Fetal hypoxia early warning method and device combining health condition and fetal heart rate

Abstract

The invention provides a fetal hypoxia early warning method and device combining health conditions and fetal heart rates, and is applied to the technical field of data processing. The method comprises the following steps: preprocessing a fetal heart rate signal of a target fetus to generate target fetal heart rate signal sequence information; the target fetal heart rate signal sequence information is processed on the basis of a Gaussian second wavelet function, a target wavelet transformation diagram is generated, and the target wavelet transformation diagram is used for representing time-frequency information of the fetal heart rate signal of the target fetus; processing the physiological parameter information of the pregnant woman and the PH value information of the target fetus based on a target fetal hypoxia diagnosis model to generate a target physiological state feature; processing the target wavelet transform diagram based on the target fetal hypoxia diagnosis model to generate target image features; and processing the target physiological status features and the target image features based on the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method and device for warning of fetal hypoxia by combining health status and fetal heart rate. Background Art

[0002] In the field of obstetrics and gynecology, fetal hypoxia is a key problem that seriously threatens the health of the fetus during pregnancy and childbirth. Although electronic fetal monitors are widely used to detect fetal hypoxia, which collect fetal heart rate signals and uterine contraction signals through ultrasonic Doppler probes, there are obvious defects in the visual inspection of fetal heart rate and uterine contraction monitoring charts based on clinical guidelines. Due to the internal differences among observers, a relatively high false positive rate is likely to occur, which may lead to unnecessary medical interventions, such as cesarean section, etc., increasing the medical risks and burdens on pregnant women and fetuses.

[0003] With the development of artificial intelligence technology, machine learning and deep learning methods have gradually been applied to the classification of fetal heart rate monitoring charts to assist obstetricians in diagnosis. In terms of machine learning, feature selection usually relies on professional domain knowledge, which limits the convenience and universality of its application. In the field of deep learning, there is generally a problem of insufficient generalization, with unstable performance in different data sets or actual clinical scenarios, and a lack of in-depth exploration of the robustness of the proposed methods, making it difficult to meet the urgent clinical need for accurate, stable, and reliable fetal hypoxia diagnosis technology.

[0004] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of the present disclosure, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present application is to provide a method and device for warning of fetal hypoxia by combining health status and fetal heart rate, which at least overcome the problems existing in the prior art to a certain extent. By obtaining the fetal heart rate signal, PH value information, pregnant woman's physiological parameter information, diagnostic model, and training sample set of the target fetus. Downsample, slice, and process outliers and missing values of the fetal heart rate signal, and then generate a target wavelet transform diagram through a series of operations of the second Gaussian wavelet function. During model training, determine the sampling ratio from the sample set, generate sampling features to form a data group for training, and verify with the validation set to determine the target model. In the feature processing link, generate different physiological state information based on the information of the pregnant woman and the fetus, extract features and map them to a low-dimensional space, obtain the target physiological state features through semantic analysis, and at the same time extract frequency change features from the target wavelet transform diagram to generate target image features. Finally, fuse the two features, process them through the model to obtain the fused feature sequence and target classification head information, generate fetal hypoxia information to assist clinical judgment of whether the fetus is hypoxic, and integrate multiple information to provide an effective way for diagnosis.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or be learned in part through the practice of the present invention.

[0007] According to one aspect of the present application, a fetal hypoxia warning method combining health status and fetal heart rate is provided, including: obtaining the fetal heart rate signal of the target fetus, the PH value information of the target fetus, the physiological parameter information of the pregnant woman, the fetal hypoxia diagnosis model based on VisionTransformer, and the training sample set; preprocessing the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information; processing the target fetal heart rate signal sequence information based on the second Gaussian wavelet function to generate a target wavelet transform diagram, wherein the target wavelet transform diagram is used to characterize the time-frequency information of the fetal heart rate signal of the target fetus; processing the fetal hypoxia diagnosis model based on VisionTransformer based on the training sample set to generate a target fetal hypoxia diagnosis model; processing the physiological parameter information of the pregnant woman and the PH value information of the target fetus based on the target fetal hypoxia diagnosis model to generate target physiological state features; processing the target wavelet transform diagram based on the target fetal hypoxia diagnosis model to generate target image features; processing the target physiological state features and the target image features based on the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

[0008] In another aspect of the present application, a fetal hypoxia warning device combining health status and fetal heart rate is characterized by including: an acquisition module for obtaining the fetal heart rate signal of the target fetus, the PH value information of the target fetus, the physiological parameter information of the pregnant woman, the fetal hypoxia diagnosis model based on VisionTransformer, and the training sample set; a processing module for preprocessing the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information; processing the target fetal heart rate signal sequence information based on the second Gaussian wavelet function to generate a target wavelet transform diagram, wherein the target wavelet transform diagram is used to characterize the time-frequency information of the fetal heart rate signal of the target fetus; processing the fetal hypoxia diagnosis model based on VisionTransformer based on the training sample set to generate a target fetal hypoxia diagnosis model; processing the physiological parameter information of the pregnant woman and the PH value information of the target fetus based on the target fetal hypoxia diagnosis model to generate target physiological state features; processing the target wavelet transform diagram based on the target fetal hypoxia diagnosis model to generate target image features; processing the target physiological state features and the target image features based on the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

[0009] According to another aspect of the present application, there is provided a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a second processor, the above-mentioned fetal hypoxia warning method combining health status and fetal heart rate is implemented.

[0010] For the fetal hypoxia warning method and device combining health status and fetal heart rate provided by the present application, the server acquires the fetal heart rate signal, pH value information, pregnant woman's physiological parameter information, diagnostic model and training sample set of the target fetus. The fetal heart rate signal is downsampled, sliced, and processed for outliers and missing values, and then a series of operations using the second Gaussian wavelet function are performed to generate a target wavelet transform diagram. During model training, the sampling ratio is determined by obtaining features from the sample set, and the sampled features are formed into a data group for training, and the target model is determined by validating with a validation set. In the feature processing section, different physiological state information is generated based on the information of the pregnant woman and the fetus, and features are extracted and mapped to a low-dimensional space. Through semantic analysis, the target physiological state features are obtained. At the same time, the frequency change features are extracted from the target wavelet transform diagram to generate the target image features. Finally, the two features are fused, and through model processing, the fused feature sequence and the target classification head information are obtained, and fetal hypoxia information is generated to assist clinical judgment of whether the fetus is hypoxic, integrating multiple information to provide an effective way for diagnosis.

