Method for constructing sepsis risk prediction model based on autoencoder and prediction system
Patent Information
- Application Number
- CN202611025934.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-09-25
AI Technical Summary
[0004]有鉴于此，本发明的目的在于提供一种脓毒症风险预测模型的构建方法及预测系统，以解决现有技术中因卷积神经网络难以有效捕捉临床数据时序依赖性、循环神经网络在混合学习时空特征时易过拟合，所导致的脓毒症预测模型时空特征提取不充分、预测性能受限的技术问题
[0035]1、发明针对背景技术指出的CNN模型难以捕捉时序趋势、RNN模型因混合学习时空特征而易过拟合的问题，创造性采用了“空间编码器(Spatial-E) + 时间编码器(LSTM-E)”的级联架构，空间编码器专职于学习同一时间点上多维度临床指标间的复杂关联（空间特征），时间编码器则专注于挖掘这些聚合特征在时间轴上的动态演变规律（时间特征）。这种分工明确的级联设计，有效避免了单一网络混合学习时发生的特征混淆与干扰，使得提取的深度时空特征向量，同时包含了更纯粹、更丰富的空间关联信息与时间演变信息，为后续精准分类奠定了坚实基础。这是对现有特征提取方法的实质性改进。
Smart Images
Figure CN122822344A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and medical information technology, specifically to a method for constructing a sepsis risk prediction model for assisting clinical decision-making and a corresponding prediction system. Background Technology
[0002] Sepsis is a clinical syndrome caused by a host dysregulation due to infection, leading to life-threatening organ dysfunction. It is an acute and critical illness with a high mortality rate. Its core manifestation is an overreaction of the body's immune system after infection, damaging its own tissues. Common symptoms include fever / hypothermia, rapid breathing, confusion, extreme pain, and cold, clammy skin, requiring immediate anti-infective and organ support treatment. Early and accurate prediction of sepsis in the intensive care unit (ICU) is crucial for improving patient prognosis.
[0003] Deep learning technology offers the possibility of extracting early warning signals from complex clinical time-series data (such as vital signs and laboratory indicators). However, existing methods have limitations: 1) Models such as Convolutional Neural Networks (CNNs) typically process single-time-point snapshots of data, making it difficult to effectively model the crucial continuous temporal dependencies and disease progression trends in clinical data; 2) Models such as Recurrent Neural Networks (RNNs) tend to learn by mixing spatial features (correspondence between indicators) and temporal features (evolutionary trends) during training. This mixed learning mode easily leads to model overfitting, and the extracted spatiotemporal features lack sufficient discriminative power and robustness. Therefore, a more robust and effective model building method is needed to separate and extract key spatiotemporal features from clinical time-series data in order to train a higher-performing predictive model. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide a method for constructing a sepsis risk prediction model and a prediction system, so as to solve the technical problems in the prior art where the sepsis prediction model has insufficient spatiotemporal feature extraction and limited prediction performance due to the difficulty of convolutional neural networks in effectively capturing the temporal dependence of clinical data and the tendency of recurrent neural networks to overfit when learning spatiotemporal features in a mixed manner.
[0005] The present invention provides a method for constructing a sepsis risk prediction model based on an autoencoder, comprising the following steps:
[0006] Model input data preparation: Acquire the patient's time-series clinical monitoring data, perform missing value imputation preprocessing on the data, and divide it according to a preset time window length ω, generating a dimension of... 3D training data tensors Where B is the batch size and M is the number of clinical features;
[0007] Spatiotemporal Feature Extraction Network Construction: A cascaded feature extraction network is constructed, comprising a spatial encoder and a temporal encoder; the spatial encoder is used to process the 3D training data tensor. Perform dimensionality reduction of spatial features and output spatial feature tensors. The time encoder is used to process the spatial feature tensor. Perform time-dimensional sequence modeling and output deep spatiotemporal feature vectors. ;
[0008] Unsupervised pre-training: Constructing a spatial decoder and a temporal decoder that are symmetric to the spatial encoder and temporal encoder, respectively; using the deep spatiotemporal feature vector... The input is sequentially processed by the temporal decoder and spatial decoder to reconstruct the spatial feature tensor. The corresponding first reconstruction tensor and the three-dimensional training data tensor The corresponding second reconstruction tensor By jointly minimizing the spatial feature tensor With the first reconstruction tensor The first reconstruction error between them, and the three-dimensional training data tensor With the second reconstruction tensor The second reconstruction error between the two is used to perform unsupervised training on the feature extraction network;
[0009] Supervised classifier training: Based on the feature extraction network trained in unsupervised training, the 3D training data tensor is processed. The corresponding deep spatiotemporal feature vectors are extracted, and a classifier is trained using these feature vectors and their corresponding sepsis labels.
