0.4), constructing a bimodal feature vector. This type of fusion is adapted to the LSTM-CNN model architecture of ICSI fertilization IVF. The CNN module is responsible for parsing image features, and the LSTM module is responsible for processing numerical temporal features. After the two are fused through a fully connected layer, the combined influence of "abnormal sperm morphology (e.g., head malformation) + advanced female age" on ICSI outcome is fully explored (PR values ​​of all 6 outcomes in the test set are ≥0.92, an improvement of 0.03 compared to the model without images).

[0039] Simultaneously, the datasets corresponding to the two types of feature vectors were divided into training, validation, and test sets in an 8:1:1 ratio, ultimately generating differentiated multimodal feature vectors adapted to conventional IVF and ICSI IVF respectively. After the two types of feature vectors were integrated, the training, validation, and test sets were divided in an 8:1:1 ratio (the conventional IVF dataset has 5000 samples, divided into 4000 training samples, 500 validation samples, and 500 test samples; the ICSI IVF dataset has 3000 samples, divided into 2400 training samples, 300 validation samples, and 300 test samples), and the two datasets were completely independent to avoid data cross-contamination.

[0040] Model input validation was performed using 8-dimensional numerical feature vectors for conventional IVF. Training set samples were input into the LSTM model in batches of 32. During training, the loss function (cross-entropy) decreased from the initial 1.2 to 0.3, and the validation set accuracy increased from 55% to 88%. Furthermore, the PR values ​​of the six outcomes in the test set (such as the cancellation rate of 0.931 and the live birth rate of 0.910) all met the clinical prediction requirements. This demonstrates that the "integration of numerical feature vectors alone" approach not only did not suffer performance loss due to the lack of image modalities, but also improved model training efficiency due to the simplification of feature dimensions (8-dimensional vs. 2048-dimensional ICSI) (training time was reduced by 40% compared to the ICSI model), making it fully adaptable to the prediction scenarios and model architecture of conventional IVF.

[0041] S103, based on multimodal feature vectors, constructs and optimizes two types of in vitro fertilization prediction models for reproductive methods, and generates initial outcome prediction models and performance evaluation reference information.

[0042] In one implementation, LSTM (Long Short-Term Memory) is selected as the core architecture for the numerical feature vector specific to conventional IVF. The input layer dimension is set to the number of filtered features, and the output layer is the predicted probability of 6 outcomes. Cross-entropy is used as the loss function, Adam as the optimizer, and the batch size is set to 32. The model is trained on an 8:1:1 training set, and the loss changes and fitting trends are recorded in real time during the training process to generate an initial LSTM prediction model for conventional IVF. First, the core architecture and parameter configuration of the model are determined. LSTM is selected as the core architecture. After feature filtering, the numerical feature vector dimension of conventional IVF is 8 (including the proportion of grade A sperm, standardized values ​​such as AMH, and composite features), so the input layer dimension is set to 8. The output layer corresponds to the predicted probability of 6 IVF outcomes (fertilization rate, canceled cycle rate, biochemical pregnancy rate, clinical pregnancy rate, miscarriage rate, and live birth rate), so the output layer dimension is 6. Cross-entropy was used as the loss function (to meet the needs of multi-class probability prediction), Adam was used as the optimizer (balancing training efficiency and convergence stability), and the batch size was set to 32 (balancing computational efficiency and memory usage). Model training was then conducted on an 8:1:1 training set. The standard IVF dataset contained 5000 samples, with 4000 samples in the training set. These samples were input into the model in batches of 32. The training loss was calculated after each iteration, and changes in loss (e.g., an initial loss of 1.2 decreasing to 0.3 after 50 iterations) and the fitting trend (e.g., training accuracy increasing from 55% to 88%) were recorded in real time. Training continued until the training loss stabilized, ultimately generating an initial LSTM prediction model for standard IVF. This model can initially output the predicted probabilities of six outcomes.

[0043] To address the dual-modal feature vectors of numerical and image data in ICSI-induced IVF, an LSTM+CNN feature fusion architecture was adopted. The CNN module extracts high-dimensional visual features from sperm morphology images, while the LSTM module processes numerical temporal features. The outputs of both are fused through a fully connected layer to output the prediction result. Weighted cross-entropy was used as the loss function, Adam as the optimizer, and a batch size of 32 was set. k-fold cross-validation was employed to ensure model generalization. Specifically, the LSTM+CNN feature fusion architecture used a pre-trained ResNet50 network (with fully connected layers removed) to extract high-dimensional visual features (such as head edges and tail integrity) from 224×224 pixel sperm morphology images, with an output dimension of 2048. The LSTM module processed the 5-dimensional ICSI basic numerical feature vector (including female age, AMH, etc.), with 128 hidden layer units, and processed numerical temporal correlation features. The outputs of both were fused through a fully connected layer (with 512 nodes), and then the output layer output six outcome prediction probabilities. We used weighted cross-entropy as the loss function (because some outcome samples in the ICSI dataset are imbalanced, such as the low proportion of live birth rate samples, we improved the prediction accuracy of minority classes by adjusting the weights), Adam as the optimizer, and set the batch size to 32. We also used 5-fold cross-validation (dividing the training set into 5 parts, 4 parts for training and 1 part for validation, and looping 5 times to ensure generalization) to ensure the generalization of the model.

[0044] Multimodal collaborative training was conducted based on the training set to optimize the fusion weights and collaborative learning efficiency of the two types of features, generating an initial LSTM+CNN prediction model for ICSI fertilization IVF. Multimodal collaborative training was performed on the ICSI training set (2400 samples). During each training iteration, numerical feature vectors and corresponding sperm morphology images were simultaneously input. The parameters of the CNN and LSTM were adjusted through backpropagation to optimize the fusion weights of the two types of features (e.g., the initial image feature weight was 0.4, adjusted to 0.6 after 30 training iterations to improve the image's contribution to outcome prediction) and collaborative learning efficiency (e.g., strengthening the association between morphologically normal sperm images and live birth rate through an attention mechanism). Training continued until the validation set loss stabilized, generating an initial LSTM+CNN prediction model for ICSI fertilization IVF. This model can combine numerical and image features to output six outcome prediction results.