[0011] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The flowchart showing a fetal hypoxia warning method combining health status and fetal heart rate provided by an embodiment of the present application;

[0013] Figure 2 The structural schematic diagram showing a fetal hypoxia warning device combining health status and fetal heart rate provided by an embodiment of the present application;

[0014] Figure 3 The structural schematic diagram showing the health status encoder of a fetal hypoxia diagnostic model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0015] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for illustrating and explaining the present invention, and are not used to limit the present invention.

[0016] The following combines Figure 1 to describe the fetal hypoxia warning method combining health status and fetal heart rate according to an exemplary embodiment of the present application. In one embodiment, the present application also proposes a fetal hypoxia warning method and device combining health status and fetal heart rate. Figure 1A flowchart schematically shows a fetal hypoxia warning method combining health status and fetal heart rate according to an embodiment of the present application. As Figure 1 shown, this method is applied to a server and includes:

[0017] S101, obtaining the fetal heart rate signal of the target fetus, the pH value information of the target fetus, the physiological parameter information of the pregnant woman, the fetal hypoxia diagnosis model based on VisionTransformer, and the training sample set.

[0018] In one embodiment, in a clinical environment, an ultrasonic Doppler probe is usually placed on the abdomen of a pregnant woman to collect the fetal heart rate signal. For example, during a prenatal examination in a hospital, the pregnant woman lies flat on the examination bed as required by the doctor. After the medical staff apply an appropriate amount of coupling agent to the ultrasonic Doppler probe, they gently place it at a suitable position on the abdomen of the pregnant woman to ensure that the fetal heart beat signal can be accurately received. The instrument records the fetal heart rate at regular time intervals (such as several times per second) to form an original fetal heart rate signal data sequence. These data may be affected by factors such as the physical movement of the pregnant woman, fetal activities, and electromagnetic interference in the surrounding environment, resulting in noise and fluctuations, but they constitute the basic data source for subsequent analysis. The pH value information of the fetus needs to be obtained through invasive operations such as amniocentesis or fetal scalp blood sampling. Taking amniocentesis as an example, under ultrasonic guidance, the doctor will insert a slender puncture needle through the abdominal wall and uterine wall of the pregnant woman and into the amniotic cavity to extract an amniotic fluid sample. By detecting and analyzing the amniotic fluid sample, the pH value information of the fetus can be obtained.

[0019] The physiological parameter information of the pregnant woman covers many aspects. Basic information such as the age, height, and weight of the pregnant woman can be obtained through interrogation and routine measurements. During each prenatal examination, the nurse will use a weighing scale to measure the weight of the pregnant woman and a height measuring instrument to measure the height, and record them in the pregnant woman's prenatal examination file. For the medical history of the pregnant woman, such as whether she has chronic diseases such as diabetes and hypertension, the doctor will carefully inquire about the pregnant woman's past medical records and conduct relevant blood tests and physical examinations to confirm. For example, by detecting the blood glucose level and blood pressure value of the pregnant woman to determine whether there is a risk of diabetes or hypertension. In addition, some special situations during pregnancy, such as whether there is premature rupture of membranes, umbilical cord around the neck, TPA, TP, thalassemia, etc., are mainly detected and recorded by ultrasound examination. These rich physiological parameter information of the pregnant woman are of great reference value for the assessment of the fetal health status.

[0020] The training sample set usually comes from a large amount of clinical case data of multiple medical centers or hospitals. For example, the research team collaborated with multiple obstetrics and gynecology hospitals to collect prenatal examination data of different pregnant women at different stages of pregnancy in the past few years, including fetal heart rate signal data, various physiological parameter data of pregnant women, and fetal pH value data, etc. These data are screened and sorted to ensure the quality and integrity of the data. At the same time, in order to enable the model to learn the health status of the fetus in different situations, the sample set will include relevant data of normal fetuses and fetuses suspected or diagnosed with hypoxia, and each sample is accurately labeled to indicate whether the fetus is hypoxic and the degree of hypoxia. In the data preprocessing stage, operations such as downsampling, data slicing, outlier detection and processing are performed on the collected fetal heart rate signals, and they are transformed into a format suitable for model input. The physiological parameter information of pregnant women and the fetal pH value information will also be subjected to corresponding standardization and encoding processing. For example, the disease status of pregnant women is represented by 0 or 1 (0 means no disease, 1 means having the corresponding disease), and the fetal pH value is classified and encoded according to a certain threshold range, etc.

[0021] The key innovation of the fetal hypoxia diagnosis model based on VisionTransformer lies in the way of processing image data, which is different from traditional convolutional neural networks. It divides the input data (such as the image converted from the fetal heart rate signal, the health status information of pregnant women and fetuses, etc.) into fixed-size patches, regards these patches as tokens in the sequence, and uses the self-attention mechanism to capture the relationships between the patches, so that it can learn the global features of the data, rather than being limited to local features only. This architecture design enables the model to have unique advantages when dealing with complex medical data, and can better mine the associations and potential information between different modality data.

[0022] The model mainly includes three important stages: the feature extraction stage, the fusion stage and the classification stage. In the feature extraction stage, there are a dedicated Health Encoder (HER) and a Signal Encoder (SER). The HER module processes the text information of the health status of pregnant women and fetuses using nine embedding layers, maps it to a continuous embedding space, and obtains the health status embedding representing the comprehensive status of the mother and fetus through element-wise addition, extracting the semantic information therein. The SER module focuses on learning the visual features of the wavelet transform map (converted from the fetal heart rate signal), and effectively extracts the time-frequency information in the fetal heart rate signal through operations such as Dynamic Snake Convolution (DSConv), feature map partitioning and linear projection. Among them, DSConv can adaptively focus on slender and tortuous local structures, better perceive the time-frequency changes, and provide key signal features for subsequent analysis.

[0023] In the fusion stage, the model cleverly performs an element-wise addition operation on the semantic information and [CLS] embeddings, then concatenates them with the image features and feeds them into the Transformer module. The Transformer module deeply analyzes and fuses these features through the multi-head self-attention mechanism and the feed-forward neural network, fully considering the relationships between different position blocks in the sequence, further capturing the global information and local details of the features, so that the fused features can more comprehensively reflect the actual situation of the fetus. Finally, in the classification stage, the model extracts the first block ([CLS] token) from the fused feature sequence, which incorporates all the key information processed previously, and then feeds it into the classification head. Based on the classification boundaries and decision rules learned during the model training process, the classification head makes a final judgment on whether the fetus is hypoxic, outputs the prediction result of fetal hypoxia, usually in the form of a classification category (hypoxic or normal) or a probability value, providing an important reference for clinical diagnosis.

