Multi-time-point multi-organ failure risk prediction method and device based on deep learning

Abstract

The invention relates to the technical field of deep learning, and discloses a multi-time-point multi-organ failure risk prediction method and device based on deep learning, and the method comprises the following steps: collecting ICU patient information, carrying out the preprocessing, obtaining an initial feature data set, and selecting a plurality of features most important for prediction as model input features; screening out data corresponding to the model input features from the initial feature data set as a model training set; constructing a MultiOrgRiskNet model based on a convolutional layer and a long-short-term memory network, and training by using a model training set to obtain a plurality of independent organ failure risk prediction models, including models for predicting attenuation risks of different organs after different time periods; integrating a plurality of independent organ failure risk prediction models as a multi-time-point multi-organ failure risk prediction model, and predicting multi-time-point multi-organ failure risks in real time according to clinical data. The risk of organ failure at different time points can be accurately evaluated.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and device for predicting the risk of multi-organ failure at multiple time points based on deep learning. Background Art

[0002] With the development of critical care medicine, multiple organ dysfunction (MOD) has become a common and serious clinical problem in critical care patients. MOD is usually defined as acute dysfunction of two or more organs induced by multiple factors, involving the interaction and dysfunction of multiple physiological systems.

[0003] Due to the high mortality rate of MOD, early warning and intervention have important clinical significance for improving patient prognosis. In recent years, researchers have proposed a variety of methods to identify early signs of MOD. For example, Bose et al. used the spectral clustering method to extract waveform data features based on the MODS standard proposed by IPSCC and Proulx et al., and combined with clinical data, used multiple algorithms to construct an early warning model for MODS in children. Studies have shown that the area under the ROC curve (AUC) of the random forest algorithm can reach 0.91, and its median warning time is 22.7 hours, while the median warning time of the XGBoost model is 37 hours. In addition, Li et al. first applied utility scoring to the performance comparison of sepsis early warning models and proposed a utility scoring for MODS, aiming to more objectively evaluate model performance. However, existing technologies generally have the following deficiencies: (1) Lack of specific prediction of a single organ; existing models mainly focus on the overall status of multiple organs, but lack the specific prediction ability of a single organ, which makes it difficult for clinicians to obtain targeted advice when facing dysfunction of a specific organ and cannot formulate personalized intervention plans, thus affecting the treatment effect. (2) Insufficient real-time monitoring capabilities: Although the model can make predictions, its real-time monitoring capabilities are still insufficient; this means that when the patient's condition changes, the model cannot update and adjust the prediction results in a timely manner, reducing its practicality in emergency and critical care. (3) Lack of multi-time point prediction capabilities: The existing models have limited prediction capabilities at multiple time points and cannot fully reflect the dynamic changes in the patient's condition. This limitation makes it difficult for clinicians to fully understand the patient's health status, which in turn affects the accuracy of clinical decision-making. (4) Poor data availability: The training and prediction of the model rely on high-quality clinical data, but in actual applications, some data are less available. This may limit the application of the model in specific situations, affecting its general applicability and promotion.

[0004] In order to further improve the prediction performance, the present invention develops the MultiOrgRiskNet model, which contains multiple key modules, including a data automatic interpolation module, a time series feature representation learning module, and a full risk prediction module. The MultiOrgRiskNet model has three significant advantages. First, the data automatic interpolation module introduced by the model effectively handles missing values ​​in clinical data and improves the robustness to incomplete data, thereby ensuring the accuracy and reliability of the prediction. Secondly, the integrated deep learning model combines convolutional layers and long short-term memory networks (LSTMs), which enhances the sensitivity to local features, enabling it to accurately identify short-term fluctuations in vital signs and capture dynamic changes in time series data, thereby improving the prediction performance. Finally, MultiOrgRiskNet can make real-time predictions of major organs at multiple time points, helping medical staff to quickly assess the health status of patients and make timely intervention decisions, providing strong support for critical care. Summary of the invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a multi-time point multi-organ failure risk prediction method based on deep learning, comprising the following steps:

[0007] Data collection step: collect ICU patient information and preprocess it to obtain the initial feature data set;

[0008] In the feature selection step, the most important features for prediction are selected from the initial feature data set through occlusion analysis as the model input features; the data corresponding to the model input features are screened out from the initial feature data set as the model training set;

[0009] A model training step, constructing a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and using a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods;

[0010] In the risk prediction step, several independent organ failure risk prediction models are integrated as a multi-time point multi-organ failure risk prediction model to predict the failure risk of multiple organs at multiple time points in real time based on clinical data.