[0010] Model integration step: Integrate the trained feature extraction network with the trained classifier to obtain the sepsis risk prediction model.
[0011] Furthermore, in the preparation of the model input data, the time-series clinical monitoring data includes vital sign data, laboratory test data, and demographic data.
[0012] Furthermore, the spatial encoder processes the three-dimensional training data tensor through the following steps. Perform dimensionality reduction on spatial features to output a spatial feature tensor. :
[0013] (1)
[0014] In the formula, Here is the weight matrix of the spatial encoder. This is the offset vector of the spatial encoder. This indicates that the Sigmoid activation function is used to perform a non-linear transformation on the data within the parentheses. The specific transformation method is as follows:
[0015] (2)
[0016] The temporal encoding includes a time-aware forget gate, an input gate, an update gate, and an output gate. The temporal encoding is applied to the spatial feature tensor. Perform time-dimensional sequence modeling to output deep spatiotemporal feature vectors. The processing steps include:
[0017] (3)
[0018] (4)
[0019] (5)
[0020] (6)
[0021] (7)
[0022] (8)
[0023] In the formula, The learnable clinical time decay factor. To introduce the time missing interval matrix, , These are the depth-spatiotemporal feature vectors output by the previous and current layer temporal encoders, respectively. , , and These are the activation functions for the forget gate, input gate, update gate, and output gate, respectively. , , , These are the output values of the forget gate, input gate, update gate, and output gate, respectively. , These represent the cell states of the current layer's time encoder at this moment and the previous moment, respectively. , , and These are the weight matrices for the forget gate, input gate, update gate, and output gate, respectively. , , and These are the offset vectors for the forget gate, input gate, update gate, and output gate, respectively. This indicates that the tanh activation function is used to perform a nonlinear transformation on the data inside the parentheses. The specific transformation method is shown in equation (9):
[0024] (9)
[0025] The depth-spatial feature vector of this layer's temporal encoder As input for the next step.
[0026] Furthermore, in the supervised classifier training, the classifier is an XGBoost model.
[0027] Furthermore, the time window length ω=6, and the number of features after dimensionality reduction. =10, the depth spatiotemporal feature vector Number of features =20.
[0028] This invention also discloses a sepsis risk prediction system, which includes:
[0029] The data preprocessing module receives raw clinical time-series data, performs missing value imputation and time window segmentation, and generates a three-dimensional data tensor that meets the model input requirements.
[0030] The feature extraction module integrates a feature extraction network trained by the above-mentioned method for constructing a sepsis risk prediction model based on an autoencoder. This network is used to process the three-dimensional data tensor output by the data preprocessing module and extract depth spatiotemporal feature vectors.
[0031] The risk prediction module integrates a classifier trained by the above-mentioned method for constructing a sepsis risk prediction model based on an autoencoder. It is used to receive the deep spatiotemporal feature vector output by the feature extraction module and output the sepsis risk probability value.
[0032] The result output module is used to output the risk probability value.
[0033] Furthermore, the result output module of the sepsis risk prediction system is also used to compare the risk probability value with a preset threshold and output corresponding prompt information based on the comparison result.
[0034] The beneficial effects of this invention are:
[0035] 1. This invention addresses the problems pointed out in the background art, namely, the difficulty of CNN models in capturing temporal trends and the tendency of RNN models to overfit due to the mixed learning of spatiotemporal features. It creatively employs a cascaded architecture of "Spatial Encoder (Spatial-E) + Temporal Encoder (LSTM-E)". The spatial encoder is dedicated to learning the complex relationships (spatial features) between multi-dimensional clinical indicators at the same time point, while the temporal encoder focuses on mining the dynamic evolution patterns of these aggregated features along the time axis (temporal features). This clearly defined cascaded design effectively avoids feature confusion and interference that occurs when a single network performs mixed learning, resulting in extracted deep spatiotemporal feature vectors that simultaneously contain purer and richer spatial correlation information and temporal evolution information, laying a solid foundation for subsequent accurate classification. This is a substantial improvement over existing feature extraction methods.