[0024] The training process requires a large amount of labeled data, which is sourced from clinical cases of multiple medical centers or hospitals. The data covers detailed information of different pregnant women at different stages of pregnancy, including preprocessed fetal heart rate signals, various physiological parameters of the pregnant women (such as age, weight, whether suffering from diabetes, hypertension, etc.), and fetal pH value information, etc. After collecting the data, strict data cleaning and preprocessing work will be carried out to ensure the quality and consistency of the data, such as imputing missing values, dealing with outliers, etc., to make it meet the requirements of model input.

[0025] During training, data features are first obtained from the training sample set, and an appropriate sampling ratio is generated according to the number of data features. Preset numbers of sampled features are generated through sampling operations. Then these sampled features are combined with other data to form multiple groups of data sets. Each data set contains a certain number of data samples, and at least one data sample has identification information (used to mark whether it is a risk factor affecting fetal hypoxia, etc.). Next, these data sets are used to iteratively train the model. In each iteration, the model performs forward propagation calculations on the input data to predict the results, and calculates the error between the predicted results and the true labels through a loss function (such as focal loss). Subsequently, based on the backpropagation algorithm, the error information propagates backward from the classification head to each layer of the model, and the model adjusts its own parameters (such as the weight matrix and bias terms in each module, etc.) according to the error to continuously improve the accuracy and generalization ability of the model. During the training process, a validation sample set is also used to validate the trained model, and the performance of the model is evaluated according to the validation results, such as indicators like accuracy, recall rate, F1 score, AUC, etc. If the validation results show that the model has deficiencies in some aspects, such as overfitting or underfitting, the hyperparameters of the model (such as the learning rate, regularization parameter, etc.) or the training strategy will be adjusted, and training will continue until the performance of the model on the validation set reaches a satisfactory level, and finally a well-trained fetal hypoxia diagnosis model is obtained, which can accurately diagnose fetal hypoxia for new input data.

[0026] S102. Preprocess the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information.

[0027] In one implementation, downsampling and data slicing are performed on the fetal heart rate signal of the target fetus to generate a number of data segments. In the actual clinical scenario, when collecting the fetal heart rate signal of the target fetus, the professional fetal monitoring device used will continuously record the fetal heart rate data. Just like the 60-minute fetal heart rate signal shown as an example, its original sampling frequency is 10 Hz, which is to capture the changes in the fetal heart rate as detailed as possible, but at the same time it also brings a large amount of data. When performing the downsampling operation, it is reduced to 2 Hz, which not only reduces the storage and processing pressure of the data, but also helps to highlight the main trends and features of the signal. During the data slicing process, taking 10 minutes as a data segment is based on the research on the fetal heart rate change law and clinical experience. Usually within such a time period, the change of the fetal heart rate is relatively stable, but it can also reflect a certain degree of dynamics, which is convenient for subsequent analysis and processing. For example, within different 10-minute segments, the fetus may show different heart rate change patterns due to factors such as its own activity cycle and the change of the pregnant woman's physical state.

[0028] Outlier detection is performed on several data segments respectively to generate target data segments, where the target data segments are data segments with abnormal data. In the outlier detection process, the criterion of setting values greater than 180 bpm or less than 100 bpm as outliers is obtained through a large number of clinical data statistics and medical research. When those 5 outliers are found in the first data segment, these outliers imply that the fetus has undergone some physiological changes or been interfered by external factors at that moment. For example, the abnormal high value of the 23rd data may be due to the sudden violent fetal movement at that moment, resulting in an instantaneous increase in the measured heart rate; while the abnormality of the 112th data may be related to the slight body movement of the pregnant woman during that period, which affects the signal acquisition.

[0029] Process the target data segments to generate missing sample information, where the missing sample information is used to characterize the samples with missing values and the positions and variable names where the missing values are located. Process the missing sample information to generate imputed variable values for the missing values. For the case of missing values, such as the 456th data being missing, it may be due to a temporary signal interruption or interference of the acquisition device at that moment. The linear interpolation method is used for imputation. For example, if the fetal heart rate shows a relatively stable upward or downward trend in the time periods before and after this data segment, then the result obtained by the linear interpolation method is more reliable; but if the heart rate changes are more complex, it is necessary to further refer to the trends of adjacent data segments or combine other clinical indicators to comprehensively judge the rationality of the imputed value.

[0030] Based on the imputed variable values for the missing values, process the target data segments to generate target fetal heart rate signal sequence information, where the target fetal heart rate signal sequence information is composed of several data segments with complete numerical values. By calculating the statistical characteristics of the data segments, such as mean, standard deviation, range, etc., to evaluate the stability and reliability of the data. If the statistical characteristics of a certain data segment after processing are significantly different from those of other segments, it may be necessary to re-examine the processing process to ensure the accuracy of the data. Only the target fetal heart rate signal sequence information after strict processing and verification can provide a reliable basis for the subsequent processing based on the second Gaussian wavelet function and the input of the fetal hypoxia diagnosis model, thereby improving the accuracy and reliability of fetal hypoxia diagnosis.

[0031] S103, Process the target fetal heart rate signal sequence information based on the second Gaussian wavelet function to generate a target wavelet transform diagram.

[0032] In one implementation, the target fetal heart rate signal sequence information is normalized to generate a normalized segment signal. Suppose there is a target fetal heart rate signal sequence information S = {110, 120, 115, 130, 125, 140, 135, 150, 145, 160}. Each element is normalized according to the calculation formula of the normalized segment signal S′ = {s′ / max(S)|s′∈S}, where S′ represents the normalized segment signal, and S = {s1s2,..., s n} represents the target fetal heart rate signal sequence; S′ = [110 / 160, 120 / 160, 115 / 160, 130 / 160, 125 / 160, 140 / 160, 135 / 160, 150 / 160, 145 / 160, 160 / 160], which is the generated normalized segment signal.

[0033] Based on a preset decomposition formula, the discrete wavelet transform is performed on the normalized segment signal to generate approximation coefficients and detail coefficients. The Symlet8 wavelet packet is used for the discrete wavelet transform. The first decomposition formula for the approximation coefficients is: A j+1 (n) = ∑ k h k S j (2n - k); where A j+1 (n) represents the approximation coefficient, h k represents the low-pass filter at level j, and S j (2n - k) represents the signal value after the original signal s undergoes time-scale transformation and filter action during the j-level decomposition. The second decomposition formula for the detail coefficients is: D j+1 (n) = ∑ k g k S j (2n - k); where D j+1 (n) represents the detail coefficient, and g k represents the high-pass filter at level j.