[0011] Preferably, the data collection step comprises the following steps:

[0012] To obtain information on ICU patients, we used the organ failure record in the ICU to mark the event patients, while excluding patients who were younger than 18 years old or had been in the ICU for more than 60 days;

[0013] For all characteristic data, if multiple values ​​were recorded at the same time point, their average value was used as the representative value; if a specific variable was not recorded at a certain time point, the variable was considered a missing value; the initial missing values ​​at the time of patient admission were processed using the median imputation method; in other cases, the forward filling method was used;

[0014] The data of all features are subjected to Z-score standardization, expressed as:

[0015] ;

[0016] in, and are the mean and standard deviation of the data for each feature, represents the original collected feature data, Represents the feature data after normalization.

[0017] Preferably, the selecting of several features that are most important for prediction from the initial feature data set by occlusion analysis comprises the following steps:

[0018] Select several easily accessible features from the initial feature data set as candidate features;

[0019] A model based on long short-term memory network is used to perform occlusion analysis on candidate features, and several features that are most important for model prediction are selected as model input features.

[0020] Preferably, the model input features include:

[0021] Vital signs, including respiratory rate, heart rate, systolic blood pressure, diastolic blood pressure, temperature, peripheral oxygen saturation, Glasgow Coma Scale-eye, Glasgow Coma Scale-language, and Glasgow Coma Scale-movement;

[0022] laboratory results, including alanine aminotransferase, aspartate aminotransferase, albumin, blood urea nitrogen, bilirubin, C-reactive protein, chloride, creatinine, blood glucose, hemoglobin, prothrombin time, platelets, potassium, sodium, leukocytes, and activated partial thromboplastin time;

[0023] And age.

[0024] Preferably, the model input features also include three time increment features:

[0025] The vital sign increment feature indicates the time that has passed since the last measurement of any vital sign, reflecting the frequency of vital sign updates and the changing trend;

[0026] The laboratory increment feature, which represents the time that has passed since the last measurement of any laboratory result, captures the time interval between laboratory examinations and possible delayed effects;

[0027] The ICU time increment feature represents the total time since the patient was admitted to the ICU and is used to quantify the progression of his / her disease.

[0028] Preferably, the MultiOrgRiskNet model includes:

[0029] The data automatic interpolation module dynamically and in real time fills missing values ​​in the input data to ensure the data integrity in the entire time series;

[0030] The time series feature representation learning module integrates convolutional layers, long short-term memory networks and fully connected layers to comprehensively extract and express the time series features of the input data; the long short-term memory network is used to effectively capture the dynamic changes and trends of the input data and extract key time-dependent information; the convolutional layer is used to enhance the sensitivity to local features, more accurately identify short-term fluctuations, and make full use of state changes in a short period of time; the fully connected layer is used to fuse features at different levels to generate a comprehensive feature vector; the comprehensive feature vector is converted into the corresponding category probability through the nonlinear mapping of the output layer, and the risk of organ failure is determined based on the category probability.

[0031] Preferably, the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods, and the organs include the heart, liver, kidneys, lungs and coagulation system.

[0032] Preferably, the plurality of independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods, and the time periods include 12 hours, 24 hours, 36 hours and 48 hours.

[0033] Preferably, in the risk prediction step, data of model input features corresponding to clinical data are obtained and input into a multi-time point multi-organ failure risk prediction model, and the model outputs the risk of failure of different organs after different time periods.

[0034] The present invention also provides a multi-time point multi-organ failure risk prediction device based on deep learning, comprising:

[0035] Data collection module, collects ICU patient information and performs preprocessing to obtain the initial feature data set;

[0036] The feature selection module selects several features that are most important for prediction from the initial feature data set through occlusion analysis as model input features; and selects data corresponding to the model input features from the initial feature data set as the model training set;

[0037] A model training module, which constructs a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and uses a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods;

[0038] The risk prediction module integrates several independent organ failure risk prediction models as a multi-time point multi-organ failure risk prediction model, and predicts the failure risk of multiple organs at multiple time points in real time based on clinical data.