[0045] For the initial model of conventional IVF, hyperparameters such as learning rate, batch size, number of LSTM network layers, and regularization coefficient were adjusted using grid search or Bayesian optimization. Candidate values ​​for the learning rate were set to [0.001, 0.0005, 0.0001], batch size to [16, 32, 64] (adjusted based on the initial batch size of 32), LSTM network layers to [1, 2, 3], and L2 regularization coefficient to [0.001, 0.01, 0.1]. Grid search was used to iterate through all parameter combinations. When the learning rate was 0.001, the batch size was 32, the number of LSTM network layers was 2, and the regularization coefficient was 0.01, the average PR value of the six outcomes on the validation set reached 0.91, and the difference between the training loss and the validation loss was the smallest (only 0.03), thus determining it as the optimal hyperparameter combination. Record the model performance under different parameter combinations (e.g., when the learning rate is 0.0001, the validation set PR value drops to 0.88), generate hyperparameter optimization curves (the horizontal axis is the learning rate, and the vertical axis is the validation set PR value, showing a trend of first rising and then falling), and clarify the target hyperparameter configuration.

[0046] For the initial ICSI fertilization IVF model, additional optimizations were made to the parameters of CNN convolutional kernel number / size, LSTM layer unit number, and fully connected layer node number. Candidate values ​​for CNN convolutional kernel number / size were: convolutional layer 1 set to [32(3×3), 64(3×3)], convolutional layer 2 set to [64(3×3), 128(3×3)] (to meet the requirements for sperm morphology detail extraction); candidate values ​​for LSTM layer unit number were [64, 128, 256]; and candidate values ​​for fully connected layer node number were [256, 512, 1024]. Bayesian optimization was used to efficiently search for the optimal combination. For example, when the CNN convolutional kernel number was "32(3×3) + 64(3×3)", the LSTM layer unit number was 128, the fully connected layer node number was 512, and the learning rate was 0.0001 with a batch size of 32, the average PR value of the six outcomes on the validation set reached 0.92, and the contribution of image features to the prediction was improved by 15%. Record the impact of parameter changes on performance (e.g., when the number of convolutional kernels increases to 128, the risk of model overfitting increases and the validation set PR value drops to 0.89) to determine the target hyperparameter configuration.

[0047] During training, the validation set loss and accuracy are monitored in real time. If the validation set loss does not decrease for 10 consecutive rounds, an early stopping mechanism is triggered to prevent overfitting. For a conventional IVF model, when the iteration reaches 30 rounds, the validation set loss drops to 0.28. After that, for 10 consecutive rounds (rounds 31-40), the validation set loss remains between 0.28 and 0.29 without decreasing, triggering the early stopping mechanism to stop training and avoid the model overfitting the training data (at this point, the training loss has dropped to 0.22; if training continues, the training loss will further decrease, but the validation set loss will increase).

[0048] For the ICSI model, the validation set loss stabilized after 45 iterations, and early stopping was triggered after 10 consecutive iterations without a decrease. Simultaneously, the model performance changes under different hyperparameter combinations were recorded (e.g., the validation set accuracy of the conventional model with a learning rate of 0.001 was 3% higher than that of the model with a learning rate of 0.0001). Target hyperparameter configurations for both types of models were selected, and overfitting prevention features were generated (e.g., after early stopping, the training-validation loss difference of the conventional model stabilized at 0.06, and that of the ICSI model stabilized at 0.08, both within a reasonable range), providing a basis for further in-depth model development.

[0049] Record the performance changes of the model under different hyperparameter combinations, select the target hyperparameter configurations for the two types of models, and generate hyperparameter optimization curves and overfitting prevention effect features. For the conventional IVF initial LSTM model and the ICSI IVF initial LSTM+CNN model, record the performance changes under different hyperparameter combinations, select target configurations and generate relevant features. First, determine the hyperparameter search range, set the candidate values ​​for learning rate to [0.001, 0.0005, 0.0001], batch size to [16, 32, 64], LSTM network layer number to [1, 2, 3], and regularization coefficient (L2 regularization) to [0.001, 0.01, 0.1]. By traversing all combinations using a grid search, the validation set performance for each parameter group was recorded. When the learning rate was 0.001, batch size was 32, LSTM layers were 2, and regularization coefficient was 0.01, the average PR value for the six outcomes on the validation set was 0.91, the average F1 score was 0.89, and the difference between training loss and validation loss was 0.03 (low risk of overfitting). When the learning rate was 0.0001, batch size was 16, LSTM layers were 3, and regularization coefficient was 0.1, the average PR value on the validation set decreased to 0.87, the average F1 score decreased to 0.85, and the training time increased by 30%. After comparing all combinations, "learning rate = 0.001, batch size = 32, LSTM layers = 2, regularization coefficient = 0.01" was selected as the standard target hyperparameter configuration for the model. Meanwhile, a hyperparameter optimization curve was plotted with the learning rate on the horizontal axis and the average PR value of the validation set on the vertical axis. The curve shows that the learning rate reaches its peak at 0.001, and both values ​​below and above this value lead to performance degradation. Overfit prevention and control features were generated, and the training-validation loss difference was recorded as 0.03 when the early stopping mechanism was triggered (validation set loss did not decrease for 10 consecutive rounds), which is much lower than 0.12 when not optimized, proving that overfit prevention and control is effective.

[0050] Based on the hyperparameters of the regular model, additional CNN-related parameter ranges are set: the candidate values ​​for the number / size of CNN convolutional kernels are "convolutional layer 1: 32 (3×3), 64 (3×3); convolutional layer 2: 64 (3×3), 128 (3×3)", the candidate values ​​for the number of units in the LSTM layer are [64, 128, 256], and the candidate values ​​for the number of nodes in the fully connected layer are [256, 512, 1024]. Bayesian optimization was used to optimize the search efficiency, and the performance of key combinations was recorded. For example, when the number of CNN convolutional kernels = "32 (3×3) + 64 (3×3)", the number of LSTM units = 128, the number of fully connected layer nodes = 512, the learning rate = 0.0001, and the batch size = 32, the average PR value of the six outcomes on the validation set was 0.92, the average AUC value was 0.93, and the contribution of image features to the performance (compared to the model without images) was improved by 15%. When the number of CNN convolutional kernels = "128 (3×3) + 128 (3×3)", the number of LSTM units = 256, and the number of fully connected layer nodes = 1024, the average PR value on the validation set dropped to 0.89, and gradient vanishing occurred during training (the validation set loss fluctuated and increased). The final selected hyperparameter configuration for the ICSI model was "CNN convolutional kernel number = 32 (3×3) + 64 (3×3), LSTM unit number = 128, fully connected layer node number = 512, learning rate = 0.0001, batch size = 32". Hyperparameter optimization curves (such as the relationship between the number of CNN convolutional kernels and the validation set PR value) were plotted, showing that an excessive number of convolutional kernels leads to overfitting. The overfitting prevention feature record showed that the training-validation loss difference was 0.04 when early stopping was triggered, and the performance standard deviation of k-fold cross-validation (5-fold) was 0.02, proving the model's stable generalization ability.

