Construction method and system of fetal heart monitoring auxiliary interpretation model based on one-dimensional time sequence signal and double neural network architecture

Abstract

The present application relates to the technical field of healthcare informatics, and discloses a construction method and system of a fetal heart monitoring auxiliary interpretation model based on one-dimensional time sequence signals and a double neural network architecture. According to the fetal heart monitoring data and monitoring result label data of a group of pregnant women, the relationship between the fetal heart monitoring data and the monitoring result label data is learned by using a double neural network architecture, and a fetal heart monitoring auxiliary interpretation model is trained. The present application adopts a time-aligned double signal sampling strategy, and parallelly inputs the double-lead one-dimensional time sequence data of the fetal heart rate and uterine contraction. A double-branch neural network architecture is adopted to independently process the signals of one lead in each branch, which can fully utilize the complementary information between different lead signals, accurately identify the physiological state of the fetus and the abnormal state of the fetus of the pregnant woman, and significantly improve the recognition accuracy of the fetal heart monitoring auxiliary interpretation model for complex abnormal conditions such as fetal hypoxia and uterine contraction compression, and time-dependent pathological states such as late deceleration and variable deceleration.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and in particular to a method and system for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture. Background Technology

[0002] Fetal heart rate monitoring is a commonly used tool for continuously monitoring the health of the fetus during labor. It assesses the fetal health status in utero by monitoring real-time changes in fetal heart rate (FHR) and uterine contractions (UC) using a fetal heart rate monitor. Fetal heart rate monitoring helps to detect abnormalities such as fetal distress early, providing a basis for clinical intervention and thus reducing the incidence of adverse perinatal outcomes.

[0003] Traditional fetal heart rate monitoring relies primarily on subjective visual assessment by the obstetric medical team, but this approach has limitations. Due to the diverse variations in fetal heart rate monitoring patterns and the existence of different standards for diagnostic rules, discrepancies often exist between and within observers, leading to inconsistencies in interpretation. This subjective influence can delay clinical decisions and affect maternal and infant outcomes. While artificial intelligence methods have been introduced to analyze fetal heart rate monitoring data, many existing methods rely on single neural networks or simple feature extraction methods, making it difficult to comprehensively capture the complex features and temporal relationships within the data. In particular, they neglect the complex correlation between fetal heart rate (FHR) and uterine contractions (UC), potentially resulting in inaccurate assessments of fetal physiological status and an inability to accurately identify subtle abnormalities.

[0004] Therefore, there is an urgent need to construct a new fetal heart rate monitoring auxiliary interpretation model to help achieve more accurate and efficient fetal heart rate monitoring. Summary of the Invention

[0005] This invention provides a method and system for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, in order to address the shortcomings of current artificial intelligence methods for analyzing fetal heart rate monitoring data that are not accurate enough.

[0006] This invention provides a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, comprising:

[0007] Acquire fetal heart rate monitoring data and corresponding monitoring result label data for pregnant women. The fetal heart rate monitoring data includes one-dimensional time series data of fetal heart rate and one-dimensional time series data of uterine contractions. The monitoring result label data includes fetal physiological status labels and abnormal interval labels of pregnant women and fetus.

[0008] Based on fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, a dual neural network architecture is used to learn the relationship between fetal heart rate monitoring data and corresponding monitoring result label data, and a fetal heart rate monitoring auxiliary interpretation model is trained.

[0009] It should be noted that one-dimensional time-series data of fetal heart rate represents the sequence of changes in the frequency of the fetal heartbeat over time, while one-dimensional time-series data of uterine contractions represents the sequence of changes in the intensity and frequency of uterine muscle contractions over time.

[0010] In one embodiment, the fetal heart rate monitoring data is organized in the form of a two-dimensional array, wherein the first row of the array is used to store one-dimensional time-series signal data of FHR, the second row of the array is used to store one-dimensional time-series signal data of UC, and each column of the array represents a specific sampling data point, namely the FHR value and UC value recorded at the same time point.

[0011] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The fetal physiological state includes any one of the following or any combination thereof: fetal hypoxia state, blood pH state, blood sample state, and blood carbon dioxide state. The fetal hypoxia state is labeled in a classification form as "yes" or "no". The blood pH state, blood oxygen state, and blood carbon dioxide state are labeled in segmented intervals and divided into three categories: "good", "abnormal", and "severely abnormal".

[0012] According to the present invention, a method for constructing a fetal heart rate monitoring-assisted interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The abnormal fetal status interval label is used to indicate the specific time-series interval in the time-series array where maternal and fetal abnormalities exist. The start and end times of the abnormality are precisely marked in the time-series array and represented in the form of an abnormality index interval, such as "[start time, end time]", clearly indicating the specific time period in which the abnormality occurred. Simultaneously, a textual description or code of the abnormality type can be added, such as fetal distress, bradycardia, tachycardia, uterine contractions, etc.

[0013] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture may include the following steps after acquiring fetal heart rate monitoring data and corresponding monitoring result label data of a pregnant woman population:

[0014] Fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women are preprocessed. Preprocessing includes any one or any combination of the following: linear interpolation, sample pruning, and normalization.