[0039] The present invention aims to develop a model that can accurately predict specific organ dysfunction and provide specific predictions for individual organs so that clinicians can obtain more targeted advice and develop personalized intervention measures. The present invention has the following beneficial effects:

[0040] (1) The present invention uses two interpolation methods to flexibly handle missing data in different situations, significantly improving the integrity of the data and the reliability of the model; the initial missing values ​​at admission are filled with the median method to avoid interference from extreme values ​​on the initial prediction and ensure the stability of the model in the early stages; and the forward filling method is used for missing data at subsequent time points to maintain data continuity and reflect the dynamic changes in the patient's status. This combination of multiple interpolation methods not only adapts to different data missing scenarios, but also reduces the impact of missing data on the prediction results, improving the accuracy and applicability of the model.

[0041] (2) The present invention incorporates three time increment features; the vital sign increment feature enables the model to capture the update frequency and change trend of vital signs, and timely identify abnormal fluctuations in the patient's status; the laboratory increment feature can capture the examination frequency and delay effect, effectively identify the potential impact of the time interval of laboratory examinations on the condition, and avoid neglecting risk information due to examination delays; the ICU time increment feature reflects the overall progression of the disease, which helps the model identify the changing trend of failure risk with hospitalization time. The three time increment features provide the model with dynamic information on the patient's status and monitoring frequency, thereby improving the accuracy and timeliness of real-time prediction of organ failure risk.

[0042] (3) The present invention introduces a real-time data monitoring mechanism to ensure that the model can update the prediction results in a timely manner and effectively reflect the dynamic changes in the patient's condition. This mechanism takes advantage of the long short-term memory network (LSTM) and can make continuous predictions based on new data at each time point. When new time point data is input, the LSTM can quickly integrate the latest vital signs and laboratory indicators to reassess the patient's risk status. This dynamic update capability improves the practicality of the model in emergency and critical care, allowing medical staff to respond quickly and adjust treatment plans in a timely manner to cope with changes in the patient's condition.

[0043] (4) The present invention constructs a model that can make predictions at multiple time points, which comprehensively reflects the health status of patients. It will enable clinicians to better understand the evolution of patients' conditions and help optimize clinical decisions. It fully utilizes the advantages of time series data and constructs prediction models at different time points based on the same outcome. This method can not only capture the changing trend of the disease over time, but also provide more accurate risk assessment, thereby helping clinicians better understand the evolution of patients' conditions and provide strong support for optimizing clinical decisions.

[0044] (5) Based on the consideration of data availability, the present invention enhances the applicability of the model by integrating multiple data sources. The data used are derived from commonly used laboratory indicators and monitoring signs during intensive care, and are screened by clinical experts to ensure the representativeness and clinical relevance of the data. This method makes data acquisition less difficult, thereby supporting the model to maintain good performance in a wide range of clinical applications.

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A method step diagram of an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of model training according to an embodiment of the present invention;

[0048] Figure 3 FIG. 4 is a diagram showing the structure of a device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] Organ failure is the main cause of death in intensive care unit (ICU) patients. This invention aims to help medical professionals and researchers more efficiently and accurately assess the risk of failure of major organs in ICU patients through a real-time multi-organ failure risk prediction system based on deep learning. The system can comprehensively analyze the functions of multiple key organs and provide real-time risk assessment so that the clinical team can intervene as early as possible to reduce the possibility of disease deterioration.

[0050] The embodiment of the present invention has developed a deep learning model called MultiOrgRiskNet (Multi-Organ Failure Risk Prediction Network) based on routine clinical indicator data of more than 12,000 intensive care patients to assess the risk of failure of multiple major organs in ICU patients. This not only provides data support for personalized diagnosis and treatment management, but also lays a solid foundation for preventive treatment and clinical decision-making assistance. With the help of the MultiOrgRiskNet model, medical professionals can identify potential organ function decline trends earlier and formulate more effective intervention strategies for patients. Intervention measures may include optimizing drug treatment regimens, implementing more targeted preventive medical measures, and more flexible care plans. By reducing the risk of organ failure, MultiOrgRiskNet can not only significantly reduce the mortality rate of patients, but also effectively reduce treatment costs and reduce the economic burden on patients and their families. In addition, the system helps to improve the quality of life and overall prognosis of patients, providing a strong technical support for modern critical care medicine.