[0051] The optimized initial models were evaluated on test sets and validated using multiple metrics. Initial outcome prediction models and comprehensive performance evaluation reference information were generated. The target hyperparameter configurations of both models were applied to the test sets. Performance was evaluated using multidimensional metrics including PR curves, accuracy, F1 score, and AUC. The test results for the conventional IVF initial model showed PR values ​​≥0.91 for all six outcomes, while the ICSI IVF initial model showed PR values ​​≥0.92 for all six outcomes. The target hyperparameter configurations of the two models were then applied to independent test sets (500 samples for the conventional model and 300 samples for the ICSI model), and multidimensional evaluation was conducted using PR curves, accuracy, F1 score, and AUC. For the initial model of conventional fertilization IVF, the PR values ​​for the six outcomes in the test set were as follows: canceled cycle rate 0.931, biochemical pregnancy rate 0.924, clinical pregnancy rate 0.921, fertilization rate 0.917, miscarriage rate 0.913, and live birth rate 0.910 (all ≥0.9); the mean accuracy was 0.90, the mean F1 score was 0.89, and the mean AUC was 0.91. The PR curve shows that the PR values ​​for all outcomes were in the high range with no significant fluctuations, demonstrating the model's predictive stability.

[0052] For the initial ICSI-induced IVF model, the PR values ​​for the six outcomes in the test set were: canceled cycle rate 0.942, biochemical pregnancy rate 0.939, clinical pregnancy rate 0.932, fertilization rate 0.930, miscarriage rate 0.928, and live birth rate 0.925 (all ≥0.92); the mean accuracy was 0.92, the mean F1 score was 0.91, and the mean AUC was 0.93. Comparing the test performance with and without sperm morphology images, the former had a mean PR value 0.03 higher than the latter, demonstrating the gain effect of image features.

[0053] The test set performance data, hyperparameter configuration information, and multimodal fusion effects of the two models were summarized to generate an initial outcome prediction model and comprehensive performance evaluation reference information including prediction accuracy, stability, and feature contribution. The test set data, hyperparameter configuration, and fusion effects of the two models were summarized to form complete evaluation information. Regarding prediction accuracy: the conventional model's PR values ​​for the six outcomes are ≥0.91, and the ICSI model's are ≥0.92, both superior to models in the current research stage. Regarding stability: the standard deviation of the conventional model's test set performance is 0.01, while that of the ICSI model is 0.008, and the performance fluctuation of k-fold cross-validation is ≤0.02, proving the model's stability. Regarding feature contribution: in the conventional model, the feature importance scores of "Grade A sperm ratio," "AMH," and "antral follicle number" are the highest (all >0.2), while in the ICSI model, the importance score of "sperm morphology image features" is 0.28 (higher than the average of 0.22 for numerical features), identifying key influencing factors.

[0054] Regarding hyperparameter configuration: Detailed records of the target hyperparameters for both types of models (e.g., learning rate of 0.001 for the conventional model, CNN convolutional kernel configuration for the ICSI model, etc.) were kept to provide a benchmark for subsequent in-depth development. Regarding multimodal fusion performance: Due to the fusion of image features, the ICSI model improved the live birth rate prediction accuracy by 5% compared to a single numerical model, proving the effectiveness of the fusion strategy. Finally, initial outcome prediction models (conventional IVF initial LSTM model, ICSI IVF initial LSTM+CNN model) and the above comprehensive performance evaluation reference information were generated to provide a basis for in-depth model iteration and clinical application.

[0055] S104. The initial outcome prediction model is iteratively processed to increase the model depth, introduce new modules, improve the joint learning capability of multimodal data, and repeatedly train and validate it. The model is evaluated based on the confusion matrix, PR curve, AUC value, and F1 value to generate a state-of-the-art IVF outcome prediction model.

[0056] In one implementation, the network structure and inter-layer connection methods of the initial LSTM model for conventional IVF are deeply optimized to improve the ability to capture numerical temporal features and generate iterative optimization features for the conventional IVF model. For the initial LSTM model of conventional IVF, optimization is performed starting from the network structure and inter-layer connection methods to improve the ability to capture numerical temporal features. Specific steps and examples are as follows: First, the limitations of the initial model are analyzed. It is found that the initial LSTM model (1 hidden layer and ordinary fully connected layer) is insufficient in capturing temporal features such as "the correlation between the proportion of grade A sperm over time and fertilization rate" and "the long-term impact of AMH fluctuations on live birth rate." The F1 value for the live birth rate prediction is only 0.88. Based on this, we carried out in-depth optimization: First, we increased the LSTM hidden layers to two, with the first layer having 128 units and the second layer having 64 units. We solved the gradient vanishing problem through inter-layer residual connections, ensuring that the deep network can effectively learn long-term features. Second, we optimized the connection between the input layer and the LSTM layer by introducing an attention mechanism module. We assigned higher attention weights to key temporal parameters such as "antral follicle count" and "AMH" (e.g., the attention weight of AMH was increased from 0.3 to 0.5), strengthening the temporal correlation learning of core parameters. Third, we adjusted the activation function of the output layer, replacing the original Sigmoid function with the Softmax function, which is more suitable for the multi-class probability output requirements of 6 outcomes. After optimization, the model was retrained using the training set (4000 samples). The training loss was recorded in real time, decreasing from the initial 0.3 to 0.22. The F1 score for live birth rate prediction on the validation set increased to 0.92. Iterative optimization features of the conventional IVF model were generated, including the network structure features of "2-layer LSTM + residual connection + attention mechanism", the key parameter attention weight distribution features (such as AMH weight 0.5 and antral follicle number weight 0.4), and the features that improved the accuracy of temporal feature capture (the accuracy of live birth rate temporal correlation prediction was improved by 15%).

