A postpartum hemorrhage risk prediction method and system based on deep learning

Abstract

This invention discloses a method and system for predicting postpartum hemorrhage risk based on deep learning, belonging to the field of intelligent medical data analysis technology. The method includes the following steps: S1 acquiring raw clinical data of pregnant women; S2 preprocessing the raw clinical data; S3 performing feature engineering on the standardized data; S4 inputting the input feature set into the core module of a deep abstract network (DANet) for feature representation learning; S5 inputting the high-level semantic features into a hybrid expert model (MoE) fusion module for fusion processing; and S6 inputting the fused features into a prediction output module to output the postpartum hemorrhage risk prediction result. This invention improves data quality and feature effectiveness through data preprocessing and feature engineering, enhances model expressive power through DANet's adaptive feature selection and deep semantic extraction, improves prediction accuracy through MoE's fusion of multi-source information, and enhances model training stability through a phased training strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data analysis technology, specifically to a method and system for predicting postpartum hemorrhage risk based on deep learning. Background Technology

[0002] Postpartum hemorrhage (PPH) is a leading cause of maternal mortality. Current prediction methods mainly include traditional statistical analysis and single machine learning models. Traditional methods rely on manual feature selection, making it difficult to characterize the complex nonlinear relationships between multi-source clinical data. Existing machine learning methods (such as random forests and logistic regression) suffer from insufficient feature utilization, limited feature interaction mining capabilities, and poor adaptability to changes in data distribution when processing multi-source heterogeneous data (maternal physiological information, fetal examination data, and laboratory test indicators).

[0003] Furthermore, most models fail to effectively integrate the discrepancies in data from different sources, resulting in insufficient generalization ability and low predictive stability, making it difficult to meet the clinical demand for high-precision early warning. Therefore, it is necessary to provide a method and system for predicting postpartum hemorrhage risk that can effectively integrate multi-source clinical data and improve predictive accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting the risk of postpartum hemorrhage based on deep learning, so as to solve the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting postpartum hemorrhage risk based on deep learning, comprising the following steps: S1. Obtain raw clinical data of pregnant women, including demographic characteristics of pregnant women, fetal examination indicators and laboratory test indicators; S2. Preprocess the raw clinical data, including out-of-bounds detection and correction, outlier classification, unit error correction, consistency of expression, and data quality assessment, to obtain standardized data; S3. Perform feature engineering processing on the standardized data, including standardizing continuous variables, labeling ordered categorical variables, and selecting core features based on the random forest algorithm to construct an input feature set; S4. Input the input feature set into the core module of the deep abstract network DANet for feature representation learning to obtain high-level semantic features. The core module of DANet is composed of multiple stacked basic blocks. Each basic block includes a main path and a shortcut path, and feature fusion is achieved through element-level addition. S5. Input the high-level semantic features into the hybrid expert MoE fusion module for fusion processing. The MoE fusion module includes a gated network and multiple expert networks. The gated network dynamically weights and fuses the outputs of each expert network. S6. Input the fused features into the prediction output module and output the postpartum hemorrhage risk prediction result.

[0006] Preferably, the time window for collecting the original clinical data is from 36 weeks before delivery to the last valid record before delivery.

[0007] Preferably, the core features include fetal position, estimated fetal weight, umbilical artery resistance index, and blood routine related indicators.

[0008] Preferably, the abstraction layer in the DANet core module includes a feature selection mechanism based on learnable sparse masks.

[0009] Preferably, the abstraction layer performs weighted processing on features by introducing a fully connected network with an attention mechanism.

[0010] Preferably, the number of expert networks is three, corresponding to the pregnant woman's physiological information, fetal examination information, and laboratory test data, respectively.

[0011] Preferably, the gating network generates weight coefficients based on input features and performs weighted summation and fusion of the outputs of each expert network.

[0012] Preferably, the model training adopts a phased training strategy, including first training the gating network, and then jointly training the gating network and the expert network.

