Fetal heart rate signal classification method based on space-time diagram convolutional network

Abstract

The invention provides a fetal heart rate signal classification method based on a space-time diagram convolutional network. The method comprises the following steps: acquiring and preprocessing a fetal heart rate signal; node feature attributes are extracted from the fetal heart rate signals, and an initial static graph is generated; constructing a dynamic graph based on the initial static graph and the self-adaptive adjacency matrix, updating information propagation of nodes in the dynamic graph on each time step, and generating an updated dynamic graph; and constructing an optimized space-time diagram convolutional network model, inputting the updated dynamic diagram into the optimized space-time diagram convolutional network model, and outputting a classification result of the fetal heart rate signals. According to the method, a dynamic graph mechanism is introduced, a traditional space-time graph convolutional network is optimized, space dependence and time dependence between nodes are effectively captured, cross-time-dimension information transmission is achieved, the pathological causality of FHR data is deduced, and the accuracy of FHR signal classification is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of modern medical signal processing, and in particular to a fetal heart rate signal classification method based on a spatiotemporal graph convolutional network. Background Art

[0002] Acute and chronic fetal distress, such as fetal hypoxia, acidosis, and abnormal fetal movement, is one of the most important factors affecting fetal health and life safety. It is particularly important to detect chronic fetal distress in a timely manner and adopt effective diagnosis and treatment plans. Clinically, cardiotocography (CTG) is widely used in prenatal and intrapartum diagnosis, among which the fetal heart rate (FHR) signal can provide valuable information about the health status of the fetus and is the only source of information directly available to clinicians. However, the current mainstream CTG interpretation method is mainly achieved by the subjective diagnosis of obstetricians and gynecologists, and there are inconsistencies between different doctors and the same doctor at different times, which affects the accuracy of CTG interpretation. Therefore, it is particularly important to carry out research on intelligent CTG classification models based on FHR signals to achieve accurate and timely diagnosis of chronic fetal distress.

[0003] Although the progress of machine learning technology has provided various means for intelligent auxiliary diagnosis of chronic fetal distress, limitations such as the inability to fully capture the spatiotemporal dependencies within and between time series have also hindered the further clinical practice of data-driven methods. In addition, compared with most univariate time series classification tasks, FHR signals, as typical medical data, themselves show greater temporal instability and complex and changeable pathological characteristics. Unlike time-invariant classification tasks, the hidden information of FHR in the diagnosis of chronic fetal distress is richer and more difficult to be mined. Therefore, a method that can accurately handle the spatiotemporal dependencies of FHR is needed.

[0004] In view of this, there is an urgent need to provide a classification algorithm for FHR signals to effectively process the complex spatiotemporal dependencies of FHR signals, thereby achieving a more intelligent and accurate pathological diagnosis of chronic fetal distress. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a fetal heart rate signal classification method based on a spatiotemporal graph convolutional network, design a dynamic graph mechanism, combine the static graph structure and the adaptive adjacency matrix, and use an improved spatiotemporal graph neural network (Spatial-Temporal Graph Neural Networks, STGNNs) model architecture to process the message propagation of dynamic graph data, so as to achieve effective classification based on FHR signals.

[0006] In order to solve the above problems, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a fetal heart rate signal classification method based on a spatiotemporal graph convolutional network, comprising the following steps:

[0008] Step S1, obtaining a fetal heart rate signal and performing preprocessing;

[0009] Step S2, extracting node feature attributes from the fetal heart rate signal to generate an initial static graph;

[0010] Step S3, constructing a dynamic graph based on the initial static graph and the adaptive adjacency matrix, updating the information propagation of nodes in the dynamic graph at each time step, and generating an updated dynamic graph;

[0011] Step S4, constructing an optimized spatiotemporal graph convolutional network model, inputting the updated dynamic graph into the optimized spatiotemporal graph convolutional network model, and outputting the classification result of the fetal heart rate signal;

[0012] The optimized spatio-temporal graph convolutional network model includes three serially connected ST-GCN (Spatio-TemporalGraph Convolutional Networks) modules, an average pooling layer, a fully connected layer and a Softmax activation function; each of the ST-GCN modules includes a graph convolutional network block OGCN, a temporal convolutional network block TCN and a residual block, and the residual block is used to realize the residual connection of the graph convolutional network block OGCN and the temporal convolutional network block TCN.

[0013] Furthermore, the graph convolutional network block OGCN includes a 3×3 channel-by-channel convolution, a 1×1 point-by-point convolution and an activation function connected in series.