[0057] This paper innovatively improves the feature fusion strategy and multimodal collaborative mechanism of the initial LSTM+CNN model for ICSI fertilization IVF, enhancing the joint learning capability of numerical and image features and generating optimized fusion features for the ICSI fertilization IVF model. Specifically, the initial model uses a simple fusion method of "directly concatenating image features extracted by CNN with numerical features from LSTM," resulting in low efficiency in the collaborative learning of image features (such as sperm head malformation) and numerical features (such as the woman's age). The PR value for fertilization rate prediction in the validation set is only 0.91. The optimization measures include: First, improving the feature fusion strategy by replacing the original direct splicing with "cross-attention fusion". The 2048-dimensional image features extracted by CNN and the 64-dimensional numerical features output by LSTM interact dynamically through the cross-attention module. The collaborative weight of "sperm head integrity" in the image features and "male BMI" in the numerical features is increased from 0.25 to 0.4, strengthening the association learning between the two types of features. Second, optimizing the multimodal collaboration mechanism by adding an "image feature filtering layer" after the CNN module. This layer filters image features strongly related to IVF outcomes (such as removing blurred features at the sperm tail) through a 1×1 convolutional layer, reducing redundant information interference. At the same time, a "temporal feature smoothing layer" is added to the LSTM module, using the sliding window mean method to reduce the impact of instantaneous fluctuations in numerical parameters (such as FSH). Third, adjusting the structure of the fully connected layer after fusion by changing the original single-level structure of 512 nodes to a two-level structure of "256 nodes + 128 nodes". Overfitting is prevented by using a Dropout layer (dropout probability 0.3). After optimization, multimodal collaborative training was conducted on the ICSI training set (2400 samples). The contribution of image features to fertilization rate prediction increased from 20% to 35%, and the average PR value of the six outcomes in the validation set increased from 0.92 to 0.93. The ICSI fertilization IVF model fusion optimization features were generated, including the collaborative mechanism features of "cross-attention fusion + image feature screening + temporal smoothing", the image-numerical feature collaborative weight distribution features (such as the collaborative weight of sperm head integrity and male BMI of 0.4), and the multimodal feature redundancy reduction feature (image redundant features reduced by 30%).

[0058] For the two types of iterative models, a three-level validation logic of training set-validation set-test set is adopted. After each iteration, the model parameters are updated based on the training set, the optimization direction is adjusted based on the validation set, and the performance is initially evaluated based on the test set. For the two types of iterative models, namely the conventional model and the ICSI model, a three-level validation logic of "training set-validation set-test set" is adopted to ensure the accuracy of the direction of each iteration. The specific steps and examples are as follows: In each iteration, the optimized model is fully trained on the training set (4000 samples for the conventional model and 2400 samples for the ICSI model). Mini-batch gradient descent (batch size 32) is used to update the parameters. For example, in the 5th iteration of the conventional model, the training set loss decreased from 0.25 to 0.23, and the fertilization rate prediction accuracy increased from 0.90 to 0.92. The parameter update magnitude is recorded (e.g., the average update magnitude of LSTM layer weights is 0.02, and the update magnitude of attention module weights is 0.05).

[0059] After each training round, performance was evaluated using a validation set (500 samples for the standard model and 300 samples for the ICSI model), and the optimization direction was adjusted based on the results. In the third iteration of the ICSI model, the validation set liveness prediction PR value dropped from 0.92 to 0.91. Analysis revealed that the excessive number of CNN convolutional kernels (128) led to overfitting. The number of convolutional kernels was then adjusted to 64, and the validation set PR value rebounded to 0.93 in the fourth round, clearly indicating the optimization direction of "reducing the CNN parameter size." After every three iterations, a preliminary evaluation was conducted using a test set (500 samples for the standard model and 300 samples for the ICSI model). After six iterations of the standard model, the average PR value of the six outcomes on the test set increased from 0.91 to 0.92, with the PR value for canceling the periodicity rate reaching 0.94, proving the effectiveness of the iteration direction. If the test set performance did not improve (e.g., the ICSI model's test set PR value remained at 0.93 in the seventh iteration), the iteration in that direction was paused, and the focus shifted to other optimization dimensions (e.g., adjusting the fusion weights).

[0060] The true positive / false positive rate of the prediction results was quantified using a confusion matrix. The accuracy and generalization of the model were comprehensively evaluated using PR curves, AUC values, and F1 scores. An iteration termination condition was set to avoid overtraining, generating model iteration-performance change curves and multi-index evaluation heatmaps. Taking the 8th iteration of the conventional model as an example, the confusion matrix showed that the true positive rate (TPR) for live birth prediction increased from 0.89 to 0.93, while the false positive rate (FPR) decreased from 0.12 to 0.08, demonstrating the improved classification accuracy of the model for "live birth" and "non-live birth". The confusion matrix of the 6th iteration of the ICSI model showed that the TPR for fertilization failure prediction of samples with abnormal sperm morphology reached 0.95, indicating a significant contribution from image features in classification. Analysis of PR curves and AUC values ​​showed that after iteration of the conventional model, the PR curves for all six outcomes shifted upwards, with the live birth rate PR increasing from 0.91 to 0.93 and the AUC increasing from 0.91 to 0.94. The ICSI model, due to image feature optimization, saw an increase of 0.03 in the area under the fertilization rate PR curve, demonstrating improved multimodal fusion performance. For comprehensive F1 score evaluation, both models showed significant improvements after iteration: the conventional model's live birth rate F1 score increased from 0.88 to 0.92, and the ICSI model's fertilization rate F1 score increased from 0.90 to 0.94. Furthermore, the difference in F1 scores between the test and validation sets was ≤0.02, indicating stable model generalization.

