Intelligent prediction system and method for fetal growth limitation based on double networks

Abstract

The invention provides a fetal growth limitation intelligent prediction system and method based on double networks, and relates to the technical field of perinatal medical monitoring, and the system comprises a data collection and preprocessing module which is used for processing collected prenatal fetal heart related data and outputting high-quality data; the fetal heart signal feature extraction module is used for outputting fetal heart feature vectors by adopting a parallel architecture of a convolutional neural network and a long-short-term memory network; the multi-modal feature fusion module is used for outputting a fusion feature vector based on the double-tower structure architecture; the hierarchical classification prediction module is used for modeling a hierarchical structure of FGR clinical diagnosis by adopting a graph neural network GNN, and outputting prediction probabilities of FGR diagnosis, subtype classification and perinatal outcome in combination with an MC loss function optimization model; and the clinical decision support module is used for generating decision result suggestions and visual reports. And the clinical requirements of early accurate identification, subtype classification and perinatal outcome prediction are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of perinatal medical monitoring, in particular to a fetal growth restriction intelligent prediction system and method based on a double network. BACKGROUND

[0002] Fetal growth restriction (FGR) is a common and serious complication in the perinatal period, with an incidence of about 5-10%, and is the second leading cause of perinatal death in China. It can lead to acute complications such as neonatal asphyxia and hypoxic-ischemic encephalopathy, and is also associated with long-term neurodevelopmental disorders and increased risk of chronic diseases in adulthood. According to the time of diagnosis, FGR is divided into early-onset (before 32 weeks of gestation) and late-onset (after 32 weeks of gestation), with significant differences in pathological mechanisms and prognosis, necessitating precise identification and intervention.

[0003] Existing FGR monitoring and prediction techniques have many limitations: Ultrasound monitoring: As the gold standard for clinical diagnosis, it has measurement errors (±10-20%), diagnostic lag (requires significant fetal growth retardation to identify), low monitoring frequency (2-4 weeks / visit), and strong dependence on operators, making it difficult to achieve early warning.

[0004] Traditional electronic fetal heart monitoring (CTG): Manual interpretation is highly subjective, with inter-observer consistency coefficients (κ values) of only 0.2-0.6, high false positive rates, and low specificity, making it difficult to extract deep pathological and physiological information from fetal heart signals.

[0005] Computerized CTG analysis: Relies on manually defined features such as short-term variability (STV), making it difficult to capture complex nonlinear patterns in signals, with a prediction AUC usually <0.7, limiting its clinical utility.

[0006] Traditional machine learning methods: Require hand-designed features, have poor generalization ability, belong to shallow learning, cannot extract deep abstract features, and are difficult to integrate multi-modal data.

[0007] Based on the analysis of existing technologies, the core technical bottlenecks in FGR intelligent prediction mainly lie in the following five aspects: Firstly, the insufficient feature extraction is the key factor restricting the improvement of prediction performance. Traditional methods mainly rely on manually designed feature parameters by domain experts based on clinical experience, such as Short-Term Variability (STV), baseline variability, acceleration and deceleration patterns, and other statistical indicators. This feature engineering paradigm based on prior knowledge cannot adaptively discover and extract deep discriminative patterns hidden in complex fetal heart signals. Even if traditional machine learning methods such as support vector machines and random forests are used, the feature engineering process still highly depends on domain knowledge and dataset characteristics, which not only consumes time and effort, but also has limited generalization ability of the extracted features, making it difficult to transfer to different clinical scenarios and populations.

[0008] Secondly, the lack of utilization of time series information limits the in-depth mining of fetal heart dynamic evolution rules. Prenatal fetal heart monitoring data is essentially a time series signal containing rich physiological and pathological information. The inherent time series dynamic characteristics, short-term fluctuation patterns, and long-range correlations are of great value for FGR diagnosis. However, existing methods are mostly based on simple statistical indicators (such as mean, standard deviation, coefficient of variation, etc.) or basic time-domain features for analysis, using sliding windows or fixed time period statistics as input, without effectively modeling the dynamic dependence of fetal heart rate evolution over time and complex patterns across time scales, resulting in significant loss of time series information.

[0009] Thirdly, the lack of multi-modal fusion mechanism restricts the comprehensive utilization of multi-source heterogeneous information by the model. Existing researches mostly focus on single modality data analysis, mainly focusing on fetal heart monitoring time series, while ignoring important information sources such as maternal clinical characteristics (such as age, pregnancy number, parity, body mass index, pregnancy complications, etc.), ultrasound measurement parameters (fetal biological measurement values, amniotic fluid index, etc.), and Doppler hemodynamic indicators. Due to the lack of effective cross-modal information fusion framework and collaborative learning mechanism, multi-dimensional clinical information cannot be fully integrated and utilized, limiting the comprehensive performance of the prediction model.

[0010] Fourthly, the lack of hierarchical relationship modeling leads to logical inconsistency in the prediction results. From the perspective of clinical diagnosis and treatment logic, FGR diagnosis and perinatal outcome prediction have a clear hierarchical structure: first determine "whether there is FGR", then distinguish "early-onset or late-onset FGR", and finally predict "specific adverse perinatal outcome types" (such as perinatal death, neonatal asphyxia, acidosis, etc.). However, existing methods mostly use flat independent classification strategies, modeling each prediction task as an unrelated problem, without explicitly utilizing the hierarchical dependency between tasks and prior constraint knowledge, which may lead to logical contradictions in the prediction results (such as predicting a normal fetus but also predicting a high-risk adverse outcome), affecting the clinical credibility of the model.

[0011] Finally, the lack of model interpretability is an important obstacle for the clinical translation of deep learning methods. Although deep neural networks exhibit strong capabilities in feature learning and pattern recognition, their complex multi-layer nonlinear transformations render the model a "black box" property, making it difficult for clinicians to understand and verify the basis and reasoning process of the model's decision-making. The lack of interpretability not only affects the trust and acceptance of doctors to artificial intelligence systems, but also limits the application of the model in actual clinical scenarios, and may bring potential risks in medical safety and ethics. Therefore, how to improve the interpretability of the model while ensuring the prediction performance, so that its decision-making process conforms to the clinical cognitive logic, is a key technical problem to be solved.

[0012] In summary, the existing technology cannot meet the clinical needs of early accurate identification, subtype classification and perinatal outcome prediction of FGR, therefore, the present application proposes a fetal growth restriction intelligent prediction system and method based on double network. SUMMARY

[0013] The present application provides a fetal growth restriction intelligent prediction system and method based on double network, to solve the technical problems proposed above.