[0014] Furthermore, the temporal convolutional network block TCN includes a convolutional layer and an activation function.

[0015] Further, step S3 includes the following steps:

[0016] S3.1, constructing node embedding matrices of source nodes and target nodes, performing sparse and normalized processing on the node embedding matrices, and obtaining an adaptive adjacency matrix;

[0017] S3.2, combining the initial static graph and the adaptive adjacency matrix to construct a dynamic graph;

[0018] S3.3, introduce diffusion graph convolution, update the node features of the current dynamic time series through the bidirectional random walk model, and obtain the updated dynamic graph.

[0019] In a second aspect, the present invention provides a fetal heart rate signal classification system for implementing the above method, comprising the following modules:

[0020] A data processing module, used for acquiring fetal heart rate signals and performing preprocessing;

[0021] A graph construction module, used to generate a static graph based on the preprocessed fetal heart rate signal, construct a dynamic graph based on the static graph and update the information propagation of nodes;

[0022] The fetal heart rate signal classification module inputs the updated dynamic graph into the optimized spatiotemporal graph convolutional network model to obtain the fetal heart rate signal classification result.

[0023] In a third aspect, the present invention provides an electronic device, comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the above method.

[0024] In a fourth aspect, the present invention provides a machine-readable storage medium storing machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the above method.

[0025] Compared with the prior art, the beneficial effects of the present invention are at least:

[0026] (1) The present invention constructs a graph structure learning module, structures multiple clinical feature index graphs to obtain a static graph structure, and automatically learns hidden spatiotemporal dependencies, which is suitable for time series without a fixed graph structure. This overcomes the defect that most GNNs methods in the prior art often rely on predefined graph structures.

[0027] (2) The present invention adds a dynamic graph mechanism and constructs a dynamic graph structure based on the constructed static graph. By reducing the dependence on the static structure of the graph and enhancing the learning of dynamic changes in nodes, the problem of randomness that may occur in daily CTG monitoring and the resulting error accumulation is solved.

[0028] (3) The present invention designs an optimized spatiotemporal graph convolutional network model to process this specific dynamic graph structure, effectively captures the spatial and temporal dependencies between nodes, realizes information transmission across time dimensions, infers the pathological causality of FHR data, and improves the accuracy of FHR signal classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is a flow chart of the method of the present invention.

[0030] Figure 2 Schematic diagram of the ST-GCN module structure. DETAILED DESCRIPTION

[0031] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0032] like Figure 1 As shown, this embodiment provides a fetal heart rate signal classification method based on spatiotemporal graph neural network, comprising the following steps:

[0033] S1: Building a graph structure learning module

[0034] First, the fetal heart rate (FHR) signal is obtained and preprocessed. For the FHR signal data, the FHR data of the i-th pregnant woman is marked as X. i , set the time step to s, the sliding window interval to w, and slice to get n = X of this group of data i / s node information. For each node information, there are 4 categories with a total of 25 physiological and pathological features that are structured to form node feature attributes α i =[a i1 ,a i2 ,...,a im ](m=25). The first 23 feature attributes are the consensus of clinical experts based on the practice and summary of authoritative guidelines such as the International Federation of Gynecology and Obstetrics (see Table 1 below), and the last two are important medical records of fetuses and pregnant women. These micro data are the gold standard for clinical heart rate variability analysis. Some of the features are obtained as shown in the following formula:

[0035] Low frequency normalization (LFnorm):

[0036]

[0037] Medium frequency normalization (MFnorm):

[0038]

[0039] High frequency normalization (HFnorm):

[0040]

[0041] Among them, VLFpower, LFpower, MFpower, and HFpower represent very low frequency power, low frequency power, medium frequency power, and high frequency power respectively, and Totalpower represents the total power.

[0042] Table 1

[0043]

[0044]

[0045] Therefore, we can get Xi The node attribute matrix A i =[α 1 ,α 2 ,...,α n ], this matrix will be regarded as the initial adjacency matrix of the i-th pregnant woman, and all elements are learnable parameters. Then, the value of the adjacency matrix is ​​optimized through graph structure learning training to obtain static graph data.

[0046] S2: Provide a dynamic graph mechanism method to solve the randomness problem in CTG monitoring.

[0047] Step S2 includes the following steps:

[0048] 1) Constructing a node embedding matrix. The present invention randomly initializes two nodes E with learnable parameters 1 ,E 2 ∈R N ×d , E 1 and E 2 denote the node embedding matrices of the source node and the target node respectively, and N and d denote the number of nodes and features respectively.