[0051] For details, see Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0052] S101, data collection step, collecting ICU patient information and preprocessing to obtain an initial feature data set;

[0053] S102, feature selection step, selecting several features that are most important for prediction from the initial feature data set through occlusion analysis as model input features; filtering data corresponding to the model input features from the initial feature data set as a model training set;

[0054] S103, a model training step, constructing a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and using a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods;

[0055] S104, risk prediction step, integrates several independent organ failure risk prediction models as a multi-time point multi-organ failure risk prediction model, and predicts the failure risk of multiple organs at multiple time points in real time based on clinical data.

[0056] The specific contents are as follows.

[0057] (1) Data collection and preprocessing. Data collection includes data extraction, population definition, and data cleaning. First, relevant information was extracted from the electronic health records of ICU patients. Event patients were marked with organ failure records in the ICU, and patients under the age of 18 or hospitalized in the ICU for more than 60 days were excluded. Then, the data of different cohorts were integrated into the same table, and similar features were grouped and manually classified and sorted by clinical experts. For all features, when multiple values ​​were recorded at the same time point, their average value was used as the representative value. Samples were generated at each time point and contained at least one valid variable; if a specific variable was not recorded at a certain time point, the variable was considered to be missing. For the initial missing values ​​when the patient was admitted to the hospital, the median imputation method was used; in other cases, the forward filling method was used to replace the missing values. Finally, all features were Z-score standardized according to the statistical parameters of the training data set to ensure that the dimensional differences between different features were eliminated and that they were all within the same scale range, thereby avoiding the imbalanced impact of some features on the model due to excessive numerical ranges. The formula is:

[0058] ;

[0059] in, and are the mean and standard deviation of each feature, respectively.

[0060] (2) Feature selection stage. First, a panel of clinical experts selected 30 candidate laboratory indicators that were easy to obtain. These variables were then subjected to a block analysis using a model based on a long short-term memory network (LSTM). Block analysis is a method that evaluates the impact of input variables on the predictive performance of a model by gradually excluding or modifying them. Specifically, block analysis determines the importance of a variable in the overall model by setting the value of a single variable to zero or other specific values ​​and then observing changes in model performance. In this method, each variable is "blocked" one by one, and its value is set to zero to evaluate the contribution of the variable to the model's prediction results. Then, the importance of the variable to the prediction results is determined by comparing the changes in model performance (usually measured by the area under the receiver operating characteristic curve, or AUROC) with and without the variable. For each variable, if the AUROC value of the model changes little after being set to zero, it can be considered that the variable contributes little to the predictive performance. Next, we screen the variables based on the impact of each variable on the reduction of AUROC. The specific approach is to exclude those variables that have the smallest impact on the AUROC value. Usually, we will first delete the variables with the smallest reduction in AUROC, indicating that these variables have limited effect on model prediction. This process is carried out iteratively, that is, each time the variable that contributes the least to model performance is deleted until the number of retained features reaches an acceptable level (for example, AUROC decreases less than 0.002) and the decline in model performance remains within a reasonable range (AUROC decreases less than 0.002). This method can help us identify features that are critical to model predictions, while eliminating variables with less impact, ensuring that the model is simplified and overfitting is avoided while retaining model performance.

[0061] After feature selection was completed, the final model input features were shown in Table 1, including 9 vital signs and 16 laboratory results.

[0062] Table 1 Model input characteristics:

[0063]

[0064] In order to further improve the predictive ability of the model, three time increment features are introduced. These incremental features are used to capture the dynamic changes of features over time to improve the model's prediction accuracy for the risk of multiple organ failure. The vital sign incremental feature is defined as the time since the last measurement of any vital sign, reflecting the frequency and change trend of vital sign updates; the laboratory incremental feature records the time since the last measurement of any laboratory result, capturing the time interval and possible delay effect of laboratory examinations. In addition, the ICU time incremental feature is defined to measure the total time since the patient was admitted to the ICU to quantify the progression of his or her disease. These time incremental features provide the model with additional information about the evolution of the patient's health status over time, allowing the model to more comprehensively capture the dynamic changes of the disease, thereby improving the prediction accuracy of the risk of multiple organ failure. This design not only brings higher flexibility and dynamic adaptability to the model, but also provides clinicians with more accurate and timely diagnostic basis, which is helpful for early intervention and optimization of treatment plans.

