ECMO damage risk prediction method and system based on deep learning

Abstract

The invention discloses an ECMO injury risk prediction method and system based on deep learning. The method comprises the steps of multi-dimensional data acquisition, data processing enhancement, multi-dimensional injury risk prediction, injury treatment path prediction and ECMO injury risk prediction. The invention relates to the technical field of ECMO injury risk prediction, in particular to an ECMO injury risk prediction method and system based on deep learning, and the method employs a bidirectional prediction method combining multi-dimensional injury risk prediction and treatment path prediction, and comprises the steps: firstly predicting a specific injury type, risk and change trend; the direct availability and the result effectiveness of the ECMO injury risk prediction are improved by combining the recommended treatment mode, path and treatment effect prediction; a deep bidirectional long-short term neural network combining multi-modal fusion and multi-task learning is adopted to carry out multi-dimensional damage risk prediction; and performing injury treatment path prediction by adopting a feature-optimized treatment path prediction network, and constructing a treatment path full-name prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of ECMO injury risk prediction, and specifically refers to a method and system for predicting ECMO injury risk based on deep learning. Background Art

[0002] The method for predicting ECMO (extracorporeal membrane oxygenation) injury risk based on deep learning is a technology that uses a deep neural network model to analyze the possible complications and injury risks that patients may encounter during ECMO treatment. Through multi-level analysis of the patient's physiological data, historical medical records, and various indicators during the treatment process, the deep learning model can identify potential risk factors and predict whether the patient will suffer injuries during the treatment. This method can help doctors make more accurate decisions during ECMO treatment, improve the treatment effect, reduce the incidence of complications, and enhance the survival rate and recovery speed of patients. Through automated risk assessment, the deep learning model not only improves the diagnosis and treatment efficiency but also provides strong support for the formulation of personalized medical plans.

[0003] However, in the existing methods for predicting ECMO injury risk, there are technical problems as follows: the existing ECMO injury risk prediction involves the processing of complex and extensive multiple physiological data, experimental data, and pathological image data. Although the current technology provides many technical means for processing multi-modal data, the extraction of data information for ECMO injury prediction, as well as the overall prediction performance and result construction for ECMO injury risk subsequently, need to be improved. Although the current intelligent systems can process multi-modal data, they cannot provide effective assistance for the actual ECMO injury treatment based on the complex information in the multi-modal data; in the existing multi-dimensional injury risk prediction methods, there are technical problems that multi-dimensional injury risk prediction involves multi-task processing requirements and multi-modal data preconditions, which lead to poor interpretability of the results of multi-dimensional injury risk prediction. At the same time, the complex data scale also reduces the real-time and timeliness of the prediction; in the existing treatment path prediction methods, there is a technical problem that when the existing technology involves ECMO injury risk prediction, due to the lack of prediction support for the trend change of risks and subsequent injury treatment means, the results of ECMO injury risk prediction can only provide support to a certain extent, and cannot provide continuous support for the patient's condition change and subsequent condition. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the existing technology, the present invention provides a method and system for predicting ECMO damage risk based on deep learning. In the existing ECMO damage risk prediction methods, the existing ECMO damage risk prediction involves the processing of complex and miscellaneous physiological data, experimental data, and pathological image data. Although the current technology provides many technical means for multi-modal data processing, the extraction of data information for ECMO damage prediction and the overall prediction performance and result construction of subsequent ECMO damage risk still need to be improved. Although the current intelligent systems can process multi-modal data, they cannot provide effective help for the actual ECMO damage treatment based on the complex information in the multi-modal data. The present solution creatively adopts a two-way prediction method combining multi-dimensional damage risk prediction and treatment path prediction. By first predicting specific damage types, risks, and change trends, and combining the prediction of recommended treatment methods, approaches, and treatment effects, the direct usability and result effectiveness of ECMO damage risk prediction are improved, and a more feasible technical idea for the processing and use of multi-modal data is also provided. In the existing multi-dimensional damage risk prediction methods, the multi-dimensional damage risk prediction involves multi-task processing requirements and multi-modal data preconditions, which results in poor interpretability of the prediction results of multi-dimensional damage risk. At the same time, the complex data scale also reduces the real-time and timeliness of the prediction. The present solution creatively adopts a deep bidirectional long short-term neural network combining multi-modal fusion and multi-task learning for multi-dimensional damage risk prediction. By combining multi-modal fusion through a multi-task learning mechanism with shared parameters, the usability of feature information is improved, and the timeliness of the method is overall improved through the deep bidirectional long short-term neural network. At the same time, the combination of artificial features and machine features also improves the interpretability and referenceability of the final prediction results. In the existing treatment path prediction methods, when the existing technology involves ECMO damage risk prediction, due to the lack of prediction support for risk trend changes and subsequent damage treatment means, the results of ECMO damage risk prediction can only provide support to a certain extent, and cannot provide continuous support for the patient's condition changes and subsequent conditions. The present solution creatively adopts a treatment path prediction network with feature optimization for damage treatment path prediction. By constructing a full-name prediction model for treatment approaches, and combining the predictions in three directions: feature optimization, treatment approach optimization, and treatment effect prediction, a complete damage treatment path is constructed, providing a solid technical foundation for the result usability of ECMO damage risk prediction and the guarantee of the treatment process, and also providing practical experience for exploring the overall idea of ECMO damage risk prediction.