[0034] Let j = 0. Suppose h k = {0.1, 0.2, 0.3}, and g k = {0.3, 0.2, 0.1}.

[0035] For the approximation coefficient A1(0): A1(0) = ∑ k＝0 h k S0(2×0 - k) = h0S0(0) + h1S0(-1) + h2S0(-2). Since S0(-1) and S0(-2) do not exist (starting position of the sequence), assuming the boundary treatment is 0, then A1(0) = h0S0(0) = 0.1×0.6875 = 0.06875.

[0036] For the approximation coefficient A1(1): A1(1) = ∑ k＝0 h k S0(2×1 - k) = h0S0(2) + h1S0(1) + h2S0(0) = 0.1×0.71875 + 0.2×0.75 + 0.3×0.6875 = 0.428125. Calculate other approximation coefficients by analogy.

[0037] For the detail coefficient D1(0): D1(0) = ∑ k g k S0(2×0 - k) = g0S0(0) + g1S0(-1) + g2S0(-2), and the boundary is also processed as 0. D1(0) = g0S0(0) = 0.3×0.6875 = 0.20625.

[0038] For the detail coefficient D1(0), D1(1) = ∑ k g k S0(2×1 - k) = g0S0(2) + g1S0(1) + g2S0(0) = 0.3×0.71875 + 0.2×0.75 + 0.1×0.6875 = 0.434375. Calculate other detail coefficients by analogy.

[0039] Filter the detail coefficients to generate the target detail coefficients. Apply the VisuShrink method to filter the detail coefficient D1. Assume that through calculation, σ = 0.05 and n = 10, then For example, D1(0) = 0.20625 > 0.072, keep it; assume D1(3) = 0.05 < 0.072, then set it to 0 to obtain the target detail coefficient D′1.

[0040] Perform inverse wavelet transform on the approximation coefficient and the target detail coefficient to generate the denoised fetal heart rate signal; assume that the approximation coefficient A1 obtained through the previous steps is [0.06875, 0428125, 0.53125, 0.6171875, 0.703125] (for simplicity of the example, only 5 coefficients are taken), and the target detail coefficient D′1 is [0.20625, 0.434375, 0.328125, 0.25, 0.1875] (also 5 coefficients are taken).

[0041] The inverse wavelet transform formula used (taking the first - level inverse transform as an example) is: where h k and g k are the filter coefficients used in the previous discrete wavelet transform (the same as the previous example, h k = {0.1, 0.2, 0.3}, g k= {0.3, 0.2, 0.1}).

[0042]

[0043] Since A1(-1) and D1(-1) do not exist (boundary cases), assuming the boundary is handled as 0, then: S0(1) = h0A1(1) + h1A1(0) + g0D1(1) + g1D1(0) = 0.1×0.428125 + 0.2×0.06875 + 0.3×0.434375 + 0.2×0.20625 = 0.228125. And so on, the entire denoised fetal heart rate signal S is calculated. denoise sequence.

[0044] The Gaussian second wavelet function and the denoised fetal heart rate signal are convolved to generate a two-dimensional matrix. The method includes a calculation formula for obtaining the two-dimensional matrix, and the calculation formula is: where W (a,b) represents the two-dimensional matrix, whose first dimension a represents the scale and the second dimension b represents the time shift; is the complex conjugate of φ a,b (t), φ a,b (t) represents the Gaussian second wavelet function, S denoise represents the denoised fetal heart rate signal, and dt represents the time interval. Assuming a = 2 and b = 3, first calculate Then calculate the convolution integral to obtain the value of W (2,3) , and by changing the values of a and b, the entire two-dimensional matrix W is calculated.

[0045] The two-dimensional matrix is processed based on a preset color mapping table to generate a target wavelet transform diagram. Based on the preset color mapping table (for example, mapping the value range [0, 0.2] in the W matrix to blue, [0.2, 0.4] to green, [0.4, 0.6] to yellow, [0.6, 0.8] to orange, [0.8, 1] to red), the two-dimensional matrix is processed to convert the numerical values in the matrix into corresponding color information, thereby generating the target wavelet transform diagram. This target wavelet transform diagram can clearly show the changes in the fetal heart rate signal of the target fetus at different times and frequencies, that is, it characterizes the time-frequency information of the fetal heart rate signal of the target fetus.

[0046] S104, process the fetal hypoxia diagnosis model based on Vision Transformer using the training sample set to generate a target fetal hypoxia diagnosis model.

[0047] In one implementation, any number of data features in the training sample set are obtained, and a sampling ratio is generated based on the number of each data feature in the training sample set. Suppose our training sample set contains relevant data of 1000 pregnant women, and the data of each pregnant woman includes fetal heart rate signal sequence information, pregnant woman physiological parameter information, fetal PH value information, and a label (identification information) indicating whether the fetus is hypoxic. First, we randomly select 200 data features from these 1000 samples (here, the data features refer to the feature vectors obtained after processing the fetal heart rate signal, physiological parameters, etc.). Calculate the quantity ratio of each data feature in the training sample set. Suppose after statistics, the quantities of different types of data features are roughly as follows: there are 4000 fetal heart rate signal-related features, 1000 pregnant woman physiological parameter-related features, and 500 fetal PH value-related features. Then the total number of data features is 5500. The sampling ratio is: the ratio of the number selected from the fetal heart rate signal-related features is 200×(4000 / 5500)≈145; the ratio of the number selected from the pregnant woman physiological parameter-related features is 200×(1000 / 5500)≈36; the ratio of the number selected from the fetal PH value-related features is 200×(500 / 5500)≈18.

[0048] The training sample set is processed based on the sampling ratio to generate a preset number of sampling features. Each data feature is processed with each sampling feature to generate multiple groups of data sets, where each group of data sets contains a preset number of data samples, and at least one data sample includes identification information. According to the above sampling ratio, the corresponding number of sampling features is selected from the training sample set. For example, 145 are selected from the fetal heart rate signal-related features, 36 are selected from the pregnant woman physiological parameter-related features, and 18 are selected from the fetal PH value-related features, and combined into a preset number (here it is 200) of sampling features.