[0061] The termination conditions were set as follows: "validation set loss shows no decrease for 15 consecutive rounds" and "test set PR value fluctuates ≤0.01 for 10 consecutive rounds". For the conventional model, after 12 rounds, the validation set loss remained between 0.21 and 0.22 for 15 consecutive rounds, and the test set PR value fluctuated by 0.005, triggering the termination conditions. For the ICSI model, the termination conditions were met after 10 rounds. Simultaneously, a model iteration round-performance change curve was generated (the horizontal axis represents the iteration round, and the vertical axis represents the average PR value of the validation set; the conventional model curve tends to stabilize after 12 rounds, while the ICSI model stabilizes after 10 rounds), and a multi-index evaluation heatmap (the vertical axis represents the outcome index, the horizontal axis represents the iteration round, and the color intensity represents the F1 value, visually showing the performance changes of each outcome with iteration, such as the ICSI model's live productivity F1 value gradually increasing to 0.94 with iteration).

[0062] By integrating the iterative optimization features, model iteration rounds-performance change curves, and multi-index evaluation heatmaps of the two types of models, and through performance comparison and target configuration screening, a state-of-the-art IVF outcome prediction model is generated. The structural features of the conventional model ("2-layer LSTM + residual connection + attention mechanism") and key parameter weight distribution features, as well as the collaborative mechanism features of the ICSI model ("cross-attention fusion + image feature screening") and multimodal weight distribution features, are summarized. Simultaneously, the iteration rounds-performance change curves of the two types of models (convergence after 12 rounds for the conventional model and after 10 rounds for the ICSI model) and multi-index evaluation heatmaps (F1 values ​​for all 6 outcomes ≥ 0.92) are integrated.

[0063] The performance of the two iterated models was compared with the initial model and existing research models (such as Vitrolife's iDAscore model). The conventional model achieved a PR value ≥0.93 for all six outcomes (the initial model ≥0.91), and the ICSI model achieved a PR value ≥0.94 (the initial model ≥0.92), both outperforming existing models (the highest PR value being 0.88). Simultaneously, the target configuration was selected: the conventional model was ultimately configured as "2-layer LSTM (128+64 units) + residual connections + attention mechanism, Adam optimizer (learning rate 0.001), and cross-entropy loss function," while the ICSI model was ultimately configured as "CNN (64+32 convolutional kernels) + 2-layer LSTM (128+64 units) + cross-attention fusion, Adam optimizer (learning rate 0.0001), and weighted cross-entropy loss function."

[0064] The final model was built based on the target configuration and retrained on the full dataset (5000 regular samples and 3000 ICSI samples). The PR values ​​of all 6 outcomes on the test set of the regular model were ≥0.93, and those of the ICSI model were ≥0.94. The standard deviation of the performance in multiple rounds of validation (5-fold cross-validation) was ≤0.01, generating state-of-the-art IVF outcome prediction models, including the regular IVF OTA-LSTM model and the ICSI IVF OTA-LSTM+CNN model. It also achieved industry-leading levels in temporal feature capture and multimodal fusion.

[0065] S105, based on the SOTA-level IVF outcome prediction model, processes the target patient information to generate personalized prediction reference schemes for assisted reproductive IVF cycles and clinical application value analysis information.

[0066] In one implementation, the basic clinical parameters and fertilization method suitability features in the target patient information are classified, analyzed, and model matched. The format is validated using the parameter input specifications of the SOTA model to generate a patient information-model suitability quantification factor. First, the basic clinical parameters and fertilization method suitability features of the target patient are extracted and classified according to fertilization method suitability. For a 32-year-old target patient, the basic clinical parameters include the ages of both partners (32 and 34 years old), BMI (22.3 kg / m² and 24.1 kg / m²), female reproductive hormones (AMH 2.0 ng / mL, FSH 6.5 mIU / mL, LH 4.8 mIU / mL), number of antral follicles (9), endometrial thickness (8.2 mm), cause of infertility (severe oligoasthenospermia in the male partner), and duration of infertility (2 years). Simultaneously, an image of the male partner's sperm morphology (224×224 pixels, clearly showing the sperm head, neck, and tail structures) is provided. Based on the characteristics of the fertilization method, since the male has severe oligoasthenospermia and the success rate of conventional fertilization is low, ICSI fertilization is the appropriate method, so the SOTA-LSTM+CNN model of ICSI fertilization IVF is matched; if the patient has blocked fallopian tubes in the female (the male has normal sperm motility, with 28% of grade A sperm and 55% of grade A+B sperm), then the SOTA-LSTM model of conventional fertilization IVF is suitable.

[0067] Combine the parameter input of the corresponding SOTA model with the standard verification information format. The ICSI model requires that the numerical parameters be in Z-score standardized format (e.g., if the female's AMH is 2.0 ng / mL, it needs to be standardized to (2.0-1.9) / 0.5=0.2 according to the training set mean 1.9 ng / mL and standard deviation 0.5 ng / mL). The sperm morphology image needs to be 224×224 pixels and the pixel value needs to be normalized to the [0,1] interval. If the patient's original image is 512×512 pixels, it needs to be scaled to 224×224 pixels first, and then processed by Min-Max normalization (e.g., if the original value of a pixel is 180, it becomes (180-0) / (255-0)=0.706 after normalization). After verification, the fit quantification factor is calculated based on "parameter completeness (weight 0.4) + format compliance (weight 0.3) + fertilization method suitability (weight 0.3)". The formula is: Fit quantification factor = (number of complete parameters / total number of parameters) × 0.4 + (number of compliant parameters / total number of parameters) × 0.3 + fertilization method suitability score (suitability is 1, non-suitability is 0) × 0.3. Taking the above ICSI-fitted patient as an example, all 10 parameters (including images) are complete and compliant, and the fertilization method suitability score is 1. The calculated fit quantification factor is 1 × 0.4 + 1 × 0.3 + 1 × 0.3 = 1.0, indicating that the patient information is highly compatible with the ICSI-SOTA model. If a patient is missing one non-critical parameter (such as the male partner's BMI), the number of complete parameters is 9, compliance is 1, and suitability is 1, then the factor is 0.9 × 0.4 + 1 × 0.3 + 1 × 0.3 = 0.96, which is still within the high fit range.