[0005] The technical solution adopted by the present invention is as follows: A method for predicting ECMO damage risk based on deep learning provided by the present invention includes the following steps:

[0006] Step S1: Multidimensional data collection;

[0007] Step S2: Data processing enhancement;

[0008] Step S3: Multidimensional injury risk prediction;

[0009] Step S4: Injury treatment path prediction;

[0010] Step S5: ECMO injury risk prediction.

[0011] Furthermore, in step S1, the multidimensional data collection is used to collect the original dataset required for ECMO injury risk prediction. Specifically, from the medical system and experimental examination data, through multidimensional data collection, the original data for injury risk prediction is obtained;

[0012] The original data for injury risk prediction specifically includes physiological data, laboratory examination data, imaging data, and clinical history data.

[0013] Furthermore, in step S2, the data processing enhancement is used to preprocess and enhance the original data. Specifically, basic data enhancement and initial feature extraction operations are performed on the original data for injury risk prediction to obtain optimized data for injury risk prediction, which specifically includes the following steps:

[0014] Step S21: Data cleaning, specifically removing the invalid data in the original data for injury risk prediction, and performing deletion and filling of missing values and outliers to obtain optimized data for injury risk cleaning;

[0015] Step S22: Data standardization, specifically performing statistical feature standardization on the optimized data for injury risk cleaning to obtain optimized data for injury risk standardization;

[0016] The statistical feature standardization specifically refers to Z-score standardization;

[0017] Step S23: Initial feature extraction, specifically performing artificial feature engineering design on the optimized data for injury risk standardization and performing initial feature extraction to obtain original feature data for injury risk prediction;

[0018] Step S24: Data content enhancement, specifically using interpolation and mean sample generation methods for data oversampling to balance the class balance of categorical data and obtain enhanced feature data for injury risk prediction;

[0019] Step S25: Data processing enhancement, specifically performing data processing enhancement through the data cleaning, the data standardization, the initial feature extraction, and the data content enhancement to obtain optimized data for injury risk prediction.

[0020] Further, in step S3, the multi-dimensional injury risk prediction is used to predict the specific type, risk level, and risk change of the injury risk. Specifically, based on the optimized data for injury risk prediction, a deep bidirectional long short-term neural network combining multi-modal fusion and multi-task learning is adopted to perform multi-dimensional injury risk prediction, and comprehensive assessment data of injury risk is obtained. The specific steps are as follows:

[0021] Step S31: Construct a multi-modal fusion network. Specifically, physiological data, laboratory examination data, and imaging data are fused. Specifically, the physiological data and laboratory examination data are fused through constructing a multi-layer perceptron to obtain mathematical and physical feature data, and the data features in the imaging data are extracted through a pre-trained convolutional neural network to obtain imaging feature data. The imaging feature data, the mathematical and physical feature data, and the original feature data in the optimized data for injury risk prediction are subjected to weighted multi-modal data fusion to obtain a fused feature data set;

[0022] Step S32: Construct a multi-task learning model. Specifically, a sub-model for predicting injury type, a sub-model for predicting injury risk level, and a sub-model for predicting the change trend of injury risk are constructed, and multi-task learning and prediction are carried out based on the fused feature data set;

[0023] Step S33: Construct a deep feedforward neural subnet for processing multi-modal data features. Specifically, a standard deep neural network is constructed to perform multi-modal data feature prediction to obtain deep feature data;

[0024] Step S34: Construct an improved bidirectional long short-term neural network. Specifically, a hierarchically improved bidirectional long short-term neural network is constructed to optimize the processing of time series data in multi-modal features, and subnet integration is carried out by combining the deep feature data to obtain a multi-task multi-modal prediction output;

[0025] Step S35: Training of the multi-dimensional injury risk prediction model. Specifically, through the constructed multi-modal fusion network, the constructed multi-task learning model, the constructed deep feedforward neural subnet, and the constructed improved bidirectional long short-term neural network, the multi-dimensional injury risk prediction model is trained to obtain the multi-dimensional injury risk prediction model;

[0026] Step S36: Multi-dimensional injury risk prediction. Specifically, using the multi-dimensional injury risk prediction model, based on the optimized data for injury risk prediction, multi-dimensional injury risk prediction is carried out to obtain comprehensive assessment data of injury risk;

[0027] The comprehensive assessment data of injury risk specifically includes injury risk type data, injury risk quantification score data, and injury time series risk change data.