[0036] 2. This invention designs a two-stage reconstruction error training method based on a symmetric decoder, which not only requires the reconstruction of the original input ( It also requires reconstructing the spatial characteristics of the intermediate layer. This two-level reconstruction constraint is a powerful unsupervised learning driving signal that forces the encoder to learn hidden features. It must contain sufficient information to accurately reconstruct both the macroscopic raw data and the microscopic spatial features. This training strategy enables the model to automatically learn highly compressed and information-rich feature representations from massive amounts of unlabeled clinical data, enhancing the model's robustness to data noise and missing data, and significantly improving the discriminative power of features and the model's generalization ability. This is a creative optimization of the training objectives of standard autoencoders.
[0037] 3. This invention employs a two-stage paradigm of "unsupervised cascaded encoder pre-training + supervised XGBoost classifier training." This paradigm fully leverages the advantages of deep neural networks in complex feature extraction and the advantages of gradient boosting decision trees in structured data classification, namely efficiency, stability, and relatively good interpretability. Figure 3 As shown in the specific implementation, the model constructed using this method achieved an excellent AUC of 0.7548 on the PhysioNet 2019 challenge dataset. This directly verifies the effectiveness of the features extracted by this invention and the high prediction accuracy of the constructed model. This technical effect demonstrates that the solution of this invention can reliably identify early risk signals of sepsis from complex clinical data.
[0038] 4. The sepsis risk prediction system of this invention includes modules such as data preprocessing, feature extraction, risk calculation, and result output. The entire solution constitutes an end-to-end automated data processing and risk analysis tool. This system processes objective clinical data and outputs quantified risk probability values, providing medical staff with objective and timely auxiliary reference information, helping to shorten assessment time and improve decision-making efficiency. Attached Figure Description
[0039] Figure 1 This is a flowchart of the method of the present invention;
[0040] Figure 2 This is a structural diagram of the automatic encoder model of the present invention;
[0041] Figure 3 This is the ROC curve of the present invention on the PhysioNet / Computing in Cardiology Challenge 2019 dataset. Detailed Implementation
[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0043] Example 1: The sepsis risk prediction model construction method based on autoencoders in this example includes the following steps:
[0044] Model input data preparation: Obtain the patient's time-series clinical monitoring data, which includes vital sign data, laboratory test data, and demographic data. Specific data are shown in Table 1, totaling 39 data types, i.e., the number of features. =39. The data collection interval is 1 hour.
[0045]
[0046]
[0047]
[0048] The data underwent preprocessing with missing value imputation. First, a forward imputation method was used to fill in missing values in the collected clinical data, that is, using the patient's most recent measurement to fill in subsequent blank values. For indicators missing from the initial admission stage, the median of the batch of historical patient data was used as a baseline to ensure data integrity. (Batch size determined.) =256, time window length =6, converting clinical data into a three-dimensional training data tensor that includes temporal memory. , The dimension is ,Right now .
[0049] Spatiotemporal Feature Extraction Network Construction: A cascaded feature extraction network is constructed, comprising a spatial autoencoder and a temporal autoencoder. Each spatial encoder has... A hidden layer, in this embodiment =10, the spatial encoder is used to process the three-dimensional training data tensor. Perform dimensionality reduction of spatial dimensions. After passing through the spatial encoder, it is mapped to a spatial feature tensor. , The dimension becomes , The number of features after dimensionality reduction of the spatial encoder, i.e. :
[0050] (1)
[0051] In the formula, Here is the weight matrix of the spatial encoder. This is the offset vector of the spatial encoder. This indicates that the Sigmoid activation function is used to perform a non-linear transformation on the data within the parentheses. The specific transformation method is shown in the equation. As shown.