[0015] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional temporal signal and a dual neural network architecture is provided. The dual neural network architecture includes a classification and recognition neural network architecture and an anomaly detection and evaluation neural network architecture. The classification and recognition neural network architecture includes a feature extraction module, a classification and recognition module, and a first embedding layer connecting the feature extraction module and the classification and recognition module. The feature extraction module includes a temporal representation submodule (a one-dimensional convolutional neural network submodule), a global average pooling submodule, an encoder, and a projector. The classification and recognition module includes a fully connected layer, a batch normalization submodule, and a Softmax (normalized exponential function) activation function. The anomaly detection and evaluation neural network architecture includes a second embedding layer, an anomaly detection module based on a Transformer architecture, and a projection layer. The second embedding layer includes a mask embedding submodule and a position encoding submodule. The anomaly detection module based on a Transformer architecture includes a first encoder layer, a second encoder layer, and a third encoder layer. The first encoder layer, the second encoder layer, and the third encoder layer each include a multi-head attention layer and a multilayer perceptron.

[0016] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The method involves using a dual neural network architecture to learn the relationship between the fetal heart rate monitoring data and the corresponding monitoring result label data of a pregnant woman population, thereby training the fetal heart rate monitoring auxiliary interpretation model. The method includes:

[0017] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, feature extraction is performed through the feature extraction module of the classification and recognition neural network architecture in the dual neural network architecture to obtain comprehensive features;

[0018] Based on the comprehensive features, the comprehensive feature vector is obtained through the first embedding layer of the classification and identification neural network architecture in the dual neural network architecture;

[0019] Based on the comprehensive feature vector, the classification recognition module of the classification recognition neural network architecture in the dual neural network architecture maps the comprehensive feature vector to the output category, thereby realizing the classification of fetal physiological status.

[0020] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The method involves extracting features from fetal heart rate monitoring data of pregnant women and corresponding monitoring result label data using a feature extraction module in the classification and recognition neural network architecture of the dual neural network architecture to obtain comprehensive features, including:

[0021] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, a temporal representation submodule with feature extraction function is constructed. Different sizes and strides of convolution kernels are used to capture different temporal relationships in the fetal heart rate monitoring data and corresponding monitoring result label data, and multiple convolutional features are obtained.

[0022] Based on the features derived from multiple convolutions, the encoder in the feature extraction module encodes them separately to obtain multiple encoded features.

[0023] Based on multiple encoded features, the projector of the feature extraction module maps the encoded features to the classification space, and then concatenates the multiple mapped features along the channel dimension to obtain the comprehensive features.

[0024] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The method involves using a dual neural network architecture to learn the relationship between the fetal heart rate monitoring data and the corresponding monitoring result label data of a pregnant woman population, thereby training the fetal heart rate monitoring auxiliary interpretation model. The method includes:

[0025] Based on fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, temporal features and location information are obtained through the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture.

[0026] Based on temporal features and location information, the anomaly detection module, which is based on the Transformer architecture and is used in the anomaly detection and evaluation neural network architecture of the dual neural network architecture, calculates attention weights and correlation differences to identify anomalies in temporal features and location information, thereby realizing the identification of abnormal fetal status in pregnant women.

[0027] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The method involves obtaining time-series features and location information based on fetal heart rate monitoring data of a pregnant woman population and corresponding monitoring result label data, through the second embedding layer of the anomaly monitoring and evaluation neural network architecture in the dual neural network architecture. The method includes:

[0028] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, the temporal characteristics of the corresponding monitoring result labels are obtained through the mask embedding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture.

[0029] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, the positional information of the corresponding monitoring result label is obtained through the positional encoding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture.

[0030] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signals and a dual neural network architecture is provided. The method involves identifying abnormal points in the time-series features and location information by calculating attention weights and correlation differences through a Transformer-based abnormality monitoring module within the abnormality monitoring and evaluation neural network architecture of the dual neural network architecture, thereby achieving the identification of abnormal fetal status in pregnant women. The method includes:

[0031] Based on temporal features and location information, the error sequence is identified by using the multi-head attention layer in the anomaly detection module based on the Transformer architecture, combined with the sliding window method and minimization optimization function. The error sequence is then expanded by anomaly padding, and continuous or overlapping error sequences are merged to form an anomaly identification signal for the pregnant woman's fetal status.

[0032] The anomaly detection module based on the Transformer architecture uses a multilayer perceptron to extract features based on the output of the corresponding multi-head attention layer, thus obtaining the anomaly recognition signal features.

[0033] By using the projection layer in the anomaly detection module based on the Transformer architecture, the features of the anomaly recognition signal are mapped to the required output dimension, which is the anomaly recognition result of the pregnant woman's fetal state.

[0034] According to the present invention, a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture is provided. The method further includes, based on fetal heart rate monitoring data and corresponding monitoring result label data of a pregnant woman population, using a dual neural network architecture to learn the relationship between the fetal heart rate monitoring data and the corresponding monitoring result label data, thereby training the fetal heart rate monitoring auxiliary interpretation model.

[0035] The fetal heart rate monitoring results are obtained by integrating the fetal physiological state classification results obtained through the classification recognition neural network architecture in the dual neural network architecture and the abnormal identification results of the pregnant woman's fetal state obtained through the abnormal monitoring and evaluation neural network architecture.

[0036] The present invention also provides a fetal heart rate monitoring auxiliary system, comprising:

[0037] The data receiving module is used to receive fetal heart rate monitoring data from at least one terminal of the pregnant woman to be tested.

[0038] The fetal heart rate monitoring auxiliary module is used to: obtain the fetal heart rate monitoring results of the pregnant woman based on the fetal heart rate monitoring data of the pregnant woman to be tested, through the fetal heart rate monitoring auxiliary interpretation model obtained by the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above.