[0065] (3) Model construction. The MultiOrgRiskNet model combines deep learning technology and clinical data analysis, aiming to provide accurate organ failure risk estimation through efficient feature learning and prediction algorithms. The model includes a data automatic interpolation module and a time series feature representation learning module.

[0066] The data automatic interpolation module is used to handle missing values ​​in clinical data and ensure that the model can continue to run even when the information is incomplete. This module can dynamically and in real time fill in missing values ​​as the data is recorded and input, ensuring the integrity of the data in the entire time series; the data automatic interpolation module in the MultiOrgRiskNet model is a key link, which is used to handle the problem of missing values ​​in medical data. Clinical data often have incomplete records, such as patients not undergoing certain vital signs or laboratory tests at certain time points, and these missing values ​​may have a significant impact on the accuracy of the model. Therefore, interpolation technology is used to replace these missing values ​​to avoid performance degradation due to missing data during the calculation process. This module effectively solves the problem of missing values ​​in clinical data through interpolation technology, ensuring that the model can still provide reliable predictions even when the information is incomplete. This mechanism supports dynamic and real-time filling of missing values ​​to maintain data integrity, thereby improving the stability and applicability of the model in practical applications.

[0067] The embodiment of the present invention adopts two interpolation methods. For features with multiple values ​​recorded at the same time point, the average of these values ​​is used as the representative value. This is a robust processing method that can eliminate the influence of extreme values ​​and ensure the balance of data. Samples are generated at each time point to ensure that there is at least one valid variable, reducing the impact of data sparsity on the model. For variables that are not recorded at certain time points, they are regarded as missing values ​​and interpolated. For different missing scenarios, the model designs two interpolation strategies. First, for the initial missing values ​​when the patient is admitted to the hospital (i.e., the first time entering the ICU), the model uses the median filling method. The median is a statistic that is not affected by extreme values ​​and can effectively reflect the central tendency of the sample. Therefore, it is suitable for estimating the initial value when there is no historical record at the time of admission. The advantage of the median filling method is that it can reduce the impact of noise on the interpolation results, especially when there is large heterogeneity in the admission data. Secondly, for time series data after admission, if some features are missing at subsequent time points, the model uses forward fill. This method is based on the assumption of temporal continuity of medical data, that is, the characteristic values ​​of adjacent time points will not change dramatically, so the current missing values ​​can be filled with the observed values ​​of the previous moment. This interpolation method can maintain the continuity of the data while reducing the risk of introducing unnecessary fluctuations. The forward filling method is particularly suitable for use in medical scenarios where certain physiological characteristics of patients may remain stable for a short period of time. Through these interpolation techniques, the model ensures the integrity and continuity of the data throughout the time series, allowing the LSTM network to effectively capture the temporal characteristics of the data and avoid affecting the learning and prediction capabilities of the model due to missing values. This dynamic real-time interpolation mechanism not only improves the quality of the data, but also provides more reliable input data for subsequent feature learning and risk prediction modules, thereby significantly improving the prediction performance of the model.

[0068] The time series feature representation learning module integrates convolutional layers, long short-term memory networks (LSTMs), and fully connected layers, aiming to conduct in-depth time series feature learning of patients' clinical data. Through the LSTM network, the module can effectively capture the dynamic changes and trends of vital signs and laboratory results, and extract key time-dependent information. The introduction of convolutional layers enhances the model's sensitivity to local features, enabling it to more accurately identify short-term fluctuations. In addition, the fully connected layer further integrates features at different levels, providing the model with a more comprehensive understanding and prediction capability. The LSTM network effectively captures the dynamic changes of vital signs and laboratory results, enabling the model to extract key time-dependent information and deeply analyze the patient's time series data, thereby more accurately identifying potential risks; the introduction of convolutional layers enhances the model's sensitivity to local features, thereby more accurately identifying short-term fluctuations and changes, which is particularly important when monitoring the patient's status, and can promptly reflect potential health risks and help medical staff respond more quickly; this multi-level feature learning capability greatly improves the accuracy and reliability of predictions. Combining these components, the module provides strong support for real-time monitoring and prediction of patient health status.

[0069] The time series feature representation learning module uses convolutional layers and long short-term memory networks (LSTM) to perform in-depth time series feature learning on patients' clinical data, focusing on capturing the dynamic changes and trends of vital signs and laboratory results to extract important time-dependent information.