[0028] Further, in step S4, the prediction of the injury treatment path is used to analyze personalized treatment plans and corresponding treatment effect predictions for the comprehensive prediction of risks, and to predict the injury treatment path. Specifically, based on the comprehensive injury risk assessment data and the injury risk prediction optimization data, a treatment path prediction network with feature optimization is used to predict the injury treatment path, and injury treatment path prediction data is obtained;

[0029] The treatment path prediction network with feature optimization specifically includes a feature optimization layer, a treatment path determination layer, and a treatment optimization prediction layer;

[0030] The steps of using the treatment path prediction network with feature optimization to predict the injury treatment path and obtain injury treatment path prediction data include:

[0031] Step S41: Construct a feature optimization layer, specifically, construct a standard convolutional long short-term neural network to extract spatial and temporal features from the comprehensive injury risk assessment data and the injury risk prediction optimization data to obtain high-dimensional feature data;

[0032] Step S42: Construct a treatment path determination layer, specifically, construct a multi-layer perceptron transformer model combined with a self-attention mechanism, and based on the high-dimensional feature data, perform treatment path prediction optimization to obtain treatment path determination feature data;

[0033] Step S43: Construct a treatment optimization prediction layer, specifically, combine the treatment path determination feature data to construct a graph neural generative adversarial network to perform detailed optimization of the treatment path and generate treatment plan data. The graph neural generative adversarial network specifically constructs a graph neural network in sequence to optimize the treatment implementation order, and generates optimal clinical treatment decision data through the generative adversarial network;

[0034] Step S44: Train the injury treatment path prediction model, specifically, through the constructed feature optimization layer, the constructed treatment path determination layer, and the constructed treatment optimization prediction layer, train the injury treatment path prediction model to obtain the injury treatment path prediction model;

[0035] Step S45: Predict the injury treatment path, specifically, use the injury treatment path prediction model, and based on the comprehensive injury risk assessment data and the injury risk prediction optimization data, predict the injury treatment path to obtain injury treatment path prediction data;

[0036] The injury treatment path prediction data specifically includes treatment plan prediction data, treatment order prediction data, and treatment effect prediction data.

[0037] Further, in step S5, the ECMO injury risk prediction is used to comprehensively predict the ECMO injury risk by combining the injury risk details and the injury treatment path. Specifically, it combines the injury risk comprehensive assessment data and the injury treatment path prediction data to predict the ECMO injury risk and obtain the ECMO injury risk integrated prediction data.

[0038] An ECMO injury risk prediction system based on deep learning provided by the present invention includes a multi-dimensional data acquisition module, a data processing enhancement module, a multi-dimensional injury risk prediction module, an injury treatment path prediction module, and an ECMO injury risk prediction module;

[0039] The multi-dimensional data acquisition module is used for multi-dimensional data acquisition. Through multi-dimensional data acquisition, the original injury risk prediction data is obtained and sent to the data processing enhancement module;

[0040] The data processing enhancement module is used for data processing enhancement. Through data processing enhancement, the optimized injury risk prediction data is obtained and sent to the multi-dimensional injury risk prediction module;

[0041] The multi-dimensional injury risk prediction module is used for multi-dimensional injury risk prediction. Through multi-dimensional injury risk prediction, the injury risk comprehensive assessment data is obtained and sent to the injury treatment path prediction module and the ECMO injury risk prediction module;

[0042] The injury treatment path prediction module is used for injury treatment path prediction. Through injury treatment path prediction, the injury treatment path prediction data is obtained and sent to the ECMO injury risk integrated prediction module;

[0043] The ECMO injury risk prediction module is used for ECMO injury risk prediction. Through ECMO injury risk prediction, the ECMO injury risk prediction data is obtained.

[0044] The beneficial effects achieved by the present invention by adopting the above scheme are as follows:

[0045] (1) In view of the technical problem that in the existing ECMO injury risk prediction methods, the existing ECMO injury risk prediction involves the processing of complex and extensive physiological data, experimental data, and pathological image data. Although the current technologies provide many technical means for multi-modal data processing, the extraction of data information for ECMO injury prediction and the overall prediction performance and result construction of ECMO injury risk subsequently need to be improved. Although the current intelligent systems can process multi-modal data, they cannot provide effective assistance for the actual ECMO injury treatment based on the complex information in multi-modal data, this solution creatively adopts a two-way prediction method combining multi-dimensional injury risk prediction and treatment path prediction. By first predicting specific injury types, risks, and change trends, and combining with the prediction of recommended treatment methods, approaches, and treatment effects, it improves the direct usability and result effectiveness of ECMO injury risk prediction, and also provides a more feasible technical idea for the processing and use of multi-modal data;