[0014] The present application provides a fetal growth restriction intelligent prediction system based on double network, comprising: A data acquisition and preprocessing module is used to acquire antenatal fetal heart monitoring time series data and structured clinical feature data, and after quality control, denoising, outlier processing, missing value filling, encoding and standardization preprocessing, high-quality data is outputted; A fetal heart signal feature extraction module is used to adopt a convolutional neural network (CNN) and a long short-term memory network (LSTM) parallel architecture, wherein the CNN extracts local pattern features of the high-quality data, and the LSTM extracts long-range time sequence dependence features of the high-quality data, and outputs a fetal heart feature vector; A multi-modal feature fusion module is used to output a fusion feature vector based on a double-tower structure architecture, wherein one tower processes structured clinical feature data through an artificial neural network (ANN) and outputs a clinical feature vector, and the other tower receives the fetal heart feature vector and fuses the two modal features in a high-dimensional semantic space; A hierarchical classification prediction module is used to model the hierarchical structure of FGR clinical diagnosis by adopting a graph neural network (GNN), combine with an MC loss function to optimize the model, and output the prediction probability of FGR diagnosis, subtype classification and perinatal outcome; A clinical decision support module is used to generate FGR risk score, diagnosis conclusion, subtype classification result, perinatal outcome risk assessment, individualized clinical advice and visual report according to the prediction probability.

[0015] Preferably, the data acquisition and preprocessing module comprises: Fetal heart monitoring data acquisition unit, using standard electronic fetal heart monitor to collect 20-40 minutes continuous time series data with 4Hz sampling frequency, synchronously collecting monitoring graph and original fetal heart rate value sequence, and marking body position change, fetal movement event; Clinical feature data acquisition unit, for collecting maternal basic information, this pregnancy information, fetal ultrasound information and perinatal outcome information, wherein the maternal basic information includes age, pregnancy times, parity, pre-pregnancy body mass index BMI, and previous adverse pregnancy history, the this pregnancy information includes monitoring gestational age, gestational hypertension, gestational diabetes, the fetal ultrasound information includes ultrasound estimated fetal weight, abdominal circumference, amniotic fluid index, and umbilical artery blood flow index, and the perinatal outcome information includes delivery gestational age, delivery mode, neonatal birth weight, Apgar score, umbilical artery blood gas analysis result, and NICU admission situation; Label definition unit, using ultrasound estimated fetal weight or abdominal circumference < corresponding gestational age 10th percentile and affected by pathological factors as FGR diagnosis gold standard, diagnosing early-onset FGR before 32 weeks of gestation, diagnosing late-onset FGR at or after 32 weeks of gestation, and adverse perinatal outcomes including premature birth, neonatal acidosis, neonatal asphyxia, and hypoxic-ischemic encephalopathy; Data preprocessing unit, for preprocessing fetal heart data to obtain high-quality data; Data set division unit, for stratified random sampling dividing training set, validation set and test set according to a preset proportion, and training, validating and testing convolutional neural network CNN and long short-term memory network LSTM.

[0016] Preferably, the network architecture of the convolutional neural network CNN comprises: Input layer: receiving preprocessed fetal heart sequence with shape (batch_size, sequence_length, 1), wherein sequence_length is the number of time steps, and batch_size is the batch size; Feature extraction layer: composed of 6-8 layers of residual blocks, with convolution kernel size of 3-7, filter number of 64-256 and increasing layer by layer, step length of 1, and activation function of ReLU; Cave convolution layer: introducing a cave convolution with dilation rate sequence [1, 2, 4, 8, 16]; adding a batch normalization layer after each convolution layer, and the batch normalization formula is: , wherein, , is the mean and variance of the current batch data, , are the learnable scaling and shifting parameters, is a numerical stability constant, and the value is ; These are the intermediate feature results after batch normalization; The input features before batch normalization; The final output feature after batch normalization; Global pooling layer: Global average pooling is used to map the sequence of side lengths into a feature vector of fixed length. The formula for global average pooling is: Where T is the sequence length, The features of the t-th time series; Feature output layer: The output is a high-dimensional feature vector of 128 or 256 dimensions, which represents an abstract representation of the fetal heart signal.

[0017] Preferably, the Long Short-Term Memory (LSTM) network includes: Feature embedding: Cosine positional encoding is used to compress long sequences. First, the original sequence is divided into segments of 100 data points each. The mean, standard deviation, maximum value, minimum value, and first-order difference mean of each segment are calculated. Then, positional encoding is added. The positional encoding formula is: The final output is a compressed sequence of length 48-96, where, For the position index of the segment, For the embedded dimension; Network architecture: A 2-3 layer bidirectional LSTM structure is adopted. The first N-1 layers return the complete sequence, and the last layer returns the final hidden state. The number of hidden units is 64-128. Dropout is added between each layer. The output of the bidirectional LSTM is the concatenation of the forward hidden state h_forward_t and the backward hidden state h_backward_t, that is, h_t=[h_forward_t;h_backward_t].

[0018] Preferably, the multimodal feature fusion module includes: Clinical Data Processing Tower ANN: Input clinical feature vectors with dimensions of 20-30 are processed through 3 fully connected layers, each containing 128 neurons with ReLU activation function. Batch normalization layer and Dropout are added after each layer, and the output is a 128-dimensional clinical feature vector F_Clinical. Feature fusion: The 128-dimensional clinical feature vector F_Clinical and the fetal heart rate feature vector F_FHR are fused. Feature standardization: Before fusion, F_FHR and F_Clinical are standardized separately, with the formula: F_FHR_norm=(F_FHR-mean(F_FHR)) / std(F_FHR); F_Clinical_norm=(F_Clinical-mean(F_Clinical)) / std(F_Clinical); Wherein, F_FHR_norm is the standardized fetal heart rate feature vector; F_Clinical_norm is the standardized clinical feature vector.

[0019] Preferably, the hierarchical classification prediction module includes: Hierarchical classification structure: Clinical diagnostic logic is represented by a directed acyclic graph (DAG) and is divided into three levels; Graph convolution operation: Information propagation is performed using a Graph Convolutional Network (GCN). The graph convolution formula is: ,in, For the first The node feature matrix of the layer; Let A be the adjacency matrix; It is the identity matrix; for The degree matrix; For the first The learnable weight matrix of the layer; The activation function is ReLU; For the first The node feature matrix of the layer; GNN network architecture: The input is a fused feature vector F_Fused, processed through two layers of GCN, each with a hidden dimension of 128, and Dropout is added after each layer; the output layer calculates the predicted probability of each node using the sigmoid function, with the formula as follows: , ,in, , The weights and biases of the output layer; Let i be the probability that the label corresponding to node i is true. Let i be the feature vector of node i; Among them, GCN layer 1: Hidden dimension = 128; GCN Layer 2: Hidden dimension = 128.

[0020] Preferably, the MC loss function ,in, For cross-entropy loss, and ,in, This is a real label; To predict probabilities; For MC-constrained loss; Weights to control the constraint loss; , , Let j1 and i1 be the predicted probabilities. Let be the set of edges in the DAG; This represents a directed edge from node i1 to node j1.

[0021] Preferably, the hierarchical classification prediction module adopts a multi-task learning strategy, and the total loss function Ltotal is: ,in, Cross-entropy loss for FGR diagnostic tasks; Cross-entropy loss for the FGR subtype classification task; Cross-entropy loss for perinatal outcome prediction tasks; , , , This represents the task weighting coefficient.