[0070] First, a convolutional layer is introduced to extract local features. The convolutional layer can identify short-term change patterns by performing convolution operations on the input data, which is crucial for feature extraction of time series data. The convolution operation helps the model focus on subtle changes in local areas and enhances sensitivity to these short-term patterns, thereby providing richer feature representations for subsequent LSTM layers.

[0071] Next, LSTM is used for time series feature learning. LSTM is a special recurrent neural network (RNN) structure designed specifically for processing and predicting time series-based data. The core of LSTM lies in its three gating mechanisms: forget gate, input gate, and output gate, which control the flow of information in each time step.

[0072] A. Forget gate: controls which information needs to be forgotten.

[0073] ;

[0074] in, is the output of the forget gate, is the sigmoid activation function and is the weight matrix, is the hidden state at the previous time step, is the input of the current time step, Is the bias term. Through the output of the sigmoid function between 0 and 1, the model decides which historical information to forget (forget when it is close to 0, and retain when it is close to 1).

[0075] B. Input gate: controls what new information will be written into the memory cell.

[0076] ;

[0077] ;

[0078] in, is the output of the input gate, It is the new candidate memory value, which is determined by the input gate and the forget gate, which determines which new information will be added to the cell state.

[0079] C. Cell state update: Update the cell state by combining the results of the forget gate and the input gate.

[0080] ;

[0081] in, is the cell state at the current time step, is the cell state at the previous time step, the forget gate determines how much previous information to retain, and the input gate determines how much new information to add.

[0082] D. Output gate: controls the output of LSTM, which is the hidden state of the next time step.

[0083] ;

[0084] ;

[0085] in, is the output of the output gate, is the hidden state at the current time step.

[0086] Through this mechanism, LSTM can effectively capture the long-term time dependency and short-term dynamic changes of vital signs and laboratory results, thereby extracting features that are critical to the evolution of the patient's health status. At the same time, during the model training process, we applied appropriate regularization techniques to prevent overfitting, including dropout and early stopping. When the loss of the validation set no longer decreases within 5 epochs, the training is terminated early to prevent the model from overfitting.

[0087] Finally, fully connected layers are used to integrate the features extracted from the convolutional layers and LSTM layers to further improve the model's expressiveness. These fully connected layers enable the model to better capture complex patterns by combining the extracted features nonlinearly. By combining multiple layers of LSTM structures, the model is able to capture features on different time scales, from short-term to long-term, ensuring a comprehensive assessment of the patient's status.

[0088] (4) Model training see Figure 2As shown, the MultiOrgRiskNet model is trained using the model training set to obtain several independent organ failure risk prediction models, which can dynamically predict the multi-time point failure risks of the patient's five major organs (heart, liver, kidney, lung, and coagulation system) in real time based on the input clinical data. The model output includes risk scores for each time point and organ, which can be further converted into the probability of failure, helping clinicians identify high-risk organs and make timely intervention decisions. The module supports real-time data input and update to ensure that the latest risk assessment is provided when the patient's condition changes. The organ failure risk prediction model predicts the failure risk of multiple organs in real time based on the features learned by LSTM. This feature enables medical staff to quickly assess the patient's health status and make timely interventions at critical moments, significantly improving the efficiency and accuracy of clinical decision-making. The embodiment of the present invention can not only predict multiple organs, but also perform real-time monitoring at multiple time points. This capability enables medical staff to fully grasp the health dynamics of patients and provide more detailed data support for clinical intervention. In order to respond to different clinical events, the module can perform dynamic evaluations at multiple time points to predict the possibility of organ failure in the future, such as within 12 hours, 24 hours, 36 hours, and 48 hours. This forward-looking risk assessment enables medical personnel to identify and intervene in time before organ failure occurs, reducing the progression of multiple organ failure. Through the integration of deep learning models, the risk prediction module provides clinicians with an early warning tool, improves the efficiency of patient management and intervention in the intensive care unit (ICU), helps optimize medical decisions, and improves patient prognosis. The embodiment of the present invention can train multiple independent prediction models for the possibility of organ failure of the heart, liver, kidney, lung, and coagulation system within 12 hours, 24 hours, 36 hours, and 48 hours through a training data set collected once, and integrates these independent prediction models to obtain a multi-time point multi-organ failure risk prediction model. The multi-time point multi-organ failure risk prediction model can be used to input clinical data once and simultaneously obtain the possibility of organ failure of the above organs within 12 hours, 24 hours, 36 hours, and 48 hours.