[0046] (2) In view of the technical problem that in the existing multi-dimensional injury risk prediction methods, multi-dimensional injury risk prediction involves multi-task processing requirements and multi-modal data preconditions, which results in poor interpretability of the results of multi-dimensional injury risk prediction. At the same time, the complex data scale also reduces the real-time and timeliness of prediction, this solution creatively adopts a deep bidirectional long short-term neural network combining multi-modal fusion and multi-task learning for multi-dimensional injury risk prediction. Through the multi-task learning mechanism of sharing parameters combined with multi-modal fusion, it improves the usability of feature information, and through the deep bidirectional long short-term neural network, it improves the timeliness of the method as a whole. At the same time, the combination of artificial features and machine features also improves the interpretability and referenceability of the final prediction results;

[0047] (3) In view of the technical problem that in the existing treatment path prediction methods, when the existing technology involves ECMO injury risk prediction, due to the lack of prediction support for risk trend changes and subsequent injury treatment means, the results of ECMO injury risk prediction can only provide support to a certain extent, and cannot provide continuous support for the patient's condition changes and subsequent conditions, this solution creatively adopts a treatment path prediction network with feature optimization for injury treatment path prediction. By constructing a full-name prediction model for treatment approaches, and combining the predictions in three directions of feature optimization, treatment approach optimization, and treatment effect prediction, it constructs a complete injury treatment path, providing a solid technical foundation for the result usability of ECMO injury risk prediction and the guarantee of treatment processes, and also providing practical experience for exploring the overall idea of ECMO injury risk prediction. Brief Description of the Drawings

[0048] Figure 1Schematic flowchart of a method for predicting ECMO damage risk based on deep learning provided by the present invention;

[0049] Figure 2 Schematic diagram of a system for predicting ECMO damage risk based on deep learning provided by the present invention;

[0050] Figure 3 Schematic flowchart of data processing enhancement in step S2;

[0051] Figure 4 Schematic flowchart of multi-dimensional damage risk prediction in step S3;

[0052] Figure 5 Schematic flowchart of damage treatment path prediction in step S4.

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. Detailed implementation manners

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention.

[0056] Example 1. Refer to Figure 1 , a method for predicting ECMO damage risk based on deep learning provided by the present invention, the method comprising the following steps:

[0057] Step S1: Multi-dimensional data collection;

[0058] Step S2: Data processing enhancement;

[0059] Step S3: Multi-dimensional damage risk prediction;

[0060] Step S4: Damage treatment path prediction;

[0061] Step S5: ECMO injury risk prediction.

[0062] By performing the above operations, in the existing ECMO injury risk prediction methods, there is a need to process complex and diverse physiological data, experimental data, and pathological imaging data. Although current technologies provide many technical means for multimodal data processing, the data information extraction for ECMO injury prediction, as well as the subsequent overall prediction performance and result construction for ECMO injury risk, need to be improved. Although the current intelligent system can process multimodal data, it cannot provide effective assistance for actual ECMO injury treatment based on the complex information in the multimodal data. This solution creatively adopts a two-way prediction method that combines multi-dimensional injury risk prediction and treatment path prediction. By first predicting the specific injury type, risk, and change trend, and combining it with the recommended treatment method, approach, and treatment effect prediction, it improves the direct usability and result effectiveness of ECMO injury risk prediction, and also provides a more feasible technical idea for the processing and use of multimodal data.

[0063] Example 2, see Figure 1 and Figure 2 In step S1, the multidimensional data acquisition is used to collect the original data set required for ECMO injury risk prediction, specifically, obtaining the original data for injury risk prediction from the medical system and laboratory examination data through multidimensional data acquisition;

[0064] The injury risk prediction raw data specifically includes physiological data, laboratory test data, imaging data and clinical history data.

[0065] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the data processing enhancement is used to preprocess and enhance the original data, specifically performing basic data enhancement and initial feature extraction operations on the damage risk prediction original data to obtain damage risk prediction optimized data. The specific steps include:

[0066] Step S21: data cleaning, specifically removing invalid data from the original damage risk prediction data, and deleting and filling missing values and outliers to obtain damage risk cleansing optimized data;

[0067] Step S22: data standardization, specifically, performing statistical feature standardization on the damage risk cleaning optimization data to obtain damage risk standardized optimization data;

[0068] The statistical feature standardization specifically refers to Z-score standardization;

[0069] Step S23: Initial feature extraction, specifically, artificial feature engineering design is performed on the damage risk standardized and optimized data, and initial feature extraction is carried out to obtain the original feature data for damage risk prediction;

[0070] Step S24: Data content enhancement, specifically, interpolation and mean sample generation methods are used for data oversampling to balance the class balance of class data, and enhanced damage risk prediction feature data is obtained;

[0071] Step S25: Data processing enhancement, specifically, through the data cleaning, the data standardization, the initial feature extraction, and the data content enhancement, data processing enhancement is carried out to obtain the optimized data for damage risk prediction.