[0049] 2) Obtain the adaptive adjacency matrix. 1 ,E 2 The spatial weights obtained by point multiplication are sparsely processed by the Relu activation function to eliminate the weak connections between spatial dependencies; then Softmax is used for normalization to obtain the adaptive adjacency matrix A. adp .

[0050] 3) Get the dynamic graph structure. Combined with static graph data A s Together they form the dynamic graph structure A of each node g , the whole process can be expressed as follows:

[0051]

[0052] 4) Update the graph structure of the dynamic time series. Introduce diffusion graph convolution, simulate the information flow between nodes through the bidirectional random walk model, and update the node features of the current dynamic time series X. The modeling process of diffusion graph convolution is shown in the following formula:

[0053]

[0054] where θ 1 ,θ 2 ,θ 3 are all trainable parameters that change with different input samples; K represents the convolution kernel size; D in and D outis the degree matrix, the former represents the in-degree matrix, and the latter represents the out-degree matrix, which is calculated from the topological structure of the dynamic graph and is used to normalize the adaptive adjacency matrix; A is the original adjacency matrix; A d is the adjacency matrix obtained by Laplace transforming the original adjacency matrix, and T is the transpose.

[0055] The present invention captures the information propagation of nodes at each time step through forward and backward multi-step diffusion to obtain the spatial characteristics of nodes, thereby reducing the dependence on the static structure of the graph, enhancing the learning of the dynamic changes of nodes, and thus improving the performance of the model in graph data processing. Finally, the dynamic graph X is output. G .

[0056] S3: Establish a feature extraction model based on the optimized spatiotemporal graph convolutional network, update and process the message propagation of the dynamic graph, and realize the classification of fetal heart rate signals.

[0057] The optimized spatiotemporal graph convolutional network model proposed in this paper is a feature extraction network model with multiple optimized ST-GCN (Spatial Temporal Graph Convolutional Networks) structural units. It is derived from the basic unit of the classic STGNNs and is improved in structure to use dynamic graph X G As the input of the model; finally, the classification results of the FHR signal are output, including normal signals and disease signals, which are used to classify the health status of the fetus.

[0058] The optimized spatiotemporal graph convolutional network model includes three serially connected ST-GCN modules, an average pooling layer, a fully connected layer and a Softmax activation function.

[0059] like Figure 2 As shown, each of the ST-GCN modules includes a graph convolutional network block (Optimized Graph Convolutional Networks, OGCN), a temporal convolutional network block (Temporal Convolutional Network, TCN) and a residual block, and the OGCN block and the TCN block are residually connected through the residual block; the graph convolutional network block OGCN includes a layer of 3×3 channel-by-channel convolution, a layer of 1×1 point-by-point convolution and an activation function layer connected in series in sequence; the temporal convolutional network block TCN includes a layer of convolution layer and an activation function layer.

[0060] The OGCN block is designed to capture the spatial dependencies between different nodes, and the TCN block is used to simulate the temporal dynamics of FHR signals, so that the model can consider the evolution of these signals over time. Among them, the improved OGCN adopts a lightweight network structure. First, a channel-by-channel convolution with a convolution kernel of 3×3 is applied to extract the node information in each neighborhood, and then a 1×1 point-by-point convolution is cascaded for channel fusion. The features are further input into the TCN layer to refine and enhance the temporal dependency features of the time series.

[0061] Specifically, the node output of each convolution operation in the spatial domain can be expressed as follows:

[0062]

[0063] Where N(v ti )={v ti |d(v ti ,v tj )≤D} represents the node v ti The neighbor set of d(v ti ,v tj ) is the node v tj to v ti The minimum distance metric for any path of ti (v tj )={v tu |y ti (v tu )=y ti (v tj )} is the cardinality of the corresponding subset, y ti (v tj ) is the label mapping at the current moment; p(v ti ,v tj ) and w(v ti ,v tj ) are the sampling function and weight function respectively.

[0064] Furthermore, the optimized STGCNs model of the present invention diffuses the spatial graph convolutional network to the spatiotemporal domain, so that the neighborhood information also includes the time information of the node, thereby obtaining a new spatiotemporal neighbor set:

[0065]

[0066]

[0067] Where k represents the time dimension frame, kt defines the time neighbor range; D′ represents the node v tj to v ti The minimum distance; Η is the time range of the control neighborhood; y ST It is the label mapping in the spatiotemporal domain.