[0089] (5) Model performance verification. The dataset was divided into a development set and a test set according to a 7:3 ratio for verification, and the area under the receiver operating characteristic curve (AUROC) was used to evaluate and compare the prediction performance of the model. The developed model was compared with traditional scoring systems, including the single parameter weighted tracking and triggering system (SPTTS), the modified early warning score (MEWS), the national early warning score (NEWS), the acute physiology and chronic health evaluation (APACHE-II), the simplified acute physiology score (SAPS), and the sequential organ failure assessment (SOFA). For scoring models that require time series data, the latest available data was used, such as the real-time oxygen partial pressure and alveolar-arterial oxygen partial pressure difference in the APACHE-II score.

[0090] In addition to AUROC, the area under the precision-recall curve (AUPRC) was calculated as a secondary evaluation indicator, and the calibration analysis of the model was performed. The calibration effect of the model was further evaluated by calculating the expected calibration error (ECE).

[0091] In the test set, the 12-hour, 24-hour, 36-hour, and 48-hour failure prediction performance of major organs (heart, liver, kidney, lung, and coagulation system) was evaluated. The model showed high accuracy and calibration at each time point. Specifically, the AUC range of 12-hour prediction was 0.78-0.82, which can better capture the short-term risk of organ failure; the AUC of 24-hour prediction was greater than 0.8, showing the best prediction accuracy and balanced calibration ability; the AUC of 36-hour prediction was between 0.75-0.84, showing that the model still has high prediction performance at medium-term time points; the AUC of 48-hour prediction was between 0.76-0.81. Despite the longer prediction time, the model still showed a robust prediction effect. Overall, through these dynamic predictions at multiple time points, the model can provide reliable assessment support for the risk of organ failure in different time periods, which is helpful for clinical decision making and early intervention.

[0092] For details, see Figure 3 FIG. 1 is a diagram showing a structure of a device according to an embodiment of the present invention, comprising:

[0093] Data collection module 301, collects ICU patient information and performs preprocessing to obtain an initial feature data set;

[0094] The feature selection module 302 selects several features that are most important for prediction from the initial feature data set through occlusion analysis as model input features; and selects data corresponding to the model input features from the initial feature data set as a model training set;

[0095] The model training module 303 constructs a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and uses a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods;

[0096] The risk prediction module 304 integrates several independent organ failure risk prediction models as a multi-time point multi-organ failure risk prediction model, and predicts the failure risk of multiple organs at multiple time points in real time based on clinical data.

[0097] The embodiment of the present invention uses advanced data processing and statistical analysis technology, combined with electronic health record data from hospitals and information on more than 12,000 multi-center ICU patients from the public critical care databases MIMIC-III and MIMIC-IV, to develop a real-time multi-organ failure risk prediction method based on clinical indicators and the MultiOrgRiskNet model. Through efficient data integration and analysis, this method can accurately assess the risk of organ failure in ICU patients, achieve real-time prediction and assist decision-making. Its purpose is to provide medical personnel with more timely and reliable risk assessment support, help optimize clinical intervention strategies, and thereby improve patient treatment outcomes and survival rates.

[0098] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-time point multi-organ failure risk prediction method based on deep learning, characterized in that: The following steps are involved: Data collection step: collect ICU patient information and preprocess it to obtain the initial feature data set; Feature selection step: select several features that are most important for prediction from the initial feature data set through occlusion analysis as model input features; Filter the data corresponding to the model input features from the initial feature data set as the model training set; A model training step, constructing a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and using a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods; In the risk prediction step, several independent organ failure risk prediction models are integrated as a multi-time point multi-organ failure risk prediction model to predict the failure risk of multiple organs at multiple time points in real time based on clinical data.

