Real-time hypotension risk prediction method based on composite neural network
By building a composite neural network model and combining multiple optimization strategies, high-precision prediction of the risk of hypotension is solved, and the problem of difficulty in achieving accurate real-time prediction of traditional methods is significantly improved.
Patent Information
- Application Number
- CN202510189665.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-20
- Publication Date
- 2025-07-01
AI Technical Summary
Traditional hypotension risk prediction methods are difficult to achieve accurate real-time prediction, and the existing methods have shortcomings in feature selection, model training and real-time updates.
The real-time hypotension risk prediction method based on composite neural network is adopted, and the first prediction model and the second prediction model are constructed, combined with a variety of optimization strategies, including detailed preprocessing of historical physiological data, identification