[0068] To further verify the effectiveness of the method of the present invention, 552 groups of CTG data from the CTG delivery database CTU-UHB jointly collected by the Czech Technical University and the Brno University Hospital were selected for comparative testing. The accuracy of pathological diagnosis was compared with four benchmark models, including the LS-SVM+GA, an auxiliary diagnosis algorithm for fetal distress based on genetic algorithm and least squares support vector machine, the DT-CTNet, an auxiliary diagnosis algorithm for fetal distress based on a two-layer AI model architecture, the Hybrid-FHR, an auxiliary diagnosis algorithm for fetal distress based on cross-modal feature fusion, the LSTM, an auxiliary diagnosis algorithm for fetal distress based on a long short-term memory network, and the ViT, an auxiliary diagnosis algorithm for fetal distress based on VisionTransformer. The accuracy (ACC), F1 value and precision (P) of the test set are shown in Table 2 below.

[0069] Table 2

[0070]

[0071]

[0072] It can be seen from Table 2 that the method of the present invention has obvious advantages.

[0073] The above embodiments are not limitations of the present invention, and the present invention is not limited to the above embodiments. As long as the requirements of the present invention are met, they belong to the protection scope of the present invention.

Claims

1. A fetal heart rate signal classification method based on spatiotemporal graph convolutional network, characterized in that: The following steps are involved: Step S1, obtaining a fetal heart rate signal and performing preprocessing; Step S2, extracting node feature attributes from the fetal heart rate signal to generate an initial static graph; Step S3, constructing a dynamic graph based on the initial static graph and the adaptive adjacency matrix, updating the information propagation of nodes in the dynamic graph at each time step, and generating an updated dynamic graph; Step S4: construct an optimized spatiotemporal graph convolutional network model, input the updated dynamic graph into the optimized spatiotemporal graph convolutional network model, and output the classification result of the fetal heart rate signal.

2. The fetal heart rate signal classification method based on spatiotemporal graph convolutional network according to claim 1, characterized in that: The optimized spatiotemporal graph convolutional network model includes three serially connected ST-GCN modules, an average pooling layer, a fully connected layer, and a Softmax activation function; Each of the ST-GCN modules includes a graph convolutional network block OGCN, a temporal convolutional network block TCN and a residual block, and the residual block is used to realize the residual connection between the graph convolutional network block OGCN and the temporal convolutional network block TCN.

3. The fetal heart rate signal classification method based on spatiotemporal graph convolutional network according to claim 1, characterized in that: The graph convolutional network block OGCN includes a 3×3 channel-by-channel convolution, a 1×1 point-by-point convolution and an activation function connected in series.

4. The fetal heart rate signal classification method based on spatiotemporal graph convolutional network according to claim 1, characterized in that: The temporal convolutional network block TCN includes a convolutional layer and an activation function.

5. The fetal heart rate signal classification method based on spatiotemporal graph convolutional network according to claim 3 or 4, characterized in that: The activation function adopts the ReLU activation function.

6. The fetal heart rate signal classification method based on spatiotemporal graph convolutional network according to claim 1, characterized in that: The step S3 comprises the following steps: S3.1, constructing node embedding matrices of source nodes and target nodes, performing sparse and normalized processing on the node embedding matrices, and obtaining an adaptive adjacency matrix; S3.2, combining the initial static graph and the adaptive adjacency matrix to construct a dynamic graph; S3.3, introduce diffusion graph convolution, update the node features of the current dynamic time series through the bidirectional random walk model, and obtain the updated dynamic graph.

7. A fetal heart rate signal classification system implementing the method according to any one of claims 1 to 6, characterized in that: Includes the following modules: A data processing module, used for acquiring fetal heart rate signals and performing preprocessing; A graph construction module is used to generate a static graph based on the preprocessed fetal heart rate signal, construct a dynamic graph based on the static graph, and update the information propagation of the nodes; The fetal heart rate signal classification module inputs the updated dynamic graph into the optimized spatiotemporal graph convolutional network model to obtain the fetal heart rate signal classification result.

8. An electronic device comprising a processor and a memory, characterized in that: The memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method according to any one of claims 1 to 6.

9. A machine-readable storage medium storing machine-executable instructions, characterized in that: When the machine executable instructions are called and executed by a processor, the machine executable instructions prompt the processor to implement the method as claimed in any one of claims 1 to 6.

Images