[0039] The data output module is used to output the fetal heart rate monitoring results of the pregnant woman to at least one terminal.

[0040] It should be noted that a terminal refers to an input / output device connected to a computer system. Depending on the function, terminals can be divided into various types: smart terminals or intelligent terminals, dumb terminals, interactive terminals or online terminals. Specifically, a terminal can be various mobile communication devices, such as mobile phones and tablets. This article aims to provide users with the function of inputting data and outputting data.

[0041] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signals and a dual neural network architecture as described above.

[0042] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture as described above.

[0043] The present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute any of the above-described methods for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture.

[0044] The present invention provides a method and system for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, which can bring at least the following beneficial effects:

[0045] (1) By inputting one-dimensional time-series data of fetal heart rate and uterine contraction in parallel, the problem of signal coupling interference in traditional single-channel monitoring is solved, which greatly improves the accuracy of the fetal heart rate monitoring auxiliary interpretation model in identifying complex conditions such as fetal hypoxia and uterine contraction compression.

[0046] (2) The time-aligned dual-signal sampling strategy is adopted to preserve the time correlation characteristics of the original monitoring waveform, which can improve the prediction accuracy of the model in time-dependent pathological states such as late deceleration and variable deceleration.

[0047] (3) A dual-branch neural network architecture is adopted, with each branch independently processing the signal of one lead, making full use of the complementary information between signals from different leads.

[0048] (4) Using a classification and recognition neural network architecture, feature extraction and dimensionality reduction are performed through a feature extraction module, a first embedding layer, and a classification and recognition module. The features extracted from each branch are spliced ​​together in the channel dimension to form a comprehensive feature vector, which improves the richness and accuracy of feature expression. The comprehensive feature vector is mapped to the output category through a fully connected layer to achieve accurate classification and recognition of the fetal physiological state.

[0049] (5) Using an anomaly monitoring and evaluation neural network architecture, the signal error is calculated and the error distribution is statistically analyzed through the second embedding layer, the anomaly monitoring module based on the Transformer architecture, and the projection layer. The anomaly detection threshold is found using optimization techniques, and the degree of anomaly of the signal is comprehensively evaluated. The error sequence is found, and the accurate identification of the abnormal state of the pregnant woman and the fetus in the fetal heart monitoring signal is realized. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is one of the flowcharts illustrating a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, as provided by the present invention.

[0052] Figure 2 The second flowchart illustrates a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, as provided by the present invention. The flowchart shows the dual neural network architecture of the fetal heart rate monitoring auxiliary interpretation model.

[0053] Figure 3 This is a schematic diagram of the structure of a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, which is provided by the present invention.

[0054] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0056] Figure 1 and 2 This is a flowchart illustrating a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, as provided by the present invention.

[0057] The execution subject of the method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture provided by the present invention can be any applicable terminal-side device or network-side device, such as a device for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture.

[0058] See Figure 1 The present invention provides a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, which may include:

[0059] S110. Obtain fetal heart rate monitoring data and corresponding monitoring result label data for pregnant women. The fetal heart rate monitoring data includes one-dimensional time-series data of fetal heart rate and one-dimensional time-series data of uterine contractions. The monitoring result label data includes fetal physiological status labels and abnormal interval labels of pregnant women and fetuses.

[0060] In one embodiment, fetal heart rate monitoring data can be acquired using an external fetal heart rate monitoring device, and the monitoring result label data can be manually tagged. Specifically, one-dimensional time-series fetal heart rate data represents the sequence of changes in the frequency of the fetal heartbeat over time, and one-dimensional time-series uterine contraction data represents the sequence of changes in the intensity and frequency of the pregnant woman's uterine muscle contractions over time.

[0061] In one embodiment, the fetal heart rate monitoring data is organized in the form of a two-dimensional array. The first row of the array is dedicated to storing one-dimensional time-series FHR signal data, and the second row is used to store one-dimensional time-series UC signal data. Each column of the array represents a specific sampling data point, i.e., the FHR and UC values ​​recorded at the same time point. All signal data can be directly acquired by a professional fetal heart rate monitoring instrument with high precision and real-time performance, ensuring that the acquired FHR and UC signals accurately reflect the actual physiological state of the fetus and mother. During the acquisition process, the instrument continuously records FHR and UC signals according to a preset sampling frequency, forming continuous time-series data. Each sampling data point is an element in the array, and its value represents the measured FHR or UC value at a specific time point. The sequence order of the data points strictly corresponds to the time sequence, ensuring the effectiveness and accuracy of the time-series analysis.

[0062] In one embodiment, the fetal physiological state includes any one or any combination of the following: fetal hypoxia, blood pH, blood sample status, and blood carbon dioxide status. The fetal hypoxia is labeled in a classification form as "yes" or "no". The blood pH, blood oxygen, and blood carbon dioxide status are labeled in segmented intervals, divided into three categories: "good", "abnormal", and "severely abnormal".

[0063] In one embodiment, the abnormal fetal condition interval label is used to indicate the specific time series interval in the time series array where maternal and fetal abnormalities exist. The start and end times of the abnormality are precisely marked in the time series array and represented in the form of an abnormality index interval, such as "[start time, end time]", clearly indicating the specific time period during which the abnormality occurred. Simultaneously, a textual description or code of the abnormality type can be added, such as fetal distress, bradycardia, tachycardia, uterine contractions, etc.