[0022] This invention provides a dual-network-based intelligent prediction method for fetal growth restriction, comprising: Step 1: Collect time-series data of prenatal fetal heart rate monitoring and structured clinical feature data. After quality control, noise reduction, outlier handling, missing value imputation, coding and standardization preprocessing, high-quality data is output. Step 2: A parallel architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) is adopted, wherein CNN extracts local pattern features of the high-quality data, and LSTM extracts long-term temporal dependency features of the high-quality data, and outputs fetal heart rate feature vector. Step 3: Output fused feature vector based on dual-tower architecture, wherein one tower processes structured clinical feature data through artificial neural network (ANN) and outputs clinical feature vector, and the other tower receives the fetal heart feature vector and fuses the two modal features in high-dimensional semantic space; Step 4: Use a graph neural network (GNN) to model the hierarchical structure of FGR clinical diagnosis, and combine it with the MC loss function to optimize the model, outputting the predicted probabilities of FGR diagnosis, subtype classification, and perinatal outcomes; Step 5: Based on the predicted probabilities, generate FGR risk scores, diagnostic conclusions, subtype classification results, perinatal outcome risk assessments, personalized clinical recommendations, and visualization reports.

[0023] Compared with the prior art, the beneficial effects of this application are as follows: Traditional methods rely on manually designed features (such as STV). This invention automatically learns features through deep learning, eliminating the need for manual feature engineering and uncovering deep patterns that traditional indicators struggle to capture. Existing research often focuses on single modalities (fetal heart rate data alone or clinical features alone). This invention effectively fuses two modalities through a dual-tower structure, fully utilizing complementary information. Traditional deep learning employs planar classification (ignoring hierarchical relationships) or multi-model cascading (error accumulation). This invention uses GNN to uniformly handle hierarchical classification within a single model, avoiding logical contradictions and error propagation. Traditional cross-entropy loss independently optimizes each classification task, ignoring hierarchical constraints. MC loss guides the model to learn a reasonable prediction distribution by explicitly modeling constraint relationships. Traditional methods either directly truncate long sequences (information loss) or attempt to process complete sequences (computational difficulties and training instability). This invention balances information preservation and computational efficiency through intelligent embedding.

[0024] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a dual-network-based intelligent prediction system for fetal growth restriction in an embodiment of the present invention. Figure 2 This is a schematic diagram of a one-dimensional convolution operation in an embodiment of the present invention; Figure 3 This is a structural diagram of the residual block in an embodiment of the present invention; Figure 4 This is a schematic diagram of the LSTM structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the double-tower structure in an embodiment of the present invention; Figure 6 This is a schematic diagram of the clinical data processing tower ANN in an embodiment of the present invention; Figure 7 This is a schematic diagram of the DAG hierarchical structure in an embodiment of the present invention; Figure 8 This is a complete system workflow diagram in an embodiment of the present invention; Figure 9This is a flowchart of a dual-network-based intelligent prediction method for fetal growth restriction in an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] This invention provides an intelligent prediction system for fetal growth restriction based on a dual-network architecture, such as... Figure 1 As shown, it includes: The data acquisition and preprocessing module is used to acquire time-series data of prenatal fetal heart rate monitoring and structured clinical feature data. After quality control, noise reduction, outlier handling, missing value imputation, coding and standardization preprocessing, high-quality data is output. The fetal heart signal feature extraction module is used to employ a parallel architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM). CNN extracts local pattern features from the high-quality data, and LSTM extracts long-term temporal dependency features from the high-quality data, outputting a fetal heart feature vector. The multimodal feature fusion module is used to output a fused feature vector based on a dual-tower architecture. One tower processes structured clinical feature data through an artificial neural network (ANN) and outputs a clinical feature vector, while the other tower receives the fetal heart feature vector and fuses the two modal features in a high-dimensional semantic space. The hierarchical classification prediction module is used to model the hierarchical structure of FGR clinical diagnosis using graph neural networks (GNNs), and optimize the model by combining the MC loss function to output the predicted probabilities of FGR diagnosis, subtype classification, and perinatal outcomes. The clinical decision support module is used to generate FGR risk scores, diagnostic conclusions, subtype classification results, perinatal outcome risk assessments, personalized clinical recommendations, and visualization reports based on the predicted probabilities.

[0029] Preferably, the data acquisition and preprocessing module includes: The fetal heart rate monitoring data acquisition unit is used to acquire 20-40 minute continuous time series data using a standard electronic fetal heart rate monitor at a sampling frequency of 4Hz, simultaneously acquiring monitoring graphs and raw fetal heart rate numerical sequences, and marking changes in body position and fetal movement events. The clinical feature data acquisition unit is used to collect basic maternal information, current pregnancy information, fetal ultrasound information, and perinatal outcome information. The basic maternal information includes age, parity, delivery, pre-pregnancy body mass index (BMI), and history of adverse pregnancy outcomes. The current pregnancy information includes gestational age at monitoring, gestational hypertension, and gestational diabetes. The fetal ultrasound information includes estimated fetal weight, abdominal circumference, amniotic fluid index, and umbilical artery blood flow parameters. The perinatal outcome information includes gestational age at delivery, delivery method, newborn birth weight, Apgar score, umbilical artery blood gas analysis results, and NICU admission status. The label definition unit is used to define the gold standard for diagnosing FGR when the estimated fetal weight or abdominal circumference is less than the 10th percentile of the corresponding gestational age and is affected by pathological factors. Early-onset FGR is diagnosed before 32 weeks of gestation, and late-onset FGR is diagnosed at 32 weeks of gestation and onwards. Adverse perinatal outcomes include preterm birth, neonatal acidosis, neonatal asphyxia, and hypoxic-ischemic encephalopathy. The data preprocessing unit is used to preprocess fetal heart rate data to obtain high-quality data; The dataset partitioning unit is used to stratify and randomly sample the high-quality data according to a preset ratio to divide it into training, validation, and test sets, and to train, validate, and test the Convolutional Neural Network (CNN) and the Long Short-Term Memory (LSTM) network.

[0030] Preferably, the network architecture of a convolutional neural network (CNN) includes: Input layer: Receives a preprocessed fetal heart rate sequence of shape (batch_size, sequence_length, 1), where sequence_length is the number of time steps and batch_size is the batch size; Feature extraction layer: consists of 6-8 concatenated residual blocks, with a kernel size of 3-7, a filter number of 64-256 increasing layer by layer, a stride of 1, and the activation function is ReLU; Dilated convolutional layers: Dilated convolutions with a dilation rate sequence of [1,2,4,8,16] are introduced; a batch normalization layer is added after each convolutional layer, and the batch normalization formula is: ,in, , The mean and variance of the current batch of data. , For learnable scaling and translation parameters, Here is the numerical stability constant, which takes the value of ; These are the intermediate feature results after batch normalization; The input features before batch normalization; The final output feature after batch normalization; Global pooling layer: Global average pooling is used to map the sequence of side lengths into a feature vector of fixed length. The formula for global average pooling is: Where T is the sequence length, The features of the t-th time series; Feature output layer: The output is a high-dimensional feature vector of 128 or 256 dimensions, which represents an abstract representation of the fetal heart signal.