[0068] The predicted probabilities of six IVF outcomes output by the state-of-the-art (SOTA) model were individually interpreted and risk-stratified. Safety and risk ranges were defined based on clinical experience thresholds, generating a quantitative factor for outcome prediction and risk stratification. The six predicted IVF outcomes include fertilization rate, cycle cancellation rate, biochemical pregnancy rate, clinical pregnancy rate, miscarriage rate, and live birth rate. Based on extensive clinical data, the clinical thresholds for the six outcomes were set as follows: fertilization rate ≥70% (safe), <50% (risk); cycle cancellation rate <5% (safe), ≥15% (risk); biochemical pregnancy rate <10% (safe), ≥20% (risk); clinical pregnancy rate ≥40% (safe), <25% (risk); miscarriage rate <15% (safe), ≥25% (risk); live birth rate ≥35% (safe), <20% (risk). The state-of-the-art (SOTA) model outputs the following outcome probabilities for a suitable ICSI patient: fertilization rate 82%, cycle cancellation rate 3%, biochemical pregnancy rate 8%, clinical pregnancy rate 45%, miscarriage rate 12%, and live birth rate 38%. Personalized interpretation: The fertilization rate, clinical pregnancy rate, and live birth rate are all within a safe range; the cycle cancellation rate, biochemical pregnancy rate, and miscarriage rate are below the risk threshold, indicating a good overall outcome. If another patient's predicted live birth rate is 18% (<20%) and miscarriage rate is 28% (≥25%), this is interpreted as "high risk of live birth and significant risk of miscarriage, requiring adjustment of the treatment plan."

[0069] The risk grading quantification factor is calculated based on the "percentage of safe items (weight 0.6) + severity of high-risk items (weight 0.4)". The formula is: Risk Grading Quantification Factor = (Number of safe items / 6) × 0.6 + [1 - (mean probability of high-risk items / highest risk threshold)] × 0.4. Taking the above patients with good outcomes as an example, all 6 items are safe, and the number of high-risk items is 0, so the calculated factor is 1 × 0.6 + 1 × 0.4 = 1.0. For patients with a live birth rate of 18% and a miscarriage rate of 28%, the number of safe items is 4 (fertilization rate, canceled cycle rate, biochemical pregnancy rate, clinical pregnancy rate), the mean probability of high-risk items is (18 + 28) / 2 = 23, and the highest risk threshold is 25. The factor is (4 / 6) × 0.6 + [1 - (23 / 25)] × 0.4 ≈ 0.4 + 0.032 = 0.432, indicating a higher risk.

[0070] Based on patient information-model fit quantitative factors and outcome prediction-risk grading quantitative factors, personalized prediction results are transformed into treatment plans and their value is evaluated, generating personalized prediction reference plans for assisted reproductive IVF cycles and clinical application value analysis information. Combining the two types of quantitative factors with clinical experience, a feasible IVF cycle reference plan is transformed. Highly compatible and safe patients (fitting factor 1.0, risk factor 1.0): For patients who are compatible with ICSI and have good outcomes, the recommended plan is "to use ICSI fertilization, select a mild stimulation protocol during ovulation induction (avoid overstimulation of the ovaries), select an endometrial thickness of 8-10mm during the transfer window, transfer one high-quality embryo, and strengthen luteal support after transfer (such as progesterone supplementation)." It also indicates that "due to a predicted live birth rate of 38% (safe range), recent cycles can be prioritized without additional pretreatment."

[0071] For highly matched patients at medium risk (fitting factor 0.96, risk factor 0.5): If a patient is matched with a conventional fertilization model, the predicted live birth rate is 23% (close to the risk threshold of 20%) and the miscarriage rate is 22% (close to the risk threshold of 25%). The recommended protocol is to "perform 1-2 months of pretreatment (such as DHEA supplementation to improve ovarian reserve), and start a conventional fertilization cycle after AMH increases to above 2.2 ng / mL. Blastocyst transfer should be selected during the transfer (to reduce the risk of miscarriage). Hormone levels should be closely monitored during the cycle (such as FSH < 7 mIU / mL)".

[0072] The value was assessed from three dimensions: patient, clinical, and social. For patients with high suitability and high safety, the value analysis was that "personalized solutions can avoid ineffective cycles (cost of 30,000-80,000 yuan per cycle), reducing the economic burden; the live birth rate is clear (38%), reducing mental anxiety caused by unknown success rates, and patient satisfaction is expected to increase by more than 40%"; for patients with medium risk, "pre-treatment solutions can increase the live birth rate to more than 30%, avoid 1-2 failed cycles, save 60,000-160,000 yuan, and the degree of mental stress relief is expected to reach 35%".

[0073] From a clinical perspective, the protocol can shorten the time doctors spend developing treatment plans (from the traditional 2-3 days to within 1 hour), improving treatment efficiency. Adjusting the protocol based on model predictions (such as prioritizing blastocyst transfer) can increase the overall IVF live birth rate at the center by 5-8%, reducing waste of medical resources (e.g., the use of ineffective ovulation-inducing drugs). From a social perspective, through accurate prediction and protocol optimization, the repeated implantation failure rate can be reduced (from 35% to below 25%), reducing medical resource consumption, while helping more infertile patients achieve their fertility needs and alleviating social pressure on fertility, demonstrating significant public health value. Finally, integrating the protocol and value analysis generates a complete "Personalized Prediction Reference Protocol for Assisted Reproductive Technology IVF Cycles" and a "Clinical Application Value Analysis Report," providing decision-making support for doctors and patients.