[0064] In one embodiment, after acquiring fetal heart rate monitoring data and corresponding monitoring result label data, preprocessing can be performed. Preprocessing includes any one or any combination of the following: linear interpolation, sample pruning, and normalization. In this embodiment, the signal is first linearly interpolated to fill in any possible data gaps and ensure signal continuity. Then, the signal is sample pruning and normalization to scale the signal values ​​to a uniform range, eliminating dimensional differences and improving the algorithm's convergence speed and stability.

[0065] Simultaneous use of both fetal heart rate (FHR) and uterine contractions (UC) provides a more comprehensive reflection of the fetal-maternal interaction. Changes in fetal heart rate are physiologically correlated with the intensity of uterine contractions (e.g., changes in placental oxygen supply during contractions affect fetal heart rate). Dual-signal training helps the model capture this dynamic coupling, making it more clinically significant than single-signal analysis. The simultaneous use of fetal physiological state labels (e.g., hypoxia, pH) and abnormal interval labels (e.g., distress, bradycardia) enables joint supervision of microscopic physiological indicators and macroscopic clinical diagnosis. This hierarchical labeling system guides the neural network to learn basic physiological characteristics (e.g., baseline variability) and advanced pathological patterns (e.g., late decelerations), improving model generalization. Traditional fetal monitoring interpretation relies on expert experience (e.g., baseline, acceleration, and deceleration rules in the FIGO standard), while this approach, through direct training with raw one-dimensional time-series data, can automatically discover potential discriminative features (e.g., the phase difference between uterine contractions and fetal heart rate in specific frequency bands), potentially uncovering hidden abnormal patterns that are difficult to summarize manually.

[0066] S120. Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, a dual neural network architecture is used to learn the relationship between the fetal heart rate monitoring data and the corresponding monitoring result label data, and a fetal heart rate monitoring auxiliary interpretation model is trained.

[0067] In one embodiment, see Figure 2 The dual neural network architecture includes a classification and recognition neural network architecture and an anomaly detection and evaluation neural network architecture. The classification and recognition neural network architecture includes a feature extraction module, a classification and recognition module, and a first embedding layer connecting the feature extraction module and the classification and recognition module. The feature extraction module includes a temporal representation submodule (a one-dimensional convolutional neural network submodule), a global average pooling submodule, an encoder, and a projector. The classification and recognition module includes a fully connected layer, a batch normalization submodule, and a Softmax activation function. The anomaly detection and evaluation neural network architecture includes a second embedding layer, a Transformer-based anomaly detection module, and a projection layer. The second embedding layer includes a mask embedding submodule and a position encoding submodule. The Transformer-based anomaly detection module includes a first encoder layer, a second encoder layer, and a third encoder layer. The first encoder layer, the second encoder layer, and the third encoder layer each include a multi-head attention layer and a multilayer perceptron.

[0068] The classification and recognition neural network architecture aims to output the fetus's specific physiological state at that time, including fetal hypoxia, blood pH, blood sample status, and blood carbon dioxide status. Its input data is a preprocessed FHR-UC time series array. The preprocessed time series array is used to perform sine-cosine hybrid encoding on the extracted features through Gram angle field changes. The specific calculation process is as follows:

[0069]

[0070] in x i For time series data that has undergone normalization preprocessing, θ i For polar coordinates, G GASF ( i , j ) is the first of the Gram angles and the field matrix. i and j Matrix elements at time points, G GADF ( i , j ) is the first of the Gram angular difference field matrices. i and j The matrix elements at each time point.

[0071] The encoder extracts local features from the signal, performs downsampling, and outputs a high-dimensional feature representation, which helps the network better understand the complexity and diversity of the signal. The features output from the two branches are concatenated along the channel dimension to form a composite feature vector, which is then input into the projection layer. This maps the encoded features to a low-dimensional space, and the fused features are then mapped to the final classification result.

[0072] In one embodiment, S120 may include:

[0073] S1201. Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, feature extraction is performed through the feature extraction module of the classification and recognition neural network architecture in the dual neural network architecture to obtain comprehensive features;

[0074] S1202. Based on the comprehensive features, the comprehensive feature vector is obtained through the first embedding layer of the classification and identification neural network architecture in the dual neural network architecture. This vector is a learnable 64-dimensional embedding vector.

[0075] S1203. Based on the comprehensive feature vector, the classification recognition module of the classification recognition neural network architecture in the dual neural network architecture maps the comprehensive feature vector to the output category to realize the classification of fetal physiological status. Specifically, the comprehensive feature vector is mapped to the output category through the fully connected layer to realize the classification recognition of fetal status (such as normal, abnormal, etc.). In the fully connected layer, through the learning and adjustment of weights, the network can gradually learn how to judge the fetal status based on the input features and give the corresponding classification results.

[0076] Specifically, S1201 may include:

[0077] Based on fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, the temporal representation submodule of the feature extraction module jointly models the preprocessed dual-lead temporal signal using a fully connected layer and positional encoding. A fully connected network maps the input one-dimensional temporal signal to a high-dimensional latent space, specifically using a linear transformation matrix of dimensions [2, 128]. The input dimension 2 represents the dual-channel signal of FHR-UC, and the output dimension 128 represents the expanded feature representation space. This linear projection operation, through parameterized learning of the weight matrix, can adaptively capture multi-scale frequency components in the fetal heart rate monitoring signal, exhibiting stronger feature representation capabilities compared to a fixed convolutional kernel design.

[0078] Based on the features after multiple convolutions, the encoder of the feature extraction module encodes them separately to obtain multiple encoded features, thereby reducing the feature dimension while retaining key information.