[0031] Preferably, the Long Short-Term Memory (LSTM) network includes: Feature embedding: Cosine positional encoding is used to compress long sequences. First, the original sequence is divided into segments of 100 data points each. The mean, standard deviation, maximum value, minimum value, and first-order difference mean of each segment are calculated. Then, positional encoding is added. The positional encoding formula is: The final output is a compressed sequence of length 48-96, where, For the position index of the segment, For the embedded dimension; Network architecture: A 2-3 layer bidirectional LSTM structure is adopted. The first N-1 layers return the complete sequence, and the last layer returns the final hidden state. The number of hidden units is 64-128. Dropout is added between each layer. The output of the bidirectional LSTM is the concatenation of the forward hidden state h_forward_t and the backward hidden state h_backward_t, that is, h_t=[h_forward_t;h_backward_t].

[0032] Preferably, the multimodal feature fusion module includes: Clinical Data Processing Tower ANN: Input clinical feature vectors with dimensions of 20-30 are processed through 3 fully connected layers, each containing 128 neurons with ReLU activation function. Batch normalization layer and Dropout are added after each layer, and the output is a 128-dimensional clinical feature vector F_Clinical. Feature fusion: The 128-dimensional clinical feature vector F_Clinical and the fetal heart rate feature vector F_FHR are fused. Feature standardization: Before fusion, F_FHR and F_Clinical are standardized separately, with the formula: F_FHR_norm=(F_FHR-mean(F_FHR)) / std(F_FHR); F_Clinical_norm=(F_Clinical-mean(F_Clinical)) / std(F_Clinical); Wherein, F_FHR_norm is the standardized fetal heart rate feature vector; F_Clinical_norm is the standardized clinical feature vector.

[0033] Preferably, the hierarchical classification prediction module includes: Hierarchical classification structure: Clinical diagnostic logic is represented by a directed acyclic graph (DAG) and is divided into three levels; Graph convolution operation: Information propagation is performed using a Graph Convolutional Network (GCN). The graph convolution formula is: ,in, For the first The node feature matrix of the layer; Let A be the adjacency matrix; It is the identity matrix; for The degree matrix; For the first The learnable weight matrix of the layer; The activation function is ReLU; For the first The node feature matrix of the layer; GNN network architecture: The input is a fused feature vector F_Fused, processed through two layers of GCN, each with a hidden dimension of 128, and Dropout is added after each layer; the output layer calculates the predicted probability of each node using the sigmoid function, with the formula as follows: , ,in, , The weights and biases of the output layer; Let i be the probability that the label corresponding to node i is true. Let i be the feature vector of node i; Among them, GCN layer 1: Hidden dimension = 128; GCN Layer 2: Hidden dimension = 128.

[0034] Preferably, the MC loss function ,in, For cross-entropy loss, and ,in, This is a real label; To predict probabilities; For MC-constrained loss; Weights to control the constraint loss; , , Let j1 and i1 be the predicted probabilities. Let be the set of edges in the DAG; This represents a directed edge from node i1 to node j1.

[0035] Preferably, the hierarchical classification prediction module adopts a multi-task learning strategy, and the total loss function Ltotal is: ,in, Cross-entropy loss for FGR diagnostic tasks; Cross-entropy loss for the FGR subtype classification task; Cross-entropy loss for perinatal outcome prediction tasks; , , , This represents the task weighting coefficient.

[0036] In this embodiment, based on the cross-validation results of 176 training set samples, this coefficient combination resulted in the highest overall F1 score (0.89) for the three tasks, compared to other combinations (such as...). =0.5, =0.2), subtype classification accuracy improved by 4%, and outcome prediction recall improved by 3%.

[0037] In this embodiment, the Convolutional Neural Network (CNN) possesses translation invariance and local receptive field characteristics, making it suitable for extracting local pattern features of fetal heart rate signals (such as baseline variation, acceleration, deceleration, etc.). This invention employs a one-dimensional convolutional network to process time-series fetal heart rate data, such as... Figure 2 As shown.

[0038] In this embodiment, the feature extraction layer adopts a ResNet structure, which consists of multiple cascaded residual blocks, such as... Figure 3 As shown.

[0039] In this embodiment, the mathematical expression for dilated convolution is: Where d is the dilation rate, k is the kernel index, and a and b are the channel indices.

[0040] In this embodiment, the core formula for the one-dimensional convolution operation of the feature output layer is: ,in, To output the i2-th position and a-th channel of the feature map, The weights are the convolution kernel weights, connecting input channel b to output channel a. Let K be the (i2+k)th position and the bth channel of the input sequence, where K is the kernel size and C is the number of input channels; This is a bias term.

[0041] In this embodiment, the CNN model training strategy is as follows: Optimizer: Adam, learning rate = 0.001, β1 = 0.9, β2 = 0.999; Loss function: cross-entropy loss (for classification tasks); Regularization: Dropout (rate = 0.3), L2 regularization (λ = 0.0001); Number of training epochs: 100-200, using an early stopping strategy (Early Stopping, patience = 20); Batch size: 16-32.

[0042] In this embodiment, such as Figure 4 The diagram shown is a schematic of the LSTM structure.

[0043] In this embodiment, in the LSTM layer: Input (embedded sequence) → Bi-LSTM layer 1 (hidden units=128, return_sequences=True); → Dropout(rate=0.3); → Bi-LSTM layer 2 (hidden units=128, return_sequences=True); → Dropout(rate=0.3); → Bi-LSTM layer 3 (hidden units=64, return_sequences=False); → Output (Eigenvector).

[0044] In this embodiment, the LSTM model training strategy is as follows: Optimizer: Adam, learning rate = 0.001; loss function: cross-entropy loss; regularization: Dropout, gradient clipping (threshold = 1.0); number of training epochs: 100-200, early stopping strategy (patience = 20); batch size: 16-32.

[0045] In this embodiment, the features extracted by CNN and LSTM are complementary: CNN is good at capturing local patterns (such as sudden acceleration and deceleration of fetal heart rate) and is sensitive to short-term features; LSTM is good at modeling long-term temporal dependencies (such as long-term trends and periodic changes in fetal heart rate) and is sensitive to temporal dynamics.

[0046] In this embodiment, the expected AUC of the CNN model on the FGR classification task is ≥0.85; the expected AUC of the LSTM model on the FGR classification task is ≥0.85; and the expected AUC of the fusion model is ≥0.90 (a significant improvement compared to the AUC of approximately 0.65-0.70 of the traditional STV method).

[0047] In this embodiment, the design principle of the multimodal feature fusion module is as follows: the occurrence and prognosis of FGR are not only related to fetal heart rate signals, but also influenced by various factors such as maternal factors (age, BMI, complications) and fetal factors (gestational age, ultrasound parameters). Single-modal data is insufficient to comprehensively characterize the pathological state. Therefore, this module adopts a two-tower architecture to achieve multimodal feature fusion.