[0049] Then, these 200 sampling features are combined with other non-sampled data features to generate multiple groups of data sets. Suppose we want to generate data sets with 50 data samples in each group, then 20 groups of data sets can be generated. In each group of data sets, ensure that at least one data sample has an identification information (i.e., the label indicating whether the fetus is hypoxic). For example, in the first group of data sets, there are 30 samples of normal fetuses and 20 samples of hypoxic fetuses, and each sample contains the corresponding fetal heart rate signal features, pregnant woman physiological parameter features, and fetal PH value features, etc.

[0050] Train a fetal hypoxia diagnosis model based on Vision Transformer using data samples in multiple data groups to generate a trained fetal hypoxia diagnosis model. Process the trained fetal hypoxia diagnosis model based on a validation sample set to generate a validation result. If the data sample containing identification information in the validation result represents a risk factor affecting fetal hypoxia, then use the trained fetal hypoxia diagnosis model as the target fetal hypoxia diagnosis model. Train the fetal hypoxia diagnosis model based on Vision Transformer using data samples in these data groups. During the training process, the model will learn the relationship between different data features and whether the fetus has hypoxia. For example, the model may find that when certain specific fetal heart rate change patterns, the pregnant woman's history of hypertension, and a lower fetal pH value are combined, the possibility of fetal hypoxia is relatively high.

[0051] After a certain number of training iterations, use the validation sample set (assuming the validation sample set contains 200 samples) to process the trained fetal hypoxia diagnosis model. The model makes predictions for each sample in the validation sample set and generates a validation result. For example, if the model predicts that the fetus in a certain sample has hypoxia and the true label of this sample is also hypoxia, then the prediction of this sample is correct; otherwise, it is incorrect. Analyze the validation result. If it is found that among the data samples containing identification information in the validation result, most are correctly predicted as representing risk factors affecting fetal hypoxia (such as the accuracy rate reaching more than 80%), then use the trained fetal hypoxia diagnosis model at this time as the target fetal hypoxia diagnosis model. If the accuracy rate does not meet the requirements, it may be necessary to adjust the model parameters, increase the amount of training data, or improve the training method, etc., and retrain and validate until a target fetal hypoxia diagnosis model that meets the requirements is obtained.

[0052] S105, process the physiological parameter information of the pregnant woman and the pH value information of the target fetus based on the target fetal hypoxia diagnosis model to generate target physiological state features.

[0053] In one implementation, process the physiological parameter information of the pregnant woman and the pH value information of the target fetus based on the target fetal hypoxia diagnosis model to generate high-risk pregnancy factor information and normal physiological state information. Assume that the physiological parameter information of the pregnant woman includes 32 years old, having gestational diabetes (marked as 1, 0 if not), blood pressure 130 / 85 mmHg (systolic pressure / diastolic pressure), having had 1 premature birth history (marked as 1, 0 if not), and the pH value of the target fetus is 7.2. Such as Figure 3As shown, the nine embedding layers of the Health Encoder (HER) process this information respectively: for the age of 32, its corresponding embedding layer maps 32 years old into a vector that can reflect potential risk characteristics according to the association pattern between age and fetal health learned during model training. For example, it is found that there is a certain similarity in the fetal health risks of pregnant women aged 30 - 35 in a large amount of data, so 32 years old may be mapped to [0.12, 0.23, 0.31].

[0054] For a pregnant woman with gestational diabetes (labeled as 1), the corresponding embedding layer, based on the impact of diabetes on fetal nutrient supply and metabolic environment, converts it into a vector containing relevant information, such as [0.41, 0.52, 0.63], through learning a large number of cases. Each dimension represents the potential impact degree on the fetus, such as blood glucose control and complication risk. The embedding layer for a blood pressure of 130 / 85 mmHg maps it to [0.22, 0.33, 0.41] according to the normal blood pressure range and the statistical relationship with the risk of fetal hypoxia. These values reflect the characteristic representation of this blood pressure value in the fetal health risk assessment. Having a history of 1 premature birth (labeled as 1) is converted by a specific embedding layer into a vector containing premature birth - related risk information, such as [0.31, 0.42, 0.53], which reflects the impact characteristics of premature birth on aspects such as the uterine environment and fetal maturity. For the pH value of 7.2 of the target fetus, the embedding layer encodes it as [0.13, 0.21, 0.32] based on the normal range of fetal pH value and its correlation with hypoxia, indicating its position and risk degree in the assessment of fetal health status.

[0055] Based on the target fetal hypoxia diagnosis model, feature extraction and processing are performed on high - risk pregnancy factor information and normal physiological state information to generate initial physiological state features. The HER module performs an element - wise addition operation on the vectors output by these nine embedding layers: adding the vectors corresponding to the above - mentioned various health status information, that is, [0.12 + 0.41 + 0.22 + 0.31 + 0.13, 0.23 + 0.52 + 0.33 + 0.42 + 0.21, 0.31 + 0.63 + 0.41 + 0.53 + 0.32] = [1.19, 1.71, 2.20]. This vector represents the comprehensive health status of the mother and fetus, that is, the initial physiological state features are generated. Mapping processing is performed on the initial physiological state features to generate a continuous low - dimensional vector space. Then, mapping processing is performed on this initial physiological state feature through subsequent fully - connected layers or other transformation layers: Assume that a linear transformation matrix W (here assume W is a 3×3 matrix, and its element values are obtained through model training) is used to transform the initial physiological state feature vector, and it is converted into a continuous low - dimensional vector space. After calculation (vector - matrix multiplication), [0.34, 0.51, 0.23] is obtained.

[0056] Semantic analysis is performed on the initial physiological state characteristics to generate semantic information for each physiological state characteristic. For the vector part corresponding to diabetes, the model understands its significance in the overall health status based on the training knowledge, that is, diabetes may increase the risk of fetal hypoxia and is closely related to factors such as blood glucose control. For example, if blood glucose control is poor, it may lead to fetal nutritional metabolism disorders, which in turn affect fetal oxygen supply. For a history of premature birth, the model knows that this will make the uterine environment unstable, may affect placental blood perfusion, and thus increase the risk of fetal hypoxia. In terms of blood pressure, the model understands that high blood pressure may affect placental blood perfusion and reduce fetal oxygen supply. In terms of fetal pH value, the model knows that a slightly low pH value may indicate a certain degree of acidosis risk in the fetus, reflecting that fetal oxygen supply may be insufficient.