[0079] Based on multiple encoded features, the projector of the feature extraction module maps the encoded features to a classification space that is easier to classify, and then concatenates the multiple mapped features along the channel dimension to obtain a comprehensive feature.

[0080] In another embodiment, LSTM (Long Short-Term Memory) can be used to replace the time series representation submodule to process data, directly process the preprocessed time series array, and the LSTM can be combined with an attention mechanism to calculate the prediction error.

[0081] In another embodiment, the self-attention mechanism in the Transformer architecture can be used to replace the temporal representation submodule to achieve feature extraction. Through a multi-layer Transformer encoder, high-level features are extracted step by step, and the self-attention mechanism can simultaneously consider the relationships between all points in the sequence to capture global dependencies.

[0082] The anomaly detection and evaluation neural network architecture aims to identify anomalous signals, including their location, severity, and possible anomaly types. The data embedding layer comprises token encoding and location encoding sub-modules, used to extract temporal features and location information from the signal, respectively. This layer transforms the input one-dimensional temporal signal into a high-dimensional embedding vector for subsequent network layers. The encoder layer consists of multiple encoder blocks, each containing an attention layer and a multilayer perceptron (MLP). The attention layer identifies anomalous points in the signal by calculating attention weights and correlation differences. The anomaly detection process first uses a sliding window to calculate the mean and standard deviation changes of the prediction error within each window. Based on factors such as the mean and standard deviation changes of the error, the number of errors exceeding a threshold, and the number of consecutive sequences, a minimum optimization function is used to find the anomaly detection threshold. Finally, error sequences exceeding the threshold are identified, and an 'anomaly filling' strategy is used to expand these sequences for more comprehensive anomaly detection. Consecutive or overlapping anomaly sequences are merged to form the final anomaly detection result. The MLP further processes the output of the attention layer to extract higher-level features. The final projection layer maps the output of the encoder layer to the desired output dimension, i.e., the result of the abnormal signal identification.

[0083] In one embodiment, S120 may include:

[0084] S1201' Based on the fetal heart rate monitoring data and corresponding monitoring result label data of the pregnant women, the temporal features and location information are obtained through the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture;

[0085] S1202' Based on temporal features and location information, the anomaly detection module based on the Transformer architecture in the anomaly detection and evaluation neural network architecture of the dual neural network architecture calculates attention weights and correlation differences to identify anomalies in temporal features and location information, thereby realizing the identification of abnormal fetal status in pregnant women.

[0086] Specifically, S1201' may include:

[0087] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, the temporal characteristics of the corresponding monitoring result labels are obtained through the mask embedding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture.

[0088] Based on the fetal heart rate monitoring data and corresponding monitoring result label data of pregnant women, the positional information of the corresponding monitoring result label is obtained through the positional encoding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the dual neural network architecture.

[0089] Specifically, S1202' may include:

[0090] Based on temporal features and location information, the error sequence is identified by using the multi-head attention layer in the anomaly detection module based on the Transformer architecture, combined with the sliding window method and minimization optimization function. The error sequence is then expanded by anomaly padding, and continuous or overlapping error sequences are merged to form an anomaly identification signal for the pregnant woman's fetal status.

[0091] The anomaly detection module based on the Transformer architecture uses a multilayer perceptron to extract features based on the output of the corresponding multi-head attention layer, thus obtaining the anomaly recognition signal features.

[0092] By using the projection layer in the anomaly detection module based on the Transformer architecture, the anomaly identification signal features are projected and mapped to the required output dimension, which is the anomaly identification result of the pregnant woman's fetal state.

[0093] Specifically, the error sequence identification process can be as follows.

[0094] First, the raw fetal heart rate monitoring time-series signal from the fetal heart rate monitoring data is preprocessed. Linear interpolation is used to fill in missing values ​​in the signal to ensure signal continuity. The calculation formula is as follows:

[0095]

[0096] in,( x 0, y 0) and ( x 1, y 1) Given two points, x For the values ​​that need to be interpolated, y The output value after interpolation;

[0097] Sample pruning removes invalid or redundant parts of the signal, improving data quality; normalization adjusts the amplitude of the signal to a uniform range, eliminating dimensional differences and facilitating subsequent processing.

[0098] A 1D CNN effectively extracts local features from preprocessed temporal signals, capturing patterns and trends. These features are then processed by global average pooling, an encoder, and a projector to further compress and refine the information, ultimately yielding the input to the embedding layer. This input is fed into a fully connected layer, which integrates and transforms the features. Finally, the softmax function outputs the monitoring result (e.g., a judgment of whether the fetus is hypoxic). The softmax function converts the output of the fully connected layer into a probability distribution, representing the likelihood of the fetus being in different states (e.g., normal or hypoxic).

[0099] Beyond classification and recognition, the process also includes an anomaly detection and evaluation neural network employing the Transformer architecture. The Transformer architecture comprises components such as an embedding layer, a multi-head attention layer, a multilayer perceptron, and an encoder. The embedding layer converts the input signal into a high-dimensional vector representation. The multi-head attention layer captures dependencies between different locations in the signal, increasing the attention to important features to improve computational efficiency. The multilayer perceptron performs a nonlinear transformation on the output of the attention layer. The encoder further integrates and refines features. Finally, the network outputs evaluation metrics through a projection layer to quantitatively assess the degree of abnormality in the fetal condition. To enhance the Transformer architecture's ability to identify abnormal patterns in fetal heart rate-uterine contraction signals, anomaly detection algorithms based on statistical methods and optimization techniques can be deeply integrated with the Transformer architecture. Specifically, before inputting the signal into the Transformer architecture, statistical analysis can be performed on the signal within each window to calculate its error, and further obtain statistical quantities such as the mean and variance of the error, as well as advanced statistical features such as skewness, waveform factor, peak factor, and impulse factor of the signal.