[0072] Example 4, refer to Figure 1 、 Figure 2 and Figure 4 , this example is based on the above example. In step S3, the multi-dimensional damage risk prediction is used to predict the specific type, risk degree, and risk change of the damage risk. Specifically, based on the optimized data for damage risk prediction, a deep bidirectional long short-term neural network combining multi-modal fusion and multi-task learning is used for multi-dimensional damage risk prediction to obtain the comprehensive evaluation data of the damage risk, which specifically includes the following steps:

[0073] Step S31: Construct a multi-modal fusion network, specifically, physiological data, laboratory examination data, and imaging data are fused. Specifically, the physiological data and laboratory examination data are fused through constructing a multi-layer perceptron to obtain mathematical and physical feature data, and the data features in the imaging data are extracted through a pre-trained convolutional neural network to obtain imaging feature data. The imaging feature data, the mathematical and physical feature data, and the original feature data in the optimized data for damage risk prediction are subjected to weighted multi-modal data fusion to obtain a fused feature dataset;

[0074] The calculation formula of the fused feature dataset is:

[0075] ;

[0076] In the formula, X fusion is the fused feature dataset, is the weight of the multi-layer perceptron, X MLP is the mathematical and physical feature data, is the weight of the convolutional neural network, X CNN is the imaging feature data, is the weight of the original feature, X opt is the original feature data in the optimized data for damage risk prediction;

[0077] Step S32: constructing a multi-task learning model, specifically constructing a damage type prediction task sub-model, a damage risk degree prediction sub-model, and a damage risk change trend prediction sub-model, and performing multi-task learning and prediction based on the fused feature dataset;

[0078] Step S33: constructing a deep feedforward neural subnetwork for processing multimodal data features, specifically constructing a standard deep neural network to perform multimodal data feature prediction to obtain deep feature data;

[0079] Step S34: constructing an improved bidirectional long-term short-term neural network, specifically constructing a hierarchically improved bidirectional long-term short-term neural network for optimizing time series data processing in multimodal features, and performing subnetwork integration by combining the deep feature data to obtain a multi-task multimodal prediction output;

[0080] Step S35: multidimensional injury risk prediction model training, specifically, performing multidimensional injury risk prediction model training by constructing a multimodal fusion network, constructing a multitask learning model, constructing a deep feedforward neural subnetwork, and constructing an improved bidirectional long-short-term neural network to obtain a multidimensional injury risk prediction model;

[0081] Step S36: multi-dimensional injury risk prediction, specifically using the multi-dimensional injury risk prediction model and the injury risk prediction optimization data to perform multi-dimensional injury risk prediction and obtain comprehensive injury risk assessment data;

[0082] The comprehensive injury risk assessment data specifically includes injury risk type data, injury risk quantitative scoring data, and injury risk time series change data.

[0083] By performing the above operations, in view of the fact that in the existing multidimensional damage risk prediction methods, multidimensional damage risk prediction involves multi-task processing requirements and multi-modal data premises, which leads to poor interpretability of the results of multidimensional damage risk prediction. At the same time, the complex data scale also reduces the real-time and timeliness of the prediction. Technical problems, this scheme creatively adopts a deep bidirectional long-term and short-term neural network that combines multimodal fusion and multi-task learning to perform multidimensional damage risk prediction. The multi-task learning mechanism of shared parameters is combined with multimodal fusion to improve the availability of feature information, and the deep bidirectional long-term and short-term neural network is used to improve the timeliness of the method as a whole. At the same time, the fusion of artificial features and machine features also improves the interpretability and referenceability of the final prediction results.

[0084] Example 5, see Figure 1 , Figure 2 and Figure 5, in step S4, the prediction of the injury treatment path is used to analyze personalized treatment plans and corresponding treatment effect predictions for the comprehensive prediction of risks, and predict the injury treatment path. Specifically, based on the comprehensive injury risk assessment data and the injury risk prediction optimization data, a treatment path prediction network with feature optimization is used to predict the injury treatment path, and injury treatment path prediction data is obtained;

[0085] The treatment path prediction network with feature optimization specifically includes a feature optimization layer, a treatment path determination layer, and a treatment optimization prediction layer;

[0086] The step of using the treatment path prediction network with feature optimization to predict the injury treatment path and obtain injury treatment path prediction data includes:

[0087] Step S41: Construct a feature optimization layer, specifically by constructing a standard convolutional long short-term neural network to extract spatial and temporal features from the comprehensive injury risk assessment data and the injury risk prediction optimization data to obtain high-dimensional feature data. The calculation formula is:

[0088] ;

[0089] In the formula, X highdim is the high-dimensional feature data, ConvLSTM(·) is the standard convolutional long short-term neural network function, X risk is the comprehensive injury risk assessment data, and X opt is the original feature data in the injury risk prediction optimization data;

[0090] Step S42: Construct a treatment path determination layer, specifically by constructing a multi-layer perceptron transformer model combined with a self-attention mechanism, and based on the high-dimensional feature data, perform treatment path prediction optimization to obtain treatment path determination feature data. The calculation formula is:

[0091] ;

[0092] In the formula, X treatmentpath is the treatment path determination feature data, Transformer selfattention (·) is the multi-layer perceptron transformer model function combined with a self-attention mechanism, and X highdim is the high-dimensional feature data;

[0093] The multi-layer perceptron transformer model combined with a self-attention mechanism specifically improves the multi-layer perceptron transformer model by introducing a cross-modal attention mechanism to dynamically learn the interaction relationship between different modalities, and performs real-time learning performance for treatment path determination by introducing a dynamic path optimization method. The calculation formula is:

[0094] ;

[0095] Wherein, Y final is the output of the multi-layer perceptron transformer model combined with the self-attention mechanism, N is the total number of paths for predicting the damage treatment path, i is the path index, sigmoid(·) is the sigmoid activation function, W i is the path learning weight, X cross is the cross-modal fusion feature data, which is used to represent the damage risk prediction optimization data, Z i is the path complexity factor, CrossAttention(·) is the cross-modal attention mechanism function, X highdim is the high-dimensional feature data, X treatmentpath is the treatment approach determination feature data;

[0096] Step S43: Construct a treatment optimization prediction layer, specifically, combine the treatment approach determination feature data to construct a graph neural generative adversarial network for detailed optimization of the treatment path and generation of treatment plans. The graph neural generative adversarial network specifically constructs a graph neural network in sequence to optimize the treatment implementation order, and generates optimal clinical treatment decision data through the generative adversarial network. The calculation formula is:

[0097] ;

[0098] Wherein, Y opt is the optimal clinical treatment decision data, GAN(·) is the generative adversarial network function, GNN(·) is the graph neural network function, X treatmentpath is the treatment approach determination feature data;

[0099] Step S44: Train the damage treatment path prediction model, specifically, train the damage treatment path prediction model through the constructed feature optimization layer, the constructed treatment approach determination layer, and the constructed treatment optimization prediction layer to obtain the damage treatment path prediction model;

[0100] Step S45: Predict the damage treatment path, specifically, use the damage treatment path prediction model to predict the damage treatment path based on the comprehensive damage risk assessment data and the damage risk prediction optimization data to obtain the damage treatment path prediction data;

[0101] The damage treatment path prediction data specifically includes treatment plan prediction data, treatment order prediction data, and treatment effect prediction data.

[0102] By performing the above operations, in the existing treatment path prediction methods, when the prior art involves the prediction of ECMO injury risk, due to the lack of prediction support for the risk trend change and subsequent injury treatment means, the result of the ECMO injury risk prediction can only provide support to a certain extent, and cannot provide continuous support for the patient's condition change and subsequent condition. For this technical problem, this solution creatively adopts a treatment path prediction network with feature optimization to predict the injury treatment path. By constructing a full-name prediction model for the treatment path, and combining the predictions in three directions: feature optimization, treatment path optimization, and treatment effect prediction, a complete injury treatment path is constructed, providing a solid technical foundation for the usability of the ECMO injury risk prediction result and the guarantee of the treatment process, and also providing practical experience for exploring the overall idea of ECMO injury risk prediction.

[0103] Example Six, refer to Figure 1 and Figure 2 , based on the above example, in step S5, the ECMO injury risk prediction is used to comprehensively predict the ECMO injury risk by combining the injury risk details and the injury treatment path. Specifically, it combines the injury risk comprehensive assessment data and the injury treatment path prediction data to perform the ECMO injury risk prediction and obtain the ECMO injury risk integrated prediction data.

[0104] Example Seven, refer to Figure 1 and Figure 2 , based on the above example, a deep learning-based ECMO injury risk prediction system provided by the present invention includes a multi-dimensional data acquisition module, a data processing enhancement module, a multi-dimensional injury risk prediction module, an injury treatment path prediction module, and an ECMO injury risk prediction module;

[0105] The multi-dimensional data acquisition module is used for multi-dimensional data acquisition. Through multi-dimensional data acquisition, the original data for injury risk prediction is obtained, and the original data for injury risk prediction is sent to the data processing enhancement module;

[0106] The data processing enhancement module is used for data processing enhancement. Through data processing enhancement, the optimized data for injury risk prediction is obtained, and the optimized data for injury risk prediction is sent to the multi-dimensional injury risk prediction module;

[0107] The multi-dimensional injury risk prediction module is used for multi-dimensional injury risk prediction. Through multi-dimensional injury risk prediction, the comprehensive assessment data of injury risk is obtained, and the comprehensive assessment data of injury risk is sent to the injury treatment path prediction module and the ECMO injury risk prediction module;

[0108] The injury treatment path prediction module is used for predicting the injury treatment path. Through the prediction of the injury treatment path, injury treatment path prediction data is obtained, and the injury treatment path prediction data is sent to the ECMO injury risk integrated prediction module;

[0109] The ECMO injury risk prediction module is used for predicting the ECMO injury risk. Through the prediction of the ECMO injury risk, ECMO injury risk prediction data is obtained.