[0057] The semantic information of each physiological state characteristic and the continuous low-dimensional vector space are processed to generate the target physiological state characteristic. In the low-dimensional vector space, the first dimension 0.34 combined with the semantic information of diabetes indicates that the influence degree of this factor on the risk of fetal hypoxia in the current comprehensive physiological state is moderately low; the second dimension 0.51 combined with the semantic information of a history of premature birth indicates that the influence degree of the factor of a history of premature birth is moderate; the third dimension 0.23 combined with the semantic information of fetal pH value indicates that the influence degree of the fetal pH value factor is low. The finally generated target physiological state characteristic can more accurately reflect the association between the physiological state of the pregnant woman and the fetus and the risk of fetal hypoxia, providing an important basis for the subsequent fusion and diagnosis of the model.

[0058] S106, Process the target wavelet transform image based on the target fetal hypoxia diagnosis model to generate target image features.

[0059] In one implementation, the target wavelet transform diagram is processed based on the target fetal hypoxia diagnosis model to generate the frequency change features in the target wavelet transform diagram. Suppose we have obtained a target wavelet transform diagram, which shows the time-frequency information of the fetal heart rate signal of the target fetus in the form of an image. The abscissa of the image represents time, the ordinate represents frequency, and the color depth represents the intensity of the signal at that time and frequency. The signal encoder (SER) in the model first processes the target wavelet transform diagram. The dynamic snake convolution (DSConv) module in SER starts to work. It slides on the target wavelet transform diagram with a 3×3 convolution kernel (for example, the central coordinates are Ki=(Xi, yi)). Since DSConv can adaptively focus on slender and tortuous local structures, it can better perceive frequency changes. For example, if there are obvious fluctuations in the signal intensity in a certain frequency band (suppose it is the frequency region corresponding to 100 - 120 bpm) in the wavelet transform diagram, DSConv can capture this fluctuation through its unique convolution method and extract the frequency change features in this region. These features may include information such as the amplitude of frequency change and the rate of change, forming a feature vector about frequency change, such as [0.2 (amplitude change value), 0.3 (rate change value)].

[0060] Process the frequency change features in the target wavelet transform diagram to generate the target time-frequency feature map. Based on the extracted frequency change features, the SER module further processes them, fusing and adjusting the frequency change features with the original time-frequency information in the target wavelet transform diagram. For example, if the frequency change amplitude is large and the rate is fast in a certain time period, the model will, according to the learned knowledge, perform weighted processing on the time-frequency information around this time period to highlight these regions. After such processing, the target time-frequency feature map is generated. Compared with the original target wavelet transform diagram, this map can better highlight the key time-frequency change regions and features. Perform block and linear projection processing on the target time-frequency feature map to generate the target feature sequence information. Next, perform block and linear projection processing on the target time-frequency feature map. In the same way as ViT-B / 16, the target time-frequency feature map with a resolution of 224×224 is divided into a sequence of 196 patches, and the resolution of each patch is 16×16. Then, through the linear projection module, the dimension of each block is mapped to 768 dimensions. For example, for the first 16×16 patch, after linear projection, a 768-dimensional vector is obtained, which contains the time-frequency feature information in this patch region. Perform such processing on all 196 patches, and finally generate the target feature sequence information, which is a sequence composed of 196 768-dimensional vectors.

[0061] Process the target feature sequence information to generate query vectors, key vectors, and value vectors. When the target feature sequence information is obtained, the Transformer module in the model starts to function. The Transformer module first performs a linear transformation on the target feature sequence information, transforming the input sequence through learnable weight matrices WQ, WK, WV (these weight matrices are continuously adjusted and optimized during model training). Assume that the i-th 768-dimensional vector in the target feature sequence information is Xi, then the query vector Qi = XiWq, the key vector Ki = XiWk, and the value vector Vi = XiWv. For example, for the first vector X1, after multiplying with the weight matrix, query vector Q1, key vector K1, and value vector V1 are obtained. They are vectors of different dimensions, and these vectors will be used for subsequent attention mechanism calculations.

[0062] Perform dot product and scaling processing on the query vectors and key vectors to generate attention weight information. Then calculate the attention weights. The Transformer module performs a dot product operation between the query vector Qi and all key vectors Ki, and then performs scaling (dividing by the square root of the dimension of the query vector Qi). For example, for the query vector Q1 and key vectors K1, K2,..., Kn, first calculate the dot products Q1·K1, Q1·K2,..., Q1·Kn, and then divide each dot product result by the square root of the dimension of Q1 to obtain attention scores. Then apply the softmax function to convert the attention scores into a probability distribution, so as to obtain the attention weight AttentionWeight(Q1,Kj) = softmax(ArenionScore(Q1,Kj)). In this way, the attention weight information is generated, which represents the degree of association and importance between features at different positions.

[0063] Perform weighted summation processing on the value vectors based on the attention weight information to generate the target image features. For the attention weights and value vectors at each position, use the attention weights to perform weighted summation on the value vectors. For example, for the value vectors V1, V2,..., Vn and the corresponding attention weights AttentionWeight(Q1,K1), AttentionWeight(Q1,K2),..., AttentionWeight(Q1,Kn), calculate SelfAttention(X)1 = ∑AttentionWeight(Q1,Kj)·Vj.

[0064] For multi-head self-attention, the calculation result is obtained by concatenating the outputs of multiple single-head self-attention, MultiHeadSelfAttention(X) = Concat(head1, head2, …, headn)Wo, and finally the target image features are obtained. These target image features fuse the key information in the target wavelet transform image and highlight important time-frequency features through the attention mechanism, providing an important basis for subsequent fusion with the target physiological state features and fetal hypoxia diagnosis.

[0065] S107, process the target physiological state features and the target image features based on the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

[0066] In one implementation, the target physiological state features and the target image features are fused based on the target fetal hypoxia diagnosis model to generate a fused feature sequence. Assume that the target physiological state features have been obtained as a vector containing the pregnant woman's health status and fetal-related physiological features, such as [0.3, 0.5, 0.2, 0.4] (where each dimension represents the weight of different physiological factors. For example, 0.3 is related to the pregnant woman's blood pressure status, 0.5 is related to the pregnant woman's blood sugar condition, etc.), and the target image features are a vector containing the key time-frequency information of the fetal heart rate signal after the previous processing, such as [0.2, 0.3, 0.1, 0.4, 0.5] (each dimension reflects the signal feature weights at different times and frequencies). The fusion layer in the model will fuse the target physiological state features and the target image features, using simple concatenation or weighted summation, etc. For example, using the concatenation method, the two feature vectors are combined into a new vector [0.3, 0.5, 0.2, 0.4, 0.2, 0.3, 0.1, 0.4, 0.5], and this new vector is the fused feature sequence. It synthesizes the physiological state information of the pregnant woman and the fetus and the time-frequency information of the fetal heart rate signal, providing a more comprehensive basis for subsequent classification.