[0100] The algorithm finds a threshold that minimizes the cost function through optimization techniques, defining the threshold for anomaly detection. During this process, the algorithm dynamically adjusts the threshold using historical data and current signal characteristics to adapt to anomaly detection needs under different circumstances. The anomaly sequence identification and evaluation section is responsible for identifying error sequences exceeding the threshold, calculating the maximum error value for each anomaly sequence, and trimming possible false alarms based on the percentage change between maximum errors. An anomaly score is calculated for each trimmed anomaly sequence, representing its degree of anomaly. Error sequences exceeding the threshold are identified, and a certain "anomaly imputation" strategy is considered to expand these sequences to more accurately capture the full picture of the anomaly. The maximum error value for each anomaly sequence is calculated, and possible false alarms are trimmed based on the percentage change between maximum errors and features such as sequence length and frequency, improving the accuracy of anomaly detection. An anomaly score is calculated for each trimmed anomaly sequence, which comprehensively considers multiple factors such as error value, sequence features, and the model's confidence in the anomaly, providing a more comprehensive representation of its degree of anomaly. Furthermore, the algorithm provides doctors with early warnings and prompts regarding the trend of fetal status changes based on the distribution and changes in the anomaly score. Assuming there are... n The reference health status time series for each sampling point is as follows: X = [ x 1, x 2,…, x n The real-time monitoring time series is The formula for calculating the anomaly score of the monitoring data is as follows:

[0101]

[0102] Where Cov() calculates the covariance, and Var() calculates the variance. This represents the average value of the time series.

[0103] In the neural network architecture for anomaly detection and evaluation, an optimization technique is used to find a threshold that minimizes the cost function, which is then used to define the threshold for anomaly detection. This dynamic threshold adjustment strategy can automatically adjust the threshold according to the characteristics of different signals, improving the accuracy and robustness of anomaly detection.

[0104] An abnormality score is calculated for each pruned abnormal sequence, taking into account multiple factors to more comprehensively represent the degree of abnormality. Based on the distribution and changes in these abnormality scores, doctors are provided with early warnings and alerts regarding trends in fetal condition, helping to promptly identify and manage potential fetal health risks.

[0105] In one embodiment, the method for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture provided by the present invention may further include the following steps:

[0106] The fetal heart rate monitoring results are obtained by integrating the fetal physiological state classification results obtained through the classification recognition neural network architecture in the dual neural network architecture and the abnormal identification results of the pregnant woman's fetal state obtained through the abnormal monitoring and evaluation neural network architecture.

[0107] By integrating the outputs of the two networks, a complete AI-assisted diagnostic report on fetal heart rate monitoring is generated. The report can include information such as the fetus's real-time status, abnormal monitoring results, and abnormality assessment, providing doctors with comprehensive diagnostic information.

[0108] This embodiment uses the CTU-UHB public medical dataset for model training and validation. This dataset was jointly constructed by the Department of Biomedical Engineering at the Czech Technical University of Prague (CTU) and the Department of Obstetrics and Gynecology at Brno University Hospital (UHB), and contains complete monitoring records of 552 singleton pregnancies with a gestational age ≥36 weeks and no prior known developmental defects. Data collection strictly adhered to medical ethical guidelines. Each sample included simultaneously acquired fetal heart rate (FHR) and uterine contraction pressure (UC) dual-lead signals at a sampling frequency of 4Hz and signal durations ranging from 20 to 60 minutes. The data included 18 clinical indicators such as maternal age, gestational age, and parity. The training set (387 cases) and the test set (165 cases) were randomly divided in a 7:3 ratio to ensure no significant differences in the distribution of maternal age, gestational age, etc., between the datasets. The positional encoding employed a learnable sinusoidal positional encoding with a dimension of 128, and multi-scale feature interaction fusion was achieved through a 2-layer MLP Transformer encoder. The model performance on the test set (165 cases) is as follows:

[0109]

[0110] This invention provides a method for constructing a fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signals and a dual neural network architecture. It employs a time-aligned dual-signal sampling strategy, inputting parallel one-dimensional time-series data of fetal heart rate and uterine contractions from two leads. A dual-branch neural network architecture is used, with each branch independently processing the signal from one lead. This fully utilizes the complementary information between signals from different leads. The trained fetal heart rate monitoring auxiliary interpretation model can accurately identify abnormalities in fetal physiological state and maternal / fetal condition, significantly improving the accuracy of the model in recognizing complex conditions such as fetal hypoxia and uterine contraction compression, as well as time-dependent pathological states such as late deceleration and variable deceleration.

[0111] The present invention provides a method and system for constructing a fetal heart rate monitoring auxiliary interpretation model based on a one-dimensional time-series signal and a dual neural network architecture, which can bring at least the following beneficial effects:

[0112] (1) By inputting one-dimensional time-series data of fetal heart rate and uterine contraction in parallel, the problem of signal coupling interference in traditional single-channel monitoring is solved, which greatly improves the accuracy of the fetal heart rate monitoring auxiliary interpretation model in identifying complex conditions such as fetal hypoxia and uterine contraction compression.