[0013] A deep learning-based postpartum hemorrhage risk prediction system includes a data preprocessing module, a feature engineering module, a deep abstract network (DANet) module, a hybrid expert (MoE) fusion module, and a prediction output module. The modules are sequentially connected and used to execute the method described in any one of claims 1 to 8.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: By employing a multi-dimensional preprocessing mechanism, the consistency and reliability of clinical data are improved, thereby enhancing data quality. Through feature engineering and the DANet structure, in-depth mining of complex feature relationships is achieved, enhancing feature expression capabilities. The MoE structure enables specialized modeling and dynamic fusion of different data types, improving multi-source data fusion capabilities. Furthermore, in a multi-source data environment, the model's generalization ability and prediction reliability are enhanced, resulting in improved prediction accuracy and stability. This can provide clinical support for postpartum hemorrhage risk warning. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 is a schematic diagram of the basic block structure of the DANet core module of the present invention; Figure 2 is a schematic diagram of the structure of the MoE fusion network of the present invention; Figure 3 illustrates the training process of the MoE network of this invention; Figure 4 is a schematic diagram of the ROC curve and AUC value of the model of the present invention; Figure 5 is a line graph showing the network performance of the present invention as a function of the number of network layers.

[0017] Figure 6 is a radar chart comparing the performance of the MoE network of this invention with existing mainstream models. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-6 Example 1: This example provides a method for predicting the risk of postpartum hemorrhage based on deep learning. The overall process includes data acquisition, data preprocessing, feature engineering, deep feature learning, feature fusion, and prediction output. S1 Data Acquisition: Acquire raw clinical data of pregnant women, including: demographic characteristics of pregnant women (such as age, gestational age, previous delivery history, etc.), fetal examination indicators (such as fetal position, estimated fetal weight, umbilical artery resistance index, etc.), and laboratory test indicators (such as blood routine related indicators, etc.). In this embodiment, the data collection time window is from 36 weeks before delivery to the last valid record before delivery.

[0020] S2 Data Preprocessing: Preprocessing the raw clinical data, specifically including: out-of-bounds detection and correction, outlier classification and processing, unit error correction, consistency of expression, and data quality assessment. The above processing results in a standardized dataset, which improves data consistency and usability.

[0021] S3 Feature Engineering Processing: Perform feature engineering processing on the standardized data, including: standardizing continuous variables, labeling ordered categorical variables, and evaluating and filtering features based on the random forest algorithm; The core features are then selected to construct an input feature set. These core features include fetal position, estimated fetal weight, umbilical artery resistance index, and blood routine related indicators.

[0022] S4 Deep Abstraction Network (DANet) Feature Learning: The input feature set is input into the Deep Abstraction Network (DANet) for feature representation learning to extract high-level semantic features; 1. Overall Structure: The DANet is composed of multiple basic blocks stacked together. Each basic block achieves progressive feature abstraction through layer-by-layer transmission. Each basic block includes: feature selection module, feature abstraction module, and feature fusion module.

[0023] 2. Feature selection process: In the feature selection stage, a learnable sparse mask is used to filter the input features. The mask parameters are normalized by a sparse mapping function to achieve adaptive grouping and selection of features. This method retains features that are highly relevant to the prediction task, suppresses irrelevant or redundant features, and improves the efficiency and stability of model training. In some implementations, the feature selection can be achieved by filtering the input feature vector element by element.

[0024] 3. Feature abstraction process: In the feature abstraction stage, the filtered features are processed by a fully connected network with an attention mechanism to generate an attention weight vector, and the features are weighted based on the attention weight. This process can: highlight key feature information, enhance the expressive power of important variables, and extract high-level semantic features. In some implementations, the feature abstraction process includes normalization processing and nonlinear activation processing.

[0025] 4. Feature Fusion Process: Each basic block includes a main path and a shortcut path: the main path is used for deep transformation of features; the shortcut path is used to retain input feature information; the outputs of the two paths are fused through element-wise addition, thereby achieving complementarity of feature information, improving the model's expressive power and alleviating the gradient vanishing problem.

[0026] 5. Inter-layer transfer: The basic blocks are connected layer by layer, and the output of the previous basic block is used as the input of the next basic block, so as to realize the progressive abstraction and enhancement of features.

[0027] S5 Hybrid Expert (MoE) Fusion Processing: The high-level semantic features output by DANet are input into the Hybrid Expert (MoE) fusion module for processing; the MoE fusion module includes: a gated network and multiple expert networks. In this embodiment, there are 3 expert networks, which correspond to: pregnant woman's physiological information, fetal examination information and laboratory test data, respectively. 1. Gating mechanism: The gating network generates weight coefficients based on input features to measure the importance of each expert network.

[0028] 2. Expert Fusion: The output results of each expert network are weighted and summed to achieve collaborative modeling of multi-source information and improve prediction accuracy.

[0029] S6 Model Training: During the model training phase, a phased training strategy is adopted, including: first, pre-training the gating network to enable it to have preliminary weight allocation capabilities, and then jointly training the gating network and each expert network to achieve the overall optimal performance.