[0048] In this embodiment, such as Figure 5 The diagram shown is a schematic of a twin-tower structure.

[0049] In this embodiment, such as Figure 6 The diagram shown is a schematic of the Clinical Data Processing Tower (ANN).

[0050] In this embodiment, the present invention explored three fusion strategies and selected the optimal solution through experimental comparison: (1) Concatenation; F_Fused=Concat([F_FHR,F_Clinical]); Output dimension = dim(F_FHR) + dim(F_Clinical) = 256 + 128 = 384; Advantages: Preserves complete information from both modalities. Disadvantages: Higher dimensionality, which may increase the risk of overfitting. (2) Element-wise multiplication; Prerequisite: F_FHR and F_Clinical dimensions are the same (e.g., both are 128); F_Fused = F_FHR⊙F_Clinical; Output dimension = 128; Advantages: Captures the interaction between two modes. Disadvantages: Requires the output dimensions of the two towers to be consistent. (3) Weighted Fusion; F_Fused=W_FHR·F_FHR+W_Clinical·F_Clinical+b; Where W_FHR and W_Clinical are learnable weight matrices, and b is the bias; This approach is achieved through an additional fully connected layer, similar to an attention mechanism; Advantages: Adaptive learning of the importance of different modalities; Disadvantages: Introduces additional parameters. Recommended solution: splicing and merging, as it preserves complete information and is simple to implement.

[0051] In this embodiment, dimension mapping is performed by compressing the dimensions after concatenation using a single fully connected layer (128 neurons, ReLU activation), as shown in the formula: .

[0052] In this embodiment, the DAG hierarchical structure edge constraints of the GNN are described in text. Root node (Level0): The "Normal" node has no connection to Level1, and the "FGR" node has a directed connection to the "Early-onset FGR" and "Late-onset FGR" nodes of Level1. Subtype nodes (Level 1): The "early-onset FGR" node is connected to the "premature birth" and "hypoxic-ischemic encephalopathy" nodes in Level 2; the "late-onset FGR" node is connected to the "neonatal acidosis" and "neonatal asphyxia" nodes in Level 2. Ending Node (Level 2): ​​There are no connections between all ending nodes. They only receive information from the corresponding upper-level node and have no reverse connections (to avoid logical loops).

[0053] In this embodiment, the adjustment range of ANN model parameters is as follows: Dropoutrate: Adjust within the range of [0.2, 0.4], with a recommended value of 0.3 (when rate=0.2, the risk of model overfitting increases, and the validation set AUC decreases by 1.8%; when rate=0.4, too much feature information is lost, and the AUC decreases by 2.5%). Number of neurons in the fully connected layer: can be adjusted in the range of [100, 150], with a recommended value of 128 (feature extraction is insufficient when the number of neurons is <100, and the computational cost increases by 30% when the number of neurons is >150, with no performance improvement).

[0054] In this embodiment, multimodal fusion improves predictive performance by 5-10% (AUC increases from 0.85-0.88 to 0.90-0.92) compared to single fetal heart rate data, and the fusion model performs particularly well in high-risk cases with abnormal clinical characteristics.

[0055] In this embodiment, such as Figure 7 The diagram shown is a schematic of the DAG hierarchical structure.

[0056] In this embodiment, the MC loss function ensures that the probability distribution learned by the model conforms to hierarchical logic, avoiding contradictory prediction results. Experiments show that, compared to the standard cross-entropy loss, the MC loss can reduce the logical contradiction rate from 15-20% to <5%.

[0057] In this embodiment, the weight of the control constraint loss is typically set to 0.1-1.0.

[0058] In this embodiment, the hierarchical classification model achieved an AUC ≥ 0.90 for FGR diagnosis; an accuracy of ≥ 85% for early / late onset classification; an AUC ≥ 0.80 for perinatal outcome prediction (significantly improved compared to the AUC of approximately 0.60-0.65 of traditional methods); and a logical contradiction rate < 5%.

[0059] In this embodiment, the output format of the FGR diagnostic results is as follows: -FGR Risk Score: 0-100 points (based on predicted probability transformation); - Diagnostic conclusion: Normal / Suspected FGR / High risk of FGR; - Confidence level: The confidence interval of the model prediction.

[0060] FGR subtype classification: If diagnosed with FGR, further output: -Subtype: Early Hairstyle FGR / Late Hairstyle FGR; - Subtype probability: P(early onset) and P(late onset); Perinatal outcome prediction: For cases of febrile grouting (FGR), predicting the risk of various adverse outcomes: - Risk of premature birth: XX% (low risk / medium risk / high risk); - Risk of neonatal acidosis: XX%; - Risk of neonatal asphyxia: XX% - Risk of hypoxic-ischemic encephalopathy: XX% Clinical recommendations: Generate personalized monitoring recommendations based on prediction results. -Low risk: Routine prenatal check-ups, follow-up check-ups every 2-4 weeks; -Medium risk: Strengthen monitoring, have a follow-up examination in 1-2 weeks, and be alert to changes in fetal movement; -High risk: Close monitoring, hospitalization for observation if necessary, and consideration of the timing of termination of pregnancy; The rules for generating clinical recommendations are shown in Table 2: Table 2. Rules for Generating Clinical Recommendations In this embodiment, the visualization report includes: fetal heart rate curve visualization: plotting a fetal heart rate time series curve. -X-axis: Time (minutes); -Y-axis: Fetal heart rate (bpm); - Mark abnormal areas (deceleration, acceleration, reduced variability); - Highlight the key moments the model focuses on (visualized through attention weights); Risk radar chart; Multidimensional risk assessment radar chart: - Dimensions: FGR risk, preterm birth risk, acidosis risk, asphyxia risk, brain injury risk; - Visually displays the various risk levels; Historical trend chart: If the patient has multiple monitoring records: - Plot the time-varying curve of the FGR risk score; - Track the evolution of the disease; - Evaluate the effectiveness of intervention measures; System deployment options: Deploy as a desktop application or a web application. -Users upload fetal heart rate monitoring data and clinical information; -The system calculates and returns the prediction results in real time; -Suitable for offline analysis and retrospective studies.

[0061] In this embodiment, the prediction report generation time is less than 10 seconds (including data preprocessing and model inference); the visual report is clear and intuitive, making it easy for clinicians to understand quickly; the system is compatible with multiple deployment scenarios and can flexibly adapt to the needs of different medical institutions.

[0062] like Figure 8 The diagram shown is a complete workflow diagram of the system.

[0063] In this embodiment, stratified random sampling is used to divide the training set, validation set, and test set in an 8:1:1 ratio. For example, with a total sample size N=220, the training set consists of 176 samples (80%), the validation set consists of 22 samples (10%), and the test set consists of 22 samples (10%). Standardized data collection and rigorous preprocessing ensure the quality and consistency of the training data, laying the foundation for subsequent model training. The signal loss rate is controlled below 50%, outlier handling reduces noise interference, and normalization accelerates model convergence.