[0110] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0111] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention.

[0112] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. All in all, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of ECMO injury based on deep learning, characterized in that: The method comprises the following steps: Step S1: multi-dimensional data collection, obtaining original data for damage risk prediction through multi-dimensional data collection; Step S2: data processing enhancement, performing basic data enhancement and initial feature extraction operations on the damage risk prediction original data to obtain damage risk prediction optimized data; Step S3: Multidimensional damage risk prediction, based on the damage risk prediction optimization data, uses a deep bidirectional long-term and short-term neural network that combines multimodal fusion and multi-task learning to perform multidimensional damage risk prediction and obtain comprehensive damage risk assessment data, specifically including the following steps: Step S31: constructing a multimodal fusion network; Step S32: constructing a multi-task learning model; Step S33: constructing a deep feedforward neural subnetwork; Step S34: constructing an improved bidirectional long-term and short-term neural network; Step S35: training a multidimensional damage risk prediction model; Step S36: multidimensional damage risk prediction; Step S4: Injury treatment pathway prediction, specifically, based on the comprehensive injury risk assessment data and the injury risk prediction optimization data, a feature-optimized treatment pathway prediction network is used to predict the injury treatment pathway to obtain injury treatment pathway prediction data, including the following steps: Step S41: Constructing a feature optimization layer; Step S42: Constructing a treatment pathway determination layer; Step S43: Constructing a treatment optimization prediction layer; Step S44: Training an injury treatment pathway prediction model; Step S45: Predicting the injury treatment pathway; Step S5: ECMO injury risk prediction, obtaining integrated ECMO injury risk prediction data.

2. The method for predicting ECMO injury risk based on deep learning according to claim 1, wherein: In step S1, the multidimensional data acquisition is used to collect the original data set required for ECMO injury risk prediction, specifically, obtaining the original data for injury risk prediction from the medical system and laboratory examination data through multidimensional data acquisition; The injury risk prediction raw data specifically includes physiological data, laboratory test data, imaging data and clinical history data.

3. The method for predicting ECMO damage risk based on deep learning according to claim 2, wherein: In step S2, the data processing enhancement is used to preprocess and enhance the original data, specifically performing basic data enhancement and initial feature extraction operations on the damage risk prediction original data to obtain damage risk prediction optimized data, which specifically includes the following steps: Step S21: data cleaning, specifically removing invalid data from the original damage risk prediction data, and deleting and filling missing values and outliers to obtain damage risk cleansing optimized data; Step S22: data standardization, specifically, performing statistical feature standardization on the damage risk cleaning optimization data to obtain damage risk standardized optimization data; The statistical feature standardization specifically refers to Z-score standardization; Step S23: initial feature extraction, specifically performing artificial feature engineering design on the damage risk standardized optimization data and performing initial feature extraction to obtain original feature data for damage risk prediction; Step S24: data content enhancement, specifically, using interpolation and mean sample generation methods to perform data oversampling to balance the category balance of category data and obtain enhanced injury risk prediction feature data; Step S25: Data processing enhancement, specifically, through the data cleaning, the data standardization, the initial feature extraction, and the data content enhancement, perform data processing enhancement to obtain optimized data for injury risk prediction.

4. The ECMO injury risk prediction method based on deep learning according to claim 3, characterized in that: In step S3, the multi-dimensional injury risk prediction is used to predict the specific type, risk level, and risk change of the injury risk. Specifically, based on the optimized data for injury risk prediction, a deep bidirectional long short-term neural network combining multi-modal fusion and multi-task learning is adopted to perform multi-dimensional injury risk prediction to obtain comprehensive injury risk assessment data, which specifically includes the following steps: Step S31: Construct a multi-modal fusion network. Specifically, fuse physiological data, laboratory examination data, and imaging data. Specifically, fuse the physiological data and laboratory examination data through constructing a multi-layer perceptron to obtain mathematical and physical feature data, and extract the data features in the imaging data through a pre-trained convolutional neural network to obtain imaging feature data. Perform weighted multi-modal data fusion on the imaging feature data, the mathematical and physical feature data, and the original feature data in the optimized data for injury risk prediction to obtain a fused feature data set; Step S32: Construct a multi-task learning model. Specifically, construct a sub-model for injury type prediction, a sub-model for injury risk level prediction, and a sub-model for injury risk change trend prediction, and perform multi-task learning and prediction based on the fused feature data set; Step S33: Construct a deep feedforward neural subnet for processing multi-modal data features. Specifically, construct a standard deep neural network to perform multi-modal data feature prediction to obtain deep feature data; Step S34: Construct an improved bidirectional long short-term neural network. Specifically, construct a hierarchically improved bidirectional long short-term neural network for optimizing the processing of time series data in multi-modal features, and perform subnet integration by combining the deep feature data to obtain a multi-task multi-modal prediction output; Step S35: Training of the multi-dimensional injury risk prediction model. Specifically, through the constructed multi-modal fusion network, the constructed multi-task learning model, the constructed deep feedforward neural subnet, and the constructed improved bidirectional long short-term neural network, perform training of the multi-dimensional injury risk prediction model to obtain a multi-dimensional injury risk prediction model; Step S36: Multi-dimensional injury risk prediction. Specifically, use the multi-dimensional injury risk prediction model to perform multi-dimensional injury risk prediction based on the optimized data for injury risk prediction to obtain comprehensive injury risk assessment data.