2. The multi-time point multi-organ failure risk prediction method based on deep learning according to claim 1 is characterized in that: The data collection step includes the following steps: To obtain information on ICU patients, we used the organ failure record in the ICU to mark the event patients, while excluding patients who were younger than 18 years old or had been in the ICU for more than 60 days; For all characteristic data, if multiple values ​​were recorded at the same time point, their average value was used as the representative value; if a specific variable was not recorded at a certain time point, the variable was considered a missing value; the initial missing values ​​at the time of patient admission were processed using the median imputation method; in other cases, the forward filling method was used; The data of all features are subjected to Z-score standardization, expressed as: ； in, and are the mean and standard deviation of the data for each feature, represents the original collected feature data, Represents the feature data after normalization.

3. The multi-time point multi-organ failure risk prediction method based on deep learning according to claim 1 is characterized in that: The method of selecting several features that are most important for prediction from the initial feature data set through occlusion analysis includes the following steps: Select several easily accessible features from the initial feature data set as candidate features; A model based on long short-term memory network is used to perform occlusion analysis on candidate features, and several features that are most important for model prediction are selected as model input features.

4. The multi-time point multi-organ failure risk prediction method based on deep learning according to claim 3 is characterized in that: The model input features include: Vital signs, including respiratory rate, heart rate, systolic blood pressure, diastolic blood pressure, temperature, peripheral oxygen saturation, Glasgow Coma Scale-eye, Glasgow Coma Scale-language, and Glasgow Coma Scale-movement; laboratory results, including alanine aminotransferase, aspartate aminotransferase, albumin, blood urea nitrogen, bilirubin, C-reactive protein, chloride, creatinine, blood glucose, hemoglobin, prothrombin time, platelets, potassium, sodium, leukocytes, and activated partial thromboplastin time; And age.

5. The method for predicting risk of multiple organ failure at multiple time points based on deep learning according to claim 4, characterized in that: The model input features also include three time increment features: The vital sign increment feature indicates the time that has passed since the last measurement of any vital sign, reflecting the frequency of vital sign updates and the changing trend; The laboratory increment feature, which represents the time that has passed since the last measurement of any laboratory result, captures the time interval between laboratory examinations and possible delayed effects; The ICU time increment feature represents the total time since the patient was admitted to the ICU and is used to quantify the progression of his / her disease.

6. The method for predicting risk of multiple organ failure at multiple time points based on deep learning according to claim 1, characterized in that: The MultiOrgRiskNet model includes: The data automatic interpolation module dynamically and in real time fills missing values ​​in the input data to ensure the data integrity in the entire time series; The time series feature representation learning module integrates convolutional layers, long short-term memory networks and fully connected layers to comprehensively extract and express the time series features of the input data; the long short-term memory network is used to effectively capture the dynamic changes and trends of the input data and extract key time-dependent information; the convolutional layer is used to enhance the sensitivity to local features, more accurately identify short-term fluctuations, and make full use of state changes in a short period of time; the fully connected layer is used to fuse features at different levels to generate a comprehensive feature vector; the comprehensive feature vector is converted into the corresponding category probability through the nonlinear mapping of the output layer, and the risk of organ failure is determined based on the category probability.

7. The multi-time point multi-organ failure risk prediction method based on deep learning according to claim 1, characterized in that: The several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods, and the organs include the heart, liver, kidneys, lungs and coagulation system.

8. The method for predicting risk of multiple organ failure at multiple time points based on deep learning according to claim 7, characterized in that: The several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods, and the time periods include 12 hours, 24 hours, 36 hours and 48 hours.

9. The method for predicting risk of multiple organ failure at multiple time points based on deep learning according to claim 8, characterized in that: In the risk prediction step, data of model input features corresponding to clinical data are obtained and input into a multi-time point multi-organ failure risk prediction model, and the model outputs the risk of failure of different organs after different time periods.

10. A multi-time point multi-organ failure risk prediction device based on deep learning, characterized in that: include: Data collection module, collects ICU patient information and performs preprocessing to obtain the initial feature data set; The feature selection module selects several features that are most important for prediction from the initial feature data set through occlusion analysis as model input features; Filter the data corresponding to the model input features from the initial feature data set as the model training set; A model training module, which constructs a MultiOrgRiskNet model based on a convolutional layer and a long short-term memory network, and uses a model training set to train to obtain a number of independent organ failure risk prediction models; the several independent organ failure risk prediction models include models for predicting the risk of attenuation of different organs after different time periods; The risk prediction module integrates several independent organ failure risk prediction models as a multi-time point multi-organ failure risk prediction model, and predicts the failure risk of multiple organs at multiple time points in real time based on clinical data.

Images