[0064] In this embodiment, the preprocessing of fetal heart rate data includes: removing low-quality samples with a signal loss rate >50%; using linear interpolation to repair short-term loss (<10 seconds) and cubic spline interpolation to repair long-term loss (≥10 seconds); removing outliers that exceed the physiological range (<60 or >200 bpm) and replacing them with the median of the five nearest data points. Outlier identification and handling: Absolute abnormality: fetal heart rate <60 bpm or >200 bpm, replaced by the median of the five nearest data points; Abnormal mutation: A difference of >30 bpm between two consecutive data points is considered a mutation and should be repaired using linear interpolation (e.g., ...). and The difference is normal. Mutation, then ; Abnormal consecutive identical values: 8 consecutive data points with the same value (outside the physiological fluctuation range) are replaced with the mean of the first 4 and the last 4 data points.

[0065] Missing value handling: Short-term missing data (<10 seconds): Linear interpolation repair, formula xi=xi k+(i (i k))×(xi+k xi k) / (2k) (k is the distance between valid data points before and after the missing segment); ,in, This represents the distance between valid data points before and after the missing segment.

[0066] Long-term missing data (≥10 seconds and ≤30 seconds): cubic spline interpolation is used for repair to ensure data smoothness; Missing data for more than 30 seconds: This is considered low-quality data, triggering the data acquisition unit to re-acquire or remove the sample.

[0067] In this embodiment, wavelet transform is used to remove high-frequency noise while retaining useful information such as fetal heart rate variability. Z-score standardization is used to make the data mean 0 and the standard deviation 1. Clinical feature preprocessing includes: filling missing values ​​for continuous variables with the median, filling missing values ​​for categorical variables with the mode or adding an "unknown" category; ordinal encoding is used for ordered categorical variables, and one-hot encoding is used for unordered categorical variables; continuous features are normalized to [0,1]. Wavelet transform is used to remove high-frequency noise while retaining useful information such as fetal heart rate variability. Wavelet decomposition is performed on the signal S (using Daubechies wavelet, decomposition level = 5), and soft thresholding is performed on the high-frequency coefficients, with a threshold of = σ is the noise standard deviation estimate, wavelet reconstruction yields the denoised signal S_denoised, and N is the length of a single fetal heart rate sequence. ,in, These are the first-level approximation coefficients.

[0068] In this embodiment, the data acquisition device is a standard electronic fetal heart rate monitor (such as the Philips Avalon series or the GE Corometrics series). Sampling parameters: Fetal heart rate sampling frequency: 4 Hz (4 data points per second); Monitoring duration: 20-40 minutes (routine clinical monitoring duration); Data format: continuous time series, each sample contains 4800-9600 data points; Data acquisition: Simultaneously collect fetal heart rate monitoring graphs (for manual interpretation reference) and raw fetal heart rate numerical sequences (for model training); Quality control: Signal loss rate <50% (samples with a loss rate exceeding 50% are removed); events such as changes in body position and fetal movement recorded during monitoring are annotated; Clinical characteristic data collection: Establish a standardized clinical characteristic database and collect the following information: Basic maternal information: age, parity, number of deliveries, pre-pregnancy body mass index (BMI), and history of adverse pregnancy and childbirth. Information for this pregnancy: gestational age at the time of data collection and monitoring, gestational hypertension, gestational diabetes, and other complications; Fetal information: Ultrasound estimation of fetal weight, abdominal circumference, amniotic fluid index, umbilical artery blood flow parameters, etc.; Perinatal outcome information: gestational age at delivery, mode of delivery, newborn birth weight, length, sex, Apgar score, umbilical artery blood gas analysis (pH, BE), NICU admission status, diagnosis of complications, and 6-month neurodevelopmental follow-up results; Label Definitions: Gold Standard for FGR Diagnosis: Fetal weight or abdominal circumference estimated by ultrasound < 10th percentile for corresponding gestational age, and affected by pathological factors (excluding constitutionally small fetuses); Early-onset FGR: Diagnosed before 32 weeks of gestation; Late-onset FGR: Diagnosed at 32 weeks of gestation and later; Adverse Perinatal Outcomes: Premature birth: delivery before 37 weeks of gestation; Neonatal acidosis: umbilical artery blood pH <7.1; Neonatal asphyxia: Apgar score <7 at 1 minute or 5 minutes; Hypoxic-ischemic encephalopathy: based on clinical and imaging diagnostic criteria.

[0069] This invention provides an intelligent prediction method for fetal growth restriction based on a dual-network architecture, such as... Figure 9 As shown, it includes: Step 1: Collect time-series data of prenatal fetal heart rate monitoring and structured clinical feature data. After quality control, noise reduction, outlier handling, missing value imputation, coding and standardization preprocessing, high-quality data is output. Step 2: A parallel architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) is adopted, wherein CNN extracts local pattern features of the high-quality data, and LSTM extracts long-term temporal dependency features of the high-quality data, and outputs fetal heart rate feature vector. Step 3: Output fused feature vector based on dual-tower architecture, wherein one tower processes structured clinical feature data through artificial neural network (ANN) and outputs clinical feature vector, and the other tower receives the fetal heart feature vector and fuses the two modal features in high-dimensional semantic space; Step 4: Use a graph neural network (GNN) to model the hierarchical structure of FGR clinical diagnosis, and combine it with the MC loss function to optimize the model, outputting the predicted probabilities of FGR diagnosis, subtype classification, and perinatal outcomes; Step 5: Based on the predicted probabilities, generate FGR risk scores, diagnostic conclusions, subtype classification results, perinatal outcome risk assessments, personalized clinical recommendations, and visualization reports.

[0070] The beneficial effects of the above technical solution are as follows: Traditional methods rely on manually designed features (such as STV), while this invention automatically learns features through deep learning, eliminating the need for manual feature engineering and discovering deep patterns that are difficult to capture with traditional indicators. Existing research mostly focuses on a single modality (fetal heart rate data only or clinical features only), while this invention effectively integrates two modalities through a dual-tower structure, making full use of complementary information. Traditional deep learning uses planar classification (ignoring hierarchical relationships) or multi-model cascading (error accumulation), while this invention uses GNN to uniformly handle hierarchical classification in a single model, avoiding logical contradictions and error propagation. Traditional cross-entropy loss independently optimizes each classification task, ignoring hierarchical constraints. MC loss guides the model to learn a reasonable prediction distribution by explicitly modeling constraint relationships. Traditional methods either directly truncate long sequences (information loss) or attempt to process complete sequences (computational difficulties and unstable training). This invention balances information preservation and computational efficiency through intelligent embedding.