[0052] (2)
[0053] The results obtained from the above steps As input to each temporal encoder; the temporal encoder is used to process the spatial feature tensor. Perform time-dimensional sequence modeling and output deep spatiotemporal feature vectors. Each time encoder contains four gates: a time-aware forget gate, an input gate, an update gate, and an output gate. The feature extraction formulas for each gate of the time encoder are shown in equations (3) to (8):
[0054] (3)
[0055] (4)
[0056] (5)
[0057] (6)
[0058] (7)
[0059] (8)
[0060] In the formula, The learnable clinical time decay factor. To introduce a time-missing interval matrix, when a certain indicator has not been updated for a long time, Increased size leads to forgetting gate By reducing the weight of older data, the model will decrease the weight of those older data. , These are the hidden states of the time encoder in the previous and current layers, respectively. , , and These are the activation functions for the forget gate, input gate, update gate, and output gate, respectively. , , , These are the output values of the forget gate, input gate, update gate, and output gate, respectively. , These represent the cell states of the current layer's time encoder at this moment and the previous moment, respectively. , , and These are the weight matrices for the forget gate, input gate, update gate, and output gate, respectively. , , and These are the offset vectors for the forget gate, input gate, update gate, and output gate, respectively. This indicates that the tanh activation function is used to perform a nonlinear transformation on the data inside the parentheses. The specific transformation method is shown in equation (9):
[0061] (9)
[0062] The depth-spatial feature vector of this layer's temporal encoder As input for the next step, The dimension becomes ,Right now , =20 represents the number of features after the time encoder is upgraded to higher dimensions.
[0063] As attached Figure 2 As shown, the feature extraction network includes a spatial encoder and a temporal encoder. The external spatial encoder is responsible for spatial dimensionality reduction of the multidimensional clinical input at each independent time step, eliminating redundant noise between indicators. The nested temporal encoder receives the dimensionality-reduced spatial feature sequence and extracts spatiotemporal features containing the trend of disease deterioration. This cascaded design achieves deep decoupling and collaborative optimization of spatial concurrent features and temporal evolution features.
[0064] Unsupervised pre-training: Constructing a spatial decoder and a temporal decoder that are symmetric to the spatial encoder and temporal encoder, respectively; using the deep spatiotemporal feature vector... The input is sequentially processed by the temporal decoder and spatial decoder to reconstruct the spatial feature tensor. The corresponding first reconstruction tensor and the three-dimensional training data tensor The corresponding second reconstruction tensor By jointly minimizing the spatial feature tensor With the first reconstruction tensor The first reconstruction error between and the three-dimensional training data tensor With the second reconstruction tensor The second reconstruction error between The feature extraction network is then trained unsupervised. The second reconstructed tensor... The dimension becomes and Same dimensions ,Right now .
[0065] Supervised classifier training: Based on the feature extraction network trained in unsupervised training, the 3D training data tensor is processed. Processing is performed to extract the corresponding depth spatiotemporal feature vectors. The feature vector and its corresponding sepsis label are used to train a classifier. In this embodiment, the classifier is an XGBoost model. The deep spatiotemporal feature vector... Input XGBoost, and the classifier will automatically learn the nonlinear relationship between deep spatiotemporal features and sepsis infection. At this stage, only the parameters of the XGBoost classifier are trained, and the encoder network parameters are not updated.
[0066] Model integration step: Integrate the trained feature extraction network with the trained classifier to obtain the sepsis risk prediction model.
[0067] In practical applications, after undergoing the same preprocessing, the encoder extracts the deep spatiotemporal feature vector from the data of new patients. Then the XGBoost classifier classifies according to Calculate the risk probability value.
[0068] Example 2: A sepsis risk prediction system, comprising:
[0069] The data preprocessing module receives raw clinical time-series data, performs missing value imputation and time window segmentation, and generates a three-dimensional data tensor that meets the model input requirements.
[0070] The feature extraction module integrates a feature extraction network trained by the sepsis risk prediction model construction method described in Example 1, which is used to process the three-dimensional data tensor output by the data preprocessing module and extract the depth spatiotemporal feature vector.
[0071] The risk prediction module integrates a classifier trained by the sepsis risk prediction model construction method described in Example 1, which is used to receive the deep spatiotemporal feature vector output by the feature extraction module and output the sepsis risk probability value.
[0072] The result output module is used to output the risk probability value.
[0073] As an improvement to Embodiment 2, the result output module is also used to compare the risk probability value with a preset threshold and output corresponding prompt information based on the comparison result.
[0074] Appendix Figure 3 Experimental results show that the sepsis risk prediction model proposed in this invention achieves a global discriminative power (AUC) of 0.7548 on the PhysioNet / Computing in Cardiology Challenge 2019 dataset. This fully demonstrates that this invention can capture trend features hidden in clinical data, providing physicians with valuable sepsis prediction results.