[0067] Process the fused feature sequence based on the target fetal hypoxia diagnosis model to generate target classification head information, and process the target classification head information to generate fetal hypoxia information. The fused feature sequence will be input into the Transformer module for further processing. The Transformer module analyzes and transforms the fused feature sequence through its multi-head self-attention mechanism and feed-forward neural network. It will recalculate the correlations and weights between features and extract higher-level feature representations. For example, after a series of linear transformations, attention calculations, and non-linear activation operations, the fused feature sequence is transformed into a new vector, such as [0.4, 0.3, 0.2], and this vector is the target classification head information. It abstracts and generalizes the fused features in a lower-dimensional space and is more suitable for the final classification decision.

[0068] The target classification header information is input into the classification header. The classification header usually consists of one or more fully connected layers. Suppose there is a simple two-layer fully connected layer structure. The first layer calculates the target classification header information [0.4, 0.3, 0.2] through the weight matrix W1 (assuming W1 is a 3×5 matrix) and the bias vector b1 (assuming b1 is [0.1, 0.2, 0.3, 0.4, 0.5]) to obtain an intermediate vector, such as [0.5, 0.6, 0.7, 0.8, 0.9]. Then the second layer calculates through the weight matrix W2 (assuming W2 is a 5×2 matrix) and the bias vector b2 (assuming b2 is [0.1, 0.2]) to finally obtain a two-dimensional vector, such as [0.7, 0.3]. Here, it can be set that if the value of the first dimension is greater than 0.5, it indicates fetal hypoxia, otherwise it indicates the fetus is normal. The finally generated fetal hypoxia information is fetal hypoxia. Through the above example, the generation process from the target physiological state features and target image features to the fetal hypoxia information is demonstrated, as well as the roles of structures such as the fusion layer, Transformer module, and classification header in the model.

[0069] The server obtains the fetal heart rate (FHR) signal, pH value information, pregnant woman's physiological parameter information, diagnostic model, and training sample set of the target fetus. First, downsample and slice the FHR signal, detect and process outliers and missing values to generate complete target fetal heart rate signal sequence information. Then use the second Gaussian wavelet function to process this sequence information, and through operations such as normalization, discrete wavelet transform, detail coefficient filtering, inverse transform, convolution, and color mapping, generate a target wavelet transform map representing the time-frequency information of the fetal heart rate signal.

[0070] In terms of model training, data features are obtained from the training sample set and the sampling ratio is determined. After generating sampling features, they are combined into data groups for training the model, and then verified with the validation sample set. The trained model that meets the conditions becomes the target model. For feature processing, the model generates high-risk pregnancy and normal physiological state information based on the information of the pregnant woman and the fetus, extracts features and maps them to a low-dimensional vector space, and generates target physiological state features through semantic analysis; at the same time, frequency change features are extracted from the target wavelet transform map and processed to generate target image features. Finally, the target physiological state and image features are fused, and through model processing, a fusion feature sequence and target classification header information are obtained, and then fetal hypoxia information is generated to determine whether the fetus is hypoxic. This method combines information from multiple aspects, provides an effective means for fetal hypoxia diagnosis, and is expected to assist clinical decision-making and ensure fetal health.

[0071] In one implementation, as Figure 2 shown, the present application also provides a fetal hypoxia warning device combining health status and fetal heart rate, including:

[0072] An acquisition module 201, configured to acquire the fetal heart rate signal of a target fetus, the PH value information of the target fetus, the physiological parameter information of the pregnant woman, a fetal hypoxia diagnosis model based on VisionTransformer, and a training sample set;

[0073] A processing module 202, configured to preprocess the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information; process the target fetal heart rate signal sequence information based on the second Gaussian wavelet function to generate a target wavelet transform diagram, where the target wavelet transform diagram is used to characterize the time-frequency information of the fetal heart rate signal of the target fetus; process the fetal hypoxia diagnosis model based on VisionTransformer using the training sample set to generate a target fetal hypoxia diagnosis model; process the physiological parameter information of the pregnant woman and the PH value information of the target fetus using the target fetal hypoxia diagnosis model to generate target physiological state features; process the target wavelet transform diagram using the target fetal hypoxia diagnosis model to generate target image features; process the target physiological state features and the target image features using the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

[0074] Each embodiment in this application is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the embodiments of the fetal hypoxia early warning method, electronic device, electronic equipment, and readable storage medium that evaluate the combination of health status and fetal heart rate, since they are basically similar to the embodiments of the fetal hypoxia early warning method that combines health status and fetal heart rate described above, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the embodiments of the fetal hypoxia early warning method that combines health status and fetal heart rate described above.

Claims

1. A fetal hypoxia early warning method combining health status and fetal heart rate, characterized in that: include: Obtain the fetal heart rate signal of the target fetus, the pH value information of the target fetus, the physiological parameter information of the pregnant woman, the fetal hypoxia diagnosis model based on VisionTransformer, and the training sample set; Preprocessing the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information; Processing the target fetal heart rate signal sequence information based on the Gaussian second wavelet function to generate a target wavelet transform map, wherein the target wavelet transform map is used to characterize the time-frequency information of the fetal heart rate signal of the target fetus; Processing the VisionTransformer-based fetal hypoxia diagnostic model based on the training sample set to generate a target fetal hypoxia diagnostic model; Processing the physiological parameter information of the pregnant woman and the pH value information of the target fetus based on the target fetal hypoxia diagnostic model to generate target physiological state characteristics; Processing the target wavelet transform image based on the target fetal hypoxia diagnostic model to generate target image features; The target physiological state characteristics and the target image characteristics are processed based on the target fetal hypoxia diagnostic model to generate fetal hypoxia information.

2. The method according to claim 1, characterized in that Preprocessing the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information includes: Down-sampling and data slicing processing are performed on the fetal heart rate signal of the target fetus to generate a plurality of data segments; Performing outlier detection processing on a plurality of data segments respectively to generate target data segments, wherein the target data segments are data segments with abnormal data; Processing the target data segment to generate missing sample information, wherein the missing sample information is used to characterize samples with missing values ​​and the locations and variable names of the missing values; Processing the missing sample information to generate missing value interpolation variable values; The target data segment is processed based on the missing value interpolation variable value to generate target fetal heart rate signal sequence information, wherein the target fetal heart rate signal sequence information is composed of a number of data segments with complete values.