[0074] In one implementation, such as Figure 2 As shown, this application also provides a device for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, comprising: The acquisition module 201 is used to acquire core data and key parameter information of assisted reproductive clinical stratification, as well as target patient information. Among them, the core data of assisted reproductive clinical stratification includes raw data of IVF cycles, which are divided into two datasets according to the fertilization method: conventional IVF and ICSI IVF. The key parameter information includes conventional IVF parameters, which include the proportion of grade A sperm, the proportion of grade A+B sperm, the age / BMI of both partners, female reproductive hormones, number of antral follicles, endometrial thickness, cause of female infertility, and duration of infertility. ICSI IVF parameters include sperm morphology image data in addition to the conventional parameters. Processing module 202 is used to preprocess and feature-engineer core data and key parameter information of assisted reproductive clinical stratification to generate multimodal feature vectors adapted to two types of fertilization methods; based on the multimodal feature vectors, it constructs and optimizes in vitro fertilization prediction models for the two types of fertilization methods, generating an initial outcome prediction model and performance evaluation reference information; it performs performance iteration processing on the initial outcome prediction model, increases model depth, introduces new modules, improves the joint learning ability of multimodal data, repeatedly trains and verifies, and evaluates based on confusion matrix, PR curve, AUC value, and F1 value to generate a state-of-the-art (SOTA) IVF outcome prediction model; based on the SOTA-level IVF outcome prediction model, it processes the target patient information to generate a personalized prediction reference plan for assisted reproductive IVF cycles and clinical application value analysis information.

[0075] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the method, electronic device, electronic device, and readable storage medium for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture are basically similar to the above-described embodiment of the method for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, so the description is relatively simple. Relevant parts can be referred to in the description of the above-described embodiment of the method for constructing an assisted reproductive IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture.

Claims

1. A method for constructing an assisted reproductive technology IVF outcome prediction model based on an LSTM and LSTM-CNN fusion architecture, characterized in that, The method comprises the following steps: obtaining auxiliary reproductive clinical stratification core data and key parameter information, target patient information, wherein the auxiliary reproductive clinical stratification core data comprises IVF cycle original data, which is divided into two data sets of conventional fertilization IVF and ICSI fertilization IVF according to the fertilization method, the key parameter information comprises conventional fertilization IVF parameters, which include the proportion of A-grade sperm, the proportion of A+B-grade sperm, the age / BMI of the husband and wife, female reproductive hormones, antral follicle count, endometrial thickness, female infertility reason, infertility duration, and ICSI fertilization IVF parameters include sperm morphology image data on the basis of the conventional parameters; preprocessing and feature engineering processing are performed on the auxiliary reproductive clinical stratification core data and the key parameter information to generate multi-modal feature vectors suitable for the two fertilization methods; based on the multi-modal feature vectors, a reproductive in vitro fertilization prediction model for the two fertilization methods is constructed and optimized to generate an initial outcome prediction model and performance evaluation reference information; the performance of the initial outcome prediction model is iteratively processed, the model depth is increased, new modules are introduced, the multi-modal data joint learning ability is improved, and the model is repeatedly trained and verified, and the SOTA level IVF outcome prediction model is generated according to the confusion matrix, the PR curve, the AUC value and the F1 value; based on the SOTA level IVF outcome prediction model, the target patient information is processed to generate an auxiliary reproductive IVF cycle individualized prediction reference scheme and clinical application value analysis information.

2. The method of claim 1, wherein, The preprocessing and feature engineering processing are performed on the auxiliary reproductive clinical stratification core data and the key parameter information to generate multi-modal feature vectors suitable for the two fertilization methods, comprising: for the numerical parameters of the A-grade sperm proportion, the A+B-grade sperm proportion, the age / BMI of the husband and wife, the female reproductive hormones, the antral follicle count, the endometrial thickness and the infertility duration in the conventional fertilization IVF data, data cleaning is first completed by removing duplicates, filling missing values and box plot method; then the continuous variables are standardized to eliminate the dimension influence, and the classification variables of the female infertility reason are encoded; finally, high-contribution features are mined by correlation analysis, composite features are constructed, and numerical feature vectors suitable for conventional fertilization IVF are generated; for the numerical parameters of the age / BMI of the husband and wife, the female reproductive hormones and the antral follicle count in the ICSI fertilization IVF data which are common to the conventional fertilization IVF, the same cleaning and standardization process as the conventional fertilization IVF is adopted; numerical features affecting the ICSI outcome are selected by correlation analysis and feature importance evaluation to generate an ICSI fertilization IVF basic numerical feature vector; for the sperm morphology image data specific to ICSI fertilization IVF, first, invalid images with blurred or incomplete sperm are removed to complete denoising; then image enhancement is realized by random flipping, ±15° rotation and ±20% brightness adjustment; the processed images are uniformly scaled to a fixed pixel size and normalized pixel value; finally, high-dimensional visual features of sperm morphology such as edge contour and structural integrity are extracted by using a convolutional neural network to generate a sperm morphology high-dimensional visual feature vector; The numerical feature vectors of conventional fertilization IVF, the numerical feature vectors of ICSI fertilization IVF, and the visual feature vectors of sperm morphology are classified, integrated, and modally fused to generate multi-modal feature vectors suitable for two types of fertilization methods: The numerical feature vectors of conventional fertilization IVF are integrated separately to form numerical feature vectors suitable for conventional fertilization IVF. The numerical feature vectors of ICSI fertilization IVF and the high-dimensional visual feature vectors of sperm morphology are dimensionally matched and weight allocated by the feature fusion module to construct numerical + image dual-modal feature vectors. Meanwhile, the two types of feature vectors are divided into training set, validation set, and test set in the ratio of 8:1:1 to generate differentiated multi-modal feature vectors suitable for conventional fertilization IVF and ICSI fertilization IVF respectively.

3. The method of claim 1, wherein, Based on the multi-modal feature vectors, reproductive in vitro fertilization prediction models for the two types of fertilization methods are constructed and optimized to generate initial outcome prediction models and performance evaluation reference information, including: For the numerical feature vectors specific to conventional fertilization IVF, LSTM is selected as the core architecture, the input layer dimension is set to the number of filtered features, the output layer is set to 6 outcome prediction probabilities, cross-entropy is used as the loss function, Adam is used as the optimizer, and the batch size is set to 32; Model training is carried out based on the 8:1:1 divided training set, the loss change and fitting trend in the training process are recorded in real time, and the initial LSTM prediction model for conventional fertilization IVF is generated. For the numerical and image dual-modal feature vectors of ICSI fertilization IVF, an LSTM+CNN feature fusion architecture is used, the CNN module is responsible for extracting high-dimensional visual features of sperm morphology images, the LSTM module is responsible for processing numerical time series features, and the two are fused by a fully connected layer to output the prediction results; Weighted cross-entropy is used as the loss function, Adam is used as the optimizer, the batch size is set to 32, and k-fold cross-validation is used to ensure model generalization; Based on the training set, multi-modal collaborative training is carried out to optimize the fusion weight and collaborative learning efficiency of the two types of features, and the initial LSTM+CNN prediction model for ICSI fertilization IVF is generated. For the initial model of conventional fertilization IVF, the learning rate, batch size, LSTM network layer number, and regularization coefficient are adjusted through grid search or Bayesian optimization; For the initial model of ICSI fertilization IVF, the number / size of CNN convolution kernels, the number of LSTM layer units, and the number of nodes in the fully connected layer are additionally optimized; During the training process, the validation set loss and accuracy are monitored in real time, and if the validation set loss does not decrease for 10 consecutive rounds, the early stopping mechanism is triggered to prevent overfitting; The model performance changes under different hyperparameter combinations are recorded, and the target hyperparameter configurations of the two types of models are selected to generate hyperparameter optimization curves and overfitting prevention effect features.