[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for constructing a sepsis risk prediction model based on an autoencoder, characterized in that: Includes the following steps: Model input data preparation: Obtain the patient's time-series clinical monitoring data, perform missing value imputation preprocessing on the data, and divide it according to a preset time window length ω, generating a dimension of... 3D training data tensors Where B is the batch size and M is the number of clinical features; Spatiotemporal Feature Extraction Network Construction: A cascaded feature extraction network is constructed, comprising a spatial encoder and a temporal encoder; the spatial encoder is used to process the 3D training data tensor. Perform dimensionality reduction of spatial features and output spatial feature tensors. The time encoder is used to process the spatial feature tensor. Perform time-dimensional sequence modeling and output deep spatiotemporal feature vectors. ; Unsupervised pre-training: Constructing a spatial decoder and a temporal decoder that are symmetric to the spatial encoder and temporal encoder, respectively; using the deep spatiotemporal feature vector... The input is sequentially processed by the temporal decoder and spatial decoder to reconstruct the spatial feature tensor. The corresponding first reconstruction tensor and the three-dimensional training data tensor The corresponding second reconstruction tensor By jointly minimizing the spatial feature tensor With the first reconstruction tensor The first reconstruction error between them, and the three-dimensional training data tensor With the second reconstruction tensor The second reconstruction error between the two is used to perform unsupervised training on the feature extraction network; Supervised classifier training: Based on the feature extraction network trained in unsupervised training, the 3D training data tensor is processed. The corresponding deep spatiotemporal feature vectors are extracted, and a classifier is trained using these feature vectors and their corresponding sepsis labels. Model integration step: Integrate the trained feature extraction network with the trained classifier to obtain the sepsis risk prediction model.
2. The method for constructing a sepsis risk prediction model based on an autoencoder according to claim 1, characterized in that: In the preparation of the model input data, the time-series clinical monitoring data includes vital sign data, laboratory test data, and demographic data.
3. The method for constructing a sepsis risk prediction model based on an autoencoder according to claim 1, characterized in that: The spatial encoder processes the three-dimensional training data tensor through the following steps. Perform dimensionality reduction on spatial features to output a spatial feature tensor. : （1） In the formula, Here is the weight matrix of the spatial encoder. This is the offset vector of the spatial encoder. This indicates that the Sigmoid activation function is used to perform a non-linear transformation on the data within the parentheses. The specific transformation method is as follows: （2） The temporal encoding includes a time-aware forget gate, an input gate, an update gate, and an output gate. The temporal encoding is applied to the spatial feature tensor. Perform time-dimensional sequence modeling to output deep spatiotemporal feature vectors. The processing steps include: （3） （4） （5） （6） （7） （8） In the formula, The learnable clinical time decay factor. To introduce the time missing interval matrix, , These are the depth-spatiotemporal feature vectors output by the previous and current layer temporal encoders, respectively. , , and These are the activation functions for the forget gate, input gate, update gate, and output gate, respectively. , , , These are the output values of the forget gate, input gate, update gate, and output gate, respectively. , These represent the cell states of the current layer's time encoder at this moment and the previous moment, respectively. , , and These are the weight matrices for the forget gate, input gate, update gate, and output gate, respectively. , , and These are the offset vectors for the forget gate, input gate, update gate, and output gate, respectively. This indicates that the tanh activation function is used to perform a nonlinear transformation on the data inside the parentheses. The specific transformation method is shown in equation (9): （9） The depth-spatial feature vector of this layer's temporal encoder As input for the next step.
4. The method for constructing a sepsis risk prediction model based on an autoencoder according to claim 1, characterized in that: In the supervised classifier training, the classifier is an XGBoost model.
5. The method according to claim 3, characterized in that: The time window length ω=6, and the number of features after dimensionality reduction. =10, the depth spatiotemporal feature vector Number of features =20.
6. A sepsis risk prediction system, characterized in that, include: The data preprocessing module receives raw clinical time-series data, performs missing value imputation and time window segmentation, and generates a three-dimensional data tensor that meets the model input requirements. The feature extraction module integrates a feature extraction network trained by any one of the methods described in claims 1 to 5, which is used to process the three-dimensional data tensor output by the data preprocessing module and extract the depth spatiotemporal feature vector. The risk prediction module integrates a classifier trained by the method described in any one of claims 1 to 5, which is used to receive the deep spatiotemporal feature vector output by the feature extraction module and output the sepsis risk probability value. The result output module is used to output the risk probability value.
7. The sepsis risk prediction system according to claim 6, characterized in that: The result output module is also used to compare the risk probability value with a preset threshold and output corresponding prompt information based on the comparison result.