[0113] (2) The time-aligned dual-signal sampling strategy is adopted to preserve the time correlation characteristics of the original monitoring waveform, which can improve the prediction accuracy of the model in time-dependent pathological states such as late deceleration and variable deceleration.

[0114] (3) A dual-branch neural network architecture is adopted, with each branch independently processing the signal of one lead, making full use of the complementary information between signals from different leads.

[0115] (4) Using a classification and recognition neural network architecture, feature extraction and dimensionality reduction are performed through a feature extraction module, a first embedding layer, and a classification and recognition module. The features extracted from each branch are spliced ​​together in the channel dimension to form a comprehensive feature vector, which improves the richness and accuracy of feature expression. The comprehensive feature vector is mapped to the output category through a fully connected layer to achieve accurate classification and recognition of the fetal physiological state.

[0116] (5) Using an anomaly monitoring and evaluation neural network architecture, the signal error is calculated and the error distribution is statistically analyzed through the second embedding layer, the anomaly monitoring module based on the Transformer architecture, and the projection layer. The anomaly detection threshold is found using optimization techniques, and the degree of anomaly of the signal is comprehensively evaluated. The error sequence is found, and the accurate identification of the abnormal state of the pregnant woman and the fetus in the fetal heart monitoring signal is realized.

[0117] The fetal heart rate monitoring auxiliary system provided by the present invention will be described below. The fetal heart rate monitoring auxiliary system described below can be referred to in correspondence with the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above.

[0118] The present invention provides a fetal heart rate monitoring auxiliary system, which may include:

[0119] The data receiving module is used to receive fetal heart rate monitoring data from at least one terminal of the pregnant woman to be tested.

[0120] The fetal heart rate monitoring auxiliary module is used to: obtain the fetal heart rate monitoring results of the pregnant woman based on the fetal heart rate monitoring data of the pregnant woman to be tested, through the fetal heart rate monitoring auxiliary interpretation model obtained by the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above.

[0121] The data output module is used to output the fetal heart rate monitoring results of the pregnant woman to at least one terminal.

[0122] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the following steps:

[0123] Receive fetal heart rate monitoring data from at least one terminal of the pregnant woman being tested;

[0124] Based on the fetal heart rate monitoring data of the pregnant woman to be tested, the fetal heart rate monitoring auxiliary interpretation model obtained by the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above is used to obtain the fetal heart rate monitoring results of the pregnant woman to be tested.

[0125] The fetal heart rate monitoring results of the pregnant woman to be tested are output to at least one terminal.

[0126] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps:

[0128] Receive fetal heart rate monitoring data from at least one terminal of the pregnant woman being tested;

[0129] Based on the fetal heart rate monitoring data of the pregnant woman to be tested, the fetal heart rate monitoring auxiliary interpretation model obtained by the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above is used to obtain the fetal heart rate monitoring results of the pregnant woman to be tested.

[0130] The fetal heart rate monitoring results of the pregnant woman to be tested are output to at least one terminal.

[0131] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0132] Receive fetal heart rate monitoring data from at least one terminal of the pregnant woman being tested;

[0133] Based on the fetal heart rate monitoring data of the pregnant woman to be tested, the fetal heart rate monitoring auxiliary interpretation model obtained by the construction method of the fetal heart rate monitoring auxiliary interpretation model based on one-dimensional time-series signal and dual neural network architecture described above is used to obtain the fetal heart rate monitoring results of the pregnant woman to be tested.

[0134] The fetal heart rate monitoring results of the pregnant woman to be tested are output to at least one terminal.

[0135] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a fetal heart monitoring aided interpretation model based on one-dimensional time series signals and a double neural network architecture, characterized in that, The method comprises: obtaining fetal heart monitoring data and corresponding monitoring result label data of a pregnant woman group, wherein the fetal heart monitoring data comprises one-dimensional time series data of fetal heart rate and one-dimensional time series data of uterine contraction, and the monitoring result label data comprises a fetal physiological state label and a pregnant woman fetal state abnormal interval label, the fetal physiological state comprises any one or any combination thereof of the following: fetal hypoxia state, blood pH state, blood sample state, and blood carbon dioxide state, and the pregnant woman fetal abnormal state comprises any one or any combination thereof of the following: fetal distress, fetal bradycardia, fetal tachycardia, and uterine contraction; learning the relationship between the fetal heart monitoring data and the corresponding monitoring result label data of the pregnant woman group by using a double neural network architecture, and training a fetal heart monitoring auxiliary interpretation model, wherein the double neural network architecture comprises a classification and identification neural network architecture and an abnormal monitoring and evaluation neural network architecture, the classification and identification neural network architecture comprises a feature extraction module, a classification and identification module, and a first embedding layer connecting the feature extraction module and the classification and identification module, the feature extraction module comprises a one-dimensional convolutional neural network sub-module, a global average pooling sub-module, an encoder, and a projector, the classification and identification module comprises a fully connected layer, a batch normalization sub-module, and a Softmax activation function, and the abnormal monitoring and evaluation neural network architecture comprises a second embedding layer, a Transformer architecture-based abnormal monitoring module, and a projection layer, the second embedding layer comprises a mask embedding sub-module and a position encoding sub-module, the Transformer architecture-based abnormal monitoring module comprises a first encoder layer, a second encoder layer, and a third encoder layer, and each of the first encoder layer, the second encoder layer, and the third encoder layer comprises a multi-head attention layer and a multi-layer perceptron. The method comprises: extracting features from the fetal heart monitoring data and the corresponding monitoring result label data of the pregnant woman group by using the feature extraction module of the classification and identification neural network architecture in the double neural network architecture, and obtaining comprehensive features; obtaining a comprehensive feature vector by using the first embedding layer of the classification and identification neural network architecture in the double neural network architecture according to the comprehensive features; mapping the comprehensive feature vector to an output category by using the classification and identification module of the classification and identification neural network architecture in the double neural network architecture according to the comprehensive feature vector, and realizing fetal physiological state classification.