5. The method for predicting ECMO damage risk based on deep learning according to claim 4, characterized in that: The comprehensive injury risk assessment data specifically includes injury risk type data, injury risk quantification score data, and injury time series risk change data.

6. The method for predicting ECMO injury risk based on deep learning according to claim 5, characterized in that: In step S4, the injury treatment path prediction is used to analyze personalized treatment plans and corresponding treatment effect predictions for the comprehensive prediction situation of the risk, and predict the injury treatment path. Specifically, based on the comprehensive injury risk assessment data and the optimized data for injury risk prediction, adopt a treatment path prediction network with optimized features to perform injury treatment path prediction to obtain injury treatment path prediction data; The above-mentioned treatment path prediction network with optimized features specifically includes a feature optimization layer, a treatment path determination layer, and a treatment optimization prediction layer; The steps of using the treatment path prediction network with optimized features to predict the injury treatment path and obtain the injury treatment path prediction data include: Step S41: Construct a feature optimization layer, specifically by constructing a standard convolutional long short-term neural network to extract spatial and temporal features from the comprehensive injury risk assessment data and the injury risk prediction optimization data to obtain high-dimensional feature data; Step S42: Construct a treatment path determination layer, specifically by constructing a multi-layer perceptron transformer model combined with a self-attention mechanism, and based on the high-dimensional feature data, perform treatment path prediction optimization to obtain treatment path determination feature data; Step S43: Construct a treatment optimization prediction layer, specifically by combining the treatment path determination feature data, construct a graph neural generative adversarial network to perform detailed optimization of the treatment path and generate treatment plan data. The graph neural generative adversarial network specifically constructs a graph neural network in sequence to optimize the treatment implementation order and generates optimal clinical treatment decision data through a generative adversarial network; Step S44: Training the injury treatment path prediction model, specifically by using the constructed feature optimization layer, the constructed treatment path determination layer, and the constructed treatment optimization prediction layer to train the injury treatment path prediction model to obtain the injury treatment path prediction model; Step S45: Injury treatment path prediction, specifically by using the injury treatment path prediction model, based on the comprehensive injury risk assessment data and the injury risk prediction optimization data, perform injury treatment path prediction to obtain injury treatment path prediction data; The injury treatment path prediction data specifically includes treatment plan prediction data, treatment order prediction data, and treatment effect prediction data.

7. The method for predicting ECMO injury risk based on deep learning according to claim 6, characterized in that: In step S5, the ECMO injury risk prediction is used to perform comprehensive ECMO injury risk prediction by combining the injury risk details and the injury treatment path. Specifically, by combining the comprehensive injury risk assessment data and the injury treatment path prediction data, perform ECMO injury risk prediction to obtain ECMO injury risk integrated prediction data.

8. A deep learning-based ECMO damage risk prediction system for implementing a deep learning-based ECMO damage risk prediction method according to any one of claims 1-7, characterized in that: It includes a multi-dimensional data acquisition module, a data processing enhancement module, a multi-dimensional injury risk prediction module, an injury treatment path prediction module, and an ECMO injury risk prediction module.

9. The deep learning-based ECMO injury risk prediction system according to claim 8, characterized in that: The multi-dimensional data acquisition module is used for multi-dimensional data acquisition. Through multi-dimensional data acquisition, the original injury risk prediction data is obtained and sent to the data processing enhancement module; The data processing enhancement module is used for data processing enhancement. Through data processing enhancement, the injury risk prediction optimization data is obtained and sent to the multi-dimensional injury risk prediction module; The multi-dimensional injury risk prediction module is used for multi-dimensional injury risk prediction. Through multi-dimensional injury risk prediction, the comprehensive injury risk assessment data is obtained and sent to the injury treatment path prediction module and the ECMO injury risk prediction module; The injury treatment path prediction module is used for predicting the injury treatment path. Through the prediction of the injury treatment path, injury treatment path prediction data is obtained, and the injury treatment path prediction data is sent to the ECMO injury risk integrated prediction module; The ECMO injury risk prediction module is used for predicting the ECMO injury risk. Through the prediction of the ECMO injury risk, ECMO injury risk prediction data is obtained.

Images