3. The method according to claim 1, characterized in that The target fetal heart rate signal sequence information is processed based on the Gaussian second wavelet function to generate a target wavelet transform graph, including: Performing standardization processing on the target fetal heart rate signal sequence information to generate a standardized segment signal; Performing discrete wavelet transform processing on the standardized segment signal based on a preset decomposition formula to generate approximate coefficients and detail coefficients; Performing filtering processing on the detail coefficients to generate target detail coefficients; Performing inverse wavelet transform processing on the approximate coefficient and the target detail coefficient to generate a denoised fetal heart rate signal; Performing convolution processing on the Gaussian second wavelet function and the denoised fetal heart rate signal to generate a two-dimensional matrix; Processing the two-dimensional matrix based on a preset color mapping table to generate a target wavelet transform image; The method includes a calculation formula for obtaining a standardized segment signal, the calculation formula being: S′={s′ / max(S)|s′∈S}; Where S′ represents the standardized segment signal, S={s1s2,...,s n } represents the target fetal heart rate signal sequence; The method includes a first decomposition formula for obtaining approximation coefficients, wherein the first decomposition formula is: A j+1 (n)=∑ k h k S j (2n-k); Among them, A j+1 (n) represents the approximation coefficient, h k represents a low-pass filter of level j, S j (2n-k) represents the signal value of the original signal s after time scale transformation and filter action at the jth level of decomposition; The method includes a second decomposition formula for obtaining detail coefficients, wherein the second decomposition formula is: D j+1 (n)=∑ k g k S j (2n-k); Among them, D j+1 (n) represents the detail coefficient, g k represents a high-pass filter of level j; The method includes a calculation formula for obtaining a two-dimensional matrix, and the calculation formula is: Among them, W (a,b) represents a two-dimensional matrix, where the first dimension a represents the scale and the second dimension b represents the time shift; Yes a,b The complex conjugate of (t), φ a,b (t) represents the Gaussian second wavelet function, S denoise represents the denoised fetal heart rate signal, and dt represents the time interval.

4. The method according to claim 1, characterized in that The fetal hypoxia diagnosis model based on VisionTransformer is processed based on the training sample set to generate a target fetal hypoxia diagnosis model, including: Get any number of data features in the training sample set; Generate a sampling ratio based on the number of each data feature in the training sample set; Processing the training sample set based on the sampling ratio to generate a preset number of sampling features; Processing based on any data feature and each sampling feature to generate multiple data groups, wherein each data group includes a preset number of data samples, and at least one data sample includes identification information; Training the VisionTransformer-based fetal hypoxia diagnosis model based on data samples in the multiple data groups to generate a trained fetal hypoxia diagnosis model; Processing the trained fetal hypoxia diagnosis model based on the verification sample set to generate a verification result; If the data sample containing identification information in the verification result represents a risk factor affecting fetal hypoxia, the trained fetal hypoxia diagnostic model is used as the target fetal hypoxia diagnostic model.

5. The method according to claim 1, characterized in that The physiological parameter information of the pregnant woman and the pH value information of the target fetus are processed based on the target fetal hypoxia diagnostic model to generate target physiological state characteristics, including: Processing the physiological parameter information of the pregnant woman and the pH value information of the target fetus based on the target fetal hypoxia diagnostic model to generate high-risk pregnancy factor information and normal physiological state information; Based on the target fetal hypoxia diagnostic model, feature extraction processing is performed on the high-risk pregnancy factor information and the normal physiological state information to generate initial physiological state features; Mapping the initial physiological state characteristics to generate a continuous low-dimensional vector space; Performing semantic analysis on the initial physiological state features to generate semantic information of each physiological state feature; The semantic information and continuous low-dimensional vector space of each physiological state feature are processed to generate the target physiological state feature.

6. The method according to claim 3, characterized in that The target wavelet transform image is processed based on the target fetal hypoxia diagnostic model to generate target image features, including: Processing the target wavelet transform image based on the target fetal hypoxia diagnostic model to generate frequency change features in the target wavelet transform image; Processing the frequency change characteristics in the target wavelet transform graph to generate a target time-frequency characteristic graph; Performing block division and linear projection processing on the target time-frequency feature map to generate target feature sequence information; Processing the target feature sequence information to generate a query vector, a key vector and a value vector; Perform dot product and scaling on the query vector and key vector to generate attention weight information; The value vector is weighted and summed based on the attention weight information to generate target image features.

7. The method according to claim 6, characterized in that Processing the target physiological state characteristics and the target image characteristics based on the target fetal hypoxia diagnostic model to generate fetal hypoxia information includes: Based on the target fetal hypoxia diagnostic model, the target physiological state feature and the target image feature are fused to generate a fused feature sequence; Processing the fused feature sequence based on the target fetal hypoxia diagnosis model to generate target classification header information; The target classification header information is processed to generate fetal hypoxia information.

8. A fetal hypoxia warning device combining health status and fetal heart rate, characterized in that: The device comprises: The acquisition module is used to obtain the fetal heart rate signal of the target fetus, the pH value information of the target fetus, the physiological parameter information of the pregnant woman, the fetal hypoxia diagnosis model based on VisionTransformer, and the training sample set; A processing module is used to pre-process the fetal heart rate signal of the target fetus to generate target fetal heart rate signal sequence information; process the target fetal heart rate signal sequence information based on the Gaussian second wavelet function to generate a target wavelet transform map, wherein the target wavelet transform map is used to characterize the time-frequency information of the fetal heart rate signal of the target fetus; process the fetal hypoxia diagnosis model based on VisionTransformer based on the training sample set to generate a target fetal hypoxia diagnosis model; process the pregnant woman's physiological parameter information and the target fetus's pH value information based on the target fetal hypoxia diagnosis model to generate target physiological state characteristics; process the target wavelet transform map based on the target fetal hypoxia diagnosis model to generate target image characteristics; process the target physiological state characteristics and the target image characteristics based on the target fetal hypoxia diagnosis model to generate fetal hypoxia information.

9. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the fetal hypoxia warning method combining health status and fetal heart rate as described in any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the fetal hypoxia early warning method combining health status and fetal heart rate as described in any one of claims 1 to 7 is implemented.

Images