2. The method of claim 1, wherein the method is characterized by: The method comprises: According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, through a one-dimensional convolutional neural network submodule of the feature extraction module, different sizes and steps of the convolution kernel are used to capture different frequency components and time sequence relationships in the fetal heart monitoring data and the corresponding monitoring result label data, to obtain a plurality of convolutional features; According to the plurality of convolutional features, the encoder of the feature extraction module is used for encoding respectively, to obtain a plurality of encoded features; According to the plurality of encoded features, the projector of the feature extraction module is used for mapping the encoded features to a classification space respectively, and the plurality of mapped features are spliced in the channel dimension to obtain a comprehensive feature.

3. The method of claim 2, wherein the method further comprises: According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, the relationship between the fetal heart monitoring data and the corresponding monitoring result label data is learned by using a double neural network architecture, and a fetal heart monitoring auxiliary interpretation model is trained, including: According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to obtain time sequence features and position information; According to the time sequence features and the position information, the abnormal monitoring module based on the Transformer architecture of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to calculate attention weights and correlation differences to identify abnormal points in the time sequence features and the position information, so as to realize the abnormal identification of the pregnant woman's fetal state.

4. The method of claim 3, wherein the method is characterized by: According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to obtain time sequence features and position information, including: According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, the mask embedding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to obtain time sequence features corresponding to the monitoring result label; According to the fetal heart monitoring data of the pregnant woman group and the corresponding monitoring result label data, the position encoding submodule of the second embedding layer of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to obtain position information corresponding to the monitoring result label; According to the time sequence features and the position information, the abnormal monitoring module based on the Transformer architecture of the abnormal monitoring and evaluation neural network architecture in the double neural network architecture is used to calculate attention weights and correlation differences to identify abnormal points in the time sequence features and the position information, so as to realize the abnormal identification of the pregnant woman's fetal state. According to the time sequence features and the position information, the multi-head attention layer in the abnormal monitoring module based on the Transformer architecture is used to identify error sequences by combining the sliding window method and the minimization optimization function, and to expand the error sequences by abnormal filling, and to merge continuous or overlapping error sequences to form an abnormal identification signal of the pregnant woman's fetal state; According to the output of the corresponding multi-head attention layer, the multi-layer perceptron in the abnormal monitoring module based on the Transformer architecture is used for feature extraction to obtain an abnormal identification signal feature; The abnormality recognition signal feature is projected to the mapping to the required output dimension, i.e., the abnormality recognition result of the fetus state of the pregnant woman, through a projection layer in the abnormality monitoring module based on the Transformer architecture.

5. The method of claim 4, wherein the method is characterized by: The relationship between the fetal heart monitoring data and the corresponding monitoring result label data is learned by using a double neural network architecture, and a fetal heart monitoring auxiliary interpretation model is trained, and the method further comprises: The fetal physiological state classification result obtained through the classification recognition neural network architecture in the double neural network architecture and the abnormality recognition result of the fetus state of the pregnant woman obtained through the abnormality monitoring and evaluation neural network architecture are integrated to obtain the fetal heart monitoring result.

6. A fetal heart monitoring assistance system characterized by, The method comprises: The data receiving module is configured to receive fetal heart monitoring data of a to-be-tested pregnant woman from at least one terminal; The fetal heart monitoring auxiliary module is configured to obtain a fetal heart monitoring result of the to-be-tested pregnant woman according to the fetal heart monitoring data of the to-be-tested pregnant woman by using the fetal heart monitoring auxiliary interpretation model obtained according to the construction method of the fetal heart monitoring auxiliary interpretation model based on a one-dimensional time sequence signal and a double neural network architecture according to any one of claims 1-5; The data output module is configured to output the fetal heart monitoring result of the to-be-tested pregnant woman to the at least one terminal.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the following steps when executing the program: The data receiving module is configured to receive fetal heart monitoring data of a to-be-tested pregnant woman from at least one terminal; The fetal heart monitoring auxiliary module is configured to obtain a fetal heart monitoring result of the to-be-tested pregnant woman according to the fetal heart monitoring data of the to-be-tested pregnant woman by using the fetal heart monitoring auxiliary interpretation model obtained according to the construction method of the fetal heart monitoring auxiliary interpretation model based on a one-dimensional time sequence signal and a double neural network architecture according to any one of claims 1-5; The data output module is configured to output the fetal heart monitoring result of the to-be-tested pregnant woman to the at least one terminal.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The processor implements the following steps when executing the program: The data receiving module is configured to receive fetal heart monitoring data of a to-be-tested pregnant woman from at least one terminal; The fetal heart monitoring auxiliary module is configured to obtain a fetal heart monitoring result of the to-be-tested pregnant woman according to the fetal heart monitoring data of the to-be-tested pregnant woman by using the fetal heart monitoring auxiliary interpretation model obtained according to the construction method of the fetal heart monitoring auxiliary interpretation model based on a one-dimensional time sequence signal and a double neural network architecture according to any one of claims 1-5; The data output module is configured to output the fetal heart monitoring result of the to-be-tested pregnant woman to the at least one terminal.

Images