[0071] This invention provides an intelligent prediction system for fetal growth restriction based on a dual-network architecture. The original LSTM feature extraction uses a fixed segmentation of 100 data points, failing to consider the temporal heterogeneity of fetal heart rate signals caused by factors such as gestational age, fetal movement frequency, and maternal pregnancy complications (e.g., short fluctuation cycles in early pregnancy and long fluctuation cycles in late pregnancy; low fetal heart rate variability in gestational hypertension patients; overly coarse segmentation easily loses key information, while overly fine segmentation easily introduces noise). Furthermore, traditional cosine position coding only reflects the sequential relationship of segments, without considering the physiological characteristics of the fetal heart rate signal itself (e.g., fetal heart rate variability), resulting in insufficient targeting and accuracy for long-term time-dependent feature extraction. Therefore, based on the original LSTM feature extraction module, the feature embedding and network gating mechanisms are optimized to achieve dynamic time-scale adaptation and feature enhancement, specifically including: Based on the gestational age (G) data collected during monitoring, a dynamic segment length was designed. The formula is as follows: ,in, The baseline value for the basic segment length; This is an adaptive adjustment term based on gestational age, fitting the physiological fluctuation cycle of fetal heart rate signals at different gestational ages. In early pregnancy, when G < 28, the segment length increases with gestational age, while in late pregnancy, when G ≥ 28, the rate of increase slows down, consistent with clinical fetal heart rate signal patterns. This is the variability weighting coefficient, and the optimal value verified by clinical data is 15. is the coefficient of variation of the preprocessed fetal heart rate signal, and ,in, The standard deviation of fetal heart rate. This is the mean fetal heart rate, used to dynamically adjust the segment length based on the signal's inherent variability. This is the floor function; Based on traditional cosine position coding, attention weights for the variability of fetal heart segments are introduced. The contribution of the location coding is dynamically adjusted, using the following formula: ; ; ; in, Let be the coefficient of variation of the fetal heart rate signal in segment pos, and , This refers to the fetal heart rate data in segment pos; The value of is between 0 and 1, and it is adaptively adjusted by the sigmoid function, where k is the adjustment coefficient with a value of 8. The threshold value is based on clinical experience and is set to 0.08.

[0072] Introducing [a specific feature] into the forget gate, input gate, and output gate of LSTM. The feedback term dynamically adjusts the gating threshold to enhance the ability to capture abnormal timing patterns. The formula is as follows: ; ; ; in, , , These are the outputs of the forget gate, input gate, and output gate, respectively. , , This is the gate weight matrix; , , For bias terms; This is a penalty term for insufficient variability; , , The penalty coefficients are set to 0.8, 1.2, and 1.0 respectively. Through experimental optimization, when the variability is insufficient, the forget gate is less likely to forget historical information, the input gate is more likely to receive new features, and the output gate is more likely to activate key signals. It is the sigmoid activation function.

[0073] In this embodiment, the original multimodal fusion scheme (sembrittlement, element-wise multiplication, fixed weighting) adopts a static fusion strategy, which does not consider the dynamic correlation differences between fetal heart rate features (F_FHR) and clinical features (F_Clinical) in different samples. For example, the contribution of the clinical feature of "blood glucose control" to FGR prediction is much higher in gestational diabetes patients than in ordinary pregnant women; and when the fetal heart rate signal shows a "continuous deceleration" pattern, the weight of fetal heart rate features should be significantly higher than that of clinical features. Static fusion cannot adapt to the dynamic changes in this intermodal correlation in real time, resulting in insufficient discriminative power of the fused features and affecting prediction accuracy. Based on the original dual-tower structure, dynamic modal correlation calculation and adaptive attention weight adjustment are introduced to achieve dynamic optimization fusion of multimodal features, specifically including: Based on the standardized fetal heart rate feature vector F_FHR_norm and clinical feature vector F_Clinical_norm, the dynamic Pearson correlation coefficient matrix between the two is calculated. The surface is then smoothed using a Gaussian kernel function, as shown in the following formula: ; ; in, is the original Pearson correlation coefficient between the i-th feature of F_FHR_norm and the j-th feature of F_Clinical_norm; N is the number of samples in the current batch; Let F_FHR_norm be the batch mean of the i3rd feature; Let f(j) be the batch mean of the j3rd feature of F_Clinical_norm; The smoothed dynamic correlation coefficient; The mean of the original correlation coefficient matrix R; The smoothing coefficient is set to balance the dynamics and stability of the correlation, and its value is 0.3. j ,and The dimension of F_FHR_norm; The dimension of F_Clinical_norm; Based on dynamic correlation coefficient matrix Construct modal attention weights and And introduce a clinical risk level correction item, specifically: To weight fetal heart rate characteristics, Weights are assigned to clinical features; ; ,and ; in, The average absolute value of the dynamic correlation coefficient matrix reflects the overall correlation strength between the two modes; The clinical risk level is quantified based on clinical characteristics: 0 for no pregnancy complications, 0.3 for a single complication, and 0.7 for multiple complications. This is a risk correction factor, with a value of 0.5, which increases the weight of clinical features in high-risk samples. By combining attention weights with standardized feature vectors and introducing a modal interaction calibration term, a fused feature vector is finally obtained. The formula is as follows: ; in, This is the element-wise multiplication operator; The interaction calibration coefficient, which balances the contributions of linear fusion and interaction fusion, is set to 0.2.

[0074] The beneficial effects of the above technical solution are: for fetal heart data of different gestational weeks and different signal quality, it can accurately extract local and long-term temporal features, adapt to the correlation differences between modalities in different samples in real time, strengthen the contribution of key modalities, better fit the clinical pathological logic, improve interpretability, and provide more accurate predictive support for complex cases (such as those with pregnancy complications or low-variability fetal heart signals).

[0075] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dual-network based intelligent prediction system for fetal growth restriction, characterized in that, The method comprises the following steps: A data acquisition and preprocessing module is used to collect prenatal fetal heart monitoring time series data and structured clinical feature data, and after quality control, denoising, outlier processing, missing value filling, encoding and standardization preprocessing, high-quality data is outputted; A fetal heart signal feature extraction module is used to adopt a convolutional neural network (CNN) and a long short-term memory network (LSTM) parallel architecture, wherein the CNN extracts local pattern features of the high-quality data, and the LSTM extracts long-range time series dependence features of the high-quality data, and a fetal heart feature vector is outputted; A multi-modal feature fusion module is used to output a fusion feature vector based on a double-tower structure architecture, wherein one tower processes structured clinical feature data through an artificial neural network (ANN) and outputs a clinical feature vector, and the other tower receives the fetal heart feature vector and fuses the two modal features in a high-dimensional semantic space; A hierarchical classification prediction module is used to model the hierarchical structure of FGR clinical diagnosis by adopting a graph neural network (GNN), combine an MC loss function to optimize the model, and output prediction probabilities of FGR diagnosis, subtype classification and perinatal outcome; A clinical decision support module is used to generate FGR risk score, diagnosis conclusion, subtype classification result, perinatal outcome risk assessment, individualized clinical advice and visual report according to the prediction probabilities.

2. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, The data acquisition and preprocessing module comprises: A fetal heart monitoring data acquisition unit is used to collect 20-40 minutes of continuous time series data at a sampling frequency of 4Hz by using a standard electronic fetal heart monitor, synchronously collect monitoring graphs and original fetal heart rate value sequences, and label body position changes and fetal movement events; A clinical feature data acquisition unit is used to collect maternal basic information, this pregnancy information, fetal ultrasound information and perinatal outcome information, wherein the maternal basic information includes age, pregnancy times, parity, pre-pregnancy body mass index (BMI), and previous adverse pregnancy history, the this pregnancy information includes monitoring time gestational age, gestational hypertension, gestational diabetes, the fetal ultrasound information includes ultrasound estimated fetal weight, abdominal circumference, amniotic fluid index, and umbilical artery blood flow index, and the perinatal outcome information includes delivery gestational age, delivery mode, neonatal birth weight, Apgar score, umbilical artery blood gas analysis result, and NICU admission situation; A label definition unit is used to take ultrasound estimated fetal weight or abdominal circumference < corresponding gestational age 10th percentile and pathologic factor influence as the FGR diagnosis gold standard, diagnose early-onset FGR before 32 weeks of gestation, diagnose late-onset FGR at and after 32 weeks of gestation, and take adverse perinatal outcomes including premature birth, neonatal acidosis, neonatal asphyxia and hypoxic-ischemic encephalopathy; A data preprocessing unit is used to preprocess fetal heart data to obtain high-quality data; A data set division unit is used to stratified random sample divide training set, validation set and test set according to a preset proportion, and train, validate and test the convolutional neural network (CNN) and the long short-term memory network (LSTM).

3. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, The network architecture of the convolutional neural network (CNN) comprises: The input layer receives a preprocessed fetal heart sequence in the form of (batch_size, sequence_length, 1), where sequence_length is the number of time steps, and batch_size is the batch size. The feature extraction layer is composed of 6-8 residual blocks, with a convolution kernel size of 3-7, a filter number of 64-256 increasing layer by layer, a step size of 1, and a ReLU activation function. Atrous convolution layer: atrous convolution with dilation rate sequence [1, 2, 4, 8, 16] is introduced; a batch normalization layer is added after each convolution layer, and the batch normalization formula is: wherein, , is the mean and variance of the current batch data, , is the learnable scaling and translation parameters, is a numerical stability constant, and the value is ; is the intermediate feature result after batch normalization; is the input feature before the batch normalization operation; is the final output feature after the batch normalization operation; Global pooling layer: the global average pooling is used to map the edge length sequence into a fixed length feature vector, where the global average pooling formula is: where T is the sequence length, is the feature of the t-th time sequence. The feature output layer outputs a high-dimensional feature vector of 128 or 256 dimensions, representing an abstract representation of the fetal heart signal.

4. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, The long short-term memory network (LSTM) includes: Feature embedding: long sequences are compressed using cosine positional encoding. First, the original sequence is segmented by every 100 data points, and the mean, standard deviation, maximum, minimum, and first-order difference mean statistical features of each segment are calculated. Then, the position encoding is added. The position encoding formula is: The final output length is 48-96 compressed sequences, where, is the position index of the segment, is the embedding dimension. The network architecture adopts a 2-3 layer bidirectional LSTM structure, with the first N-1 layers returning the complete sequence and the last layer returning the final hidden state. The number of hidden units is 64-128, and Dropout is added between layers. The bidirectional LSTM output is the concatenation of the forward hidden state h_forward_t and the backward hidden state h_backward_t, i.e., h_t = [h_forward_t; h_backward_t].

5. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, The multi-modal feature fusion module includes: The clinical data processing tower (ANN) inputs a 20-30 dimensional clinical feature vector, processes it through 3 fully connected layers, each containing 128 neurons, with a ReLU activation function, and adds a batch normalization layer and Dropout after each layer. It outputs a 128-dimensional clinical feature vector F_Clinical. Feature fusion: fuse the 128-dimensional clinical feature vector F_Clinical with the fetal heart feature vector F_FHR. Feature standardization: standardize F_FHR and F_Clinical before fusion, with the formula: F_FHR_norm = (F_FHR - mean(F_FHR)) / std(F_FHR) F_Clinical_norm = (F_Clinical - mean(F_Clinical)) / std(F_Clinical) 6. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, where F_FHR_norm is the standardized fetal heart feature vector, and F_Clinical_norm is the standardized clinical feature vector. The hierarchical classification prediction module includes: Graph convolution operation: Information propagation is performed using a Graph Convolutional Network (GCN). The graph convolution formula is: ,in, For the first The node feature matrix of the layer; Let A be the adjacency matrix; It is the identity matrix; for The degree matrix; For the first The learnable weight matrix of the layer; The activation function is ReLU; For the first The node feature matrix of the layer; The GNN network architecture: the input is the fusion feature vector F_Fused, which is processed by two layers of GCN, each layer has a hidden dimension of 128, and a Dropout is added after each layer; the output layer calculates the prediction probability of each node through the sigmoid function, and the formula is , , , is the weight and bias of the output layer; is the probability that the label of node i is true; is the feature vector of node i; wherein the GCN layer 1: hidden_dim = 128; GCN layer 2: , hidden_dim = 128.

7. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 1, wherein, the MC loss function wherein, is a cross-entropy loss, and wherein, is a true label; is a predicted probability; is an MC constraint loss; is a weight controlling the constraint loss; , , is a predicted probability of node j1, node i1; is a set of edges in a DAG; denotes a directed edge from node i1 to node j1.

8. The dual-network based intelligent prediction system for fetal growth restriction as claimed in claim 7, wherein, Hierarchical classification structure: represents the clinical diagnosis logic as a directed acyclic graph (DAG) and is divided into three levels. wherein, is the cross-entropy loss for the FGR diagnosis task; is the cross-entropy loss for the FGR subtype classification task; is the cross-entropy loss for the perinatal outcome prediction task; , , , is the task weight coefficient.

9. A dual-network based intelligent prediction method for fetal growth restriction, characterized in that, The hierarchical classification prediction module adopts a multi-task learning strategy, and the total loss function Ltotal is: Step 1: Collect prenatal fetal heart monitoring time series data and structured clinical feature data, and after quality control, denoising, outlier processing, missing value filling, encoding, and standardization preprocessing, output high-quality data. Step 2: Use a convolutional neural network (CNN) and a long short-term memory network (LSTM) parallel architecture, where CNN extracts local pattern features from the high-quality data, and LSTM extracts long-range temporal dependence features from the high-quality data, outputting a fetal heart feature vector. Step 3: output the fusion feature vector based on a double-tower structure architecture, wherein one tower processes the structured clinical feature data through an artificial neural network (ANN) and outputs a clinical feature vector, and the other tower receives the fetal heart feature vector and fuses the two modal features in a high-dimensional semantic space; Step 4: model the hierarchical structure of FGR clinical diagnosis by using a graph neural network (GNN), combine the MC loss function to optimize the model, and output the prediction probability of FGR diagnosis, subtype classification and perinatal outcome; Step 5: generate the FGR risk score, diagnosis conclusion, subtype classification result, perinatal outcome risk assessment, personalized clinical suggestion and visual report according to the prediction probability.

Images