4. The method of claim 3, wherein, Based on the multi-modal feature vectors, reproductive in vitro fertilization prediction models for the two types of fertilization methods are constructed and optimized to generate initial outcome prediction models and performance evaluation reference information, including: The optimized initial models of the two types are tested for performance evaluation and multi-index verification to generate initial outcome prediction models and comprehensive performance evaluation reference information: The target hyperparameter configurations of the two types of models are applied to the test set, and the performance is evaluated by multiple indicators such as PR curve, accuracy, F1 value, and AUC value; The test results of the initial model of conventional fertilization IVF show that the PR values of the six outcomes are all greater than or equal to 0.91, and the PR values of the six outcomes of the initial model of ICSI fertilization IVF are all greater than or equal to 0.92; The performance data of the test set, hyperparameter configuration information, and multi-modal fusion effect of the two types of models are summarized to generate an initial outcome prediction model and comprehensive performance evaluation reference information including prediction accuracy, stability, and feature contribution.

5. The method of claim 1, wherein, The performance of the initial outcome prediction model is iteratively processed, the model depth is increased, new modules are introduced, the multi-modal data joint learning ability is improved, and the SOTA level IVF outcome prediction model is generated based on repeated training and validation and evaluation by confusion matrix, PR curve, AUC value, and F1 value, including: The network structure and layer connection method of the initial LSTM model of conventional fertilization IVF are deeply optimized to improve the numerical time series feature capturing ability, and the iterative optimization features of the conventional fertilization IVF model are generated; The feature fusion strategy and multi-modal collaborative mechanism of the initial LSTM+CNN model of ICSI fertilization IVF are innovatively improved to enhance the numerical-image feature joint learning ability, and the fusion optimization features of the ICSI fertilization IVF model are generated; For the two types of iterative models, a three-level verification logic of training set-validation set-test set is adopted, the model parameters are updated based on the training set after each iteration, the optimization direction is adjusted based on the validation set, and the performance is preliminarily evaluated based on the test set; The true positive / false positive rate of the prediction results is quantified by confusion matrix, the accuracy and generalization of the model are comprehensively evaluated by PR curve, AUC value, and F1 value, the iteration termination condition is set to avoid overtraining, the model iteration round-performance change curve and multi-index evaluation heat map are generated; The iterative optimization features of the two types of models, the model iteration round-performance change curve, and the multi-index evaluation heat map are integrated, and the SOTA level IVF outcome prediction model is generated by performance comparison and target configuration screening.

6. The method of claim 5, wherein, Based on the SOTA level IVF outcome prediction model, the target patient information is processed to generate an assisted reproductive IVF cycle personalized prediction reference scheme and clinical application value analysis information, including: The basic clinical parameters and fertilization method adaptation features in the target patient information are classified and analyzed, and the model matching processing is performed, the format is verified according to the parameter input specification of the SOTA model, and the patient information-model adaptation quantization factor is generated; The six IVF outcome prediction probabilities output by the SOTA model are individually interpreted and risk classified, the safety and risk intervals are determined based on the clinical diagnosis and treatment experience threshold, and the outcome prediction-risk classification quantization factor is generated, wherein the six IVF outcome prediction probabilities include fertilization rate, cycle cancellation rate, biochemical pregnancy rate, clinical pregnancy rate, abortion rate, and live birth rate; Based on the patient information-model adaptation quantization factor and the outcome prediction-risk classification quantization factor, the individualized prediction results are converted into a scheme and evaluated in value to generate an assisted reproductive IVF cycle personalized prediction reference scheme and clinical application value analysis information.

7. An assisted reproductive technology IVF outcome prediction model construction device based on an LSTM and LSTM-CNN fusion architecture, characterized in that, The device comprises: The acquisition module is used for acquiring assisted reproductive clinical stratification core data and key parameter information, target patient information, wherein the assisted reproductive clinical stratification core data includes IVF cycle original data, which is divided into two data sets of conventional fertilization IVF and ICSI fertilization IVF according to fertilization methods, the key parameter information includes conventional fertilization IVF parameters, which include A-grade sperm proportion, A+B-grade sperm proportion, age / BMI of both husband and wife, female reproductive hormone, antral follicle number, endometrial thickness, female infertility reason, infertility duration, and ICSI fertilization IVF parameters including sperm morphology image data on the basis of conventional parameters; The processing module is used for pre-processing and feature engineering processing of the assisted reproductive clinical stratification core data and the key parameter information, and generating multi-modal feature vectors adaptive to the two fertilization methods; constructing and optimizing reproductive in vitro fertilization prediction models of the two fertilization methods based on the multi-modal feature vectors, generating initial outcome prediction models and performance evaluation reference information; performing performance iteration processing on the initial outcome prediction models, increasing model depth, introducing new modules, improving multi-modal data joint learning ability, repeatedly training and verifying and evaluating based on a confusion matrix, a PR curve, an AUC value and an F1 value, and generating an SOTA level IVF outcome prediction model; processing target patient information based on the SOTA level IVF outcome prediction model, and generating an assisted reproductive IVF cycle personalized prediction reference scheme and clinical application value analysis information.

8. An electronic device, comprising: It comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to perform the assisted reproductive IVF outcome prediction model construction method based on the LSTM and LSTM-CNN fusion architecture according to any one of claims 1-6. ​

Images