[0030] S7 Model Training: Input the fused features into the prediction output module to output the postpartum hemorrhage risk prediction result; the prediction result can be the hemorrhage risk probability value and risk level classification result to assist clinical decision-making.

[0031] The specific application is as follows: Data preparation and preprocessing: Collect raw clinical data, including maternal demographic characteristics (age, blood type, etc.), fetal examination indicators (fetal position, estimated fetal weight, umbilical artery RI, etc.), and laboratory test indicators (blood routine related parameters, etc.). The data collection range is from 36 weeks to delivery. According to the data preprocessing process of this invention, outlier correction, abnormal value handling, unit correction, and standardized representation are performed on the data. After assessing the data quality, 1000 complete data points are selected from each set to construct an experimental dataset, which is then divided into training and test sets in an 8:2 ratio.

[0032] Feature engineering implementation: Feature encoding is performed on the preprocessed data, continuous variables are standardized, and ordered categorical variables are labeled; the importance of each feature is calculated using the random forest algorithm, and the core features are selected to form the model input feature set.

[0033] Model Training: Initialize the parameters of the DANet core module and the MoE fusion module, set the number of DANet base blocks to 64 layers, and the number of expert networks in the MoE fusion module to 3; use the Adam optimizer and cross-entropy loss function to train the model on the training set. The training process is divided into two stages: the first stage trains only the gating network (freezing the expert network parameters), and the second stage trains the gating network and the expert network together until the model converges.

[0034] Predictive Application: The preprocessed features of the sample to be predicted are input into the trained model. The DANet core module generates high-level semantic features, which are then weighted and fused by the MoE fusion module. The prediction output module then outputs the probability of postpartum hemorrhage risk and the classification result. Clinicians can develop targeted prevention and intervention plans based on the prediction results to reduce the risk of postpartum hemorrhage.

[0035] Model performance evaluation: Accuracy (ACC) and area under the ROC curve (AUC) are used as the core evaluation metrics, combined with precision, recall and F1 score for comprehensive evaluation.

[0036] In the description of this invention, unless otherwise expressly specified and limited, the terminology should be interpreted broadly. Unless otherwise expressly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for predicting the risk of postpartum hemorrhage, characterized in that, Includes the following steps: S1. Obtain raw clinical data of pregnant women, including demographic characteristics of pregnant women, fetal examination indicators and laboratory test indicators; S2. Preprocess the raw clinical data, including out-of-bounds detection and correction, outlier classification, unit error correction, consistency of expression, and data quality assessment, to obtain standardized data; S3. Perform feature engineering processing on the standardized data, including standardizing continuous variables, labeling ordered categorical variables, and selecting core features based on the random forest algorithm to construct an input feature set; S4. Input the input feature set into the core module of the deep abstract network DANet for feature representation learning to obtain high-level semantic features. The core module of DANet is composed of multiple stacked basic blocks. Each basic block includes a main path and a shortcut path, and feature fusion is achieved through element-level addition. S5. Input the high-level semantic features into the hybrid expert MoE fusion module for fusion processing. The MoE fusion module includes a gated network and multiple expert networks. The gated network dynamically weights and fuses the outputs of each expert network. S6. Input the fused features into the prediction output module and output the postpartum hemorrhage risk prediction result.

2. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The time window for collecting the original clinical data is from 36 weeks before delivery to the last valid record before delivery.

3. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The core features include fetal position, estimated fetal weight, umbilical artery resistance index, and blood routine related indicators.

4. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The abstraction layer in the DANet core module includes a feature selection mechanism based on learnable sparse masks.

5. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The abstraction layer performs weighted processing on features by introducing a fully connected network with an attention mechanism.

6. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The number of expert networks is three, corresponding to the pregnant woman's physiological information, fetal examination information, and laboratory test data, respectively.

7. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The gated network generates weight coefficients based on input features and performs weighted summation and fusion of the outputs of each expert network.

8. The method for predicting postpartum hemorrhage risk based on deep learning according to claim 1, characterized in that: The model training adopts a phased training strategy, which includes first training the gating network, and then jointly training the gating network and the expert network.

9. A deep learning-based postpartum hemorrhage risk prediction system, characterized in that: It includes a data preprocessing module, a feature engineering module, a deep abstract network (DANet) module, a hybrid expert (MoE) fusion module, and a prediction output module. Each module is connected in sequence and used to perform the method described in any one of claims 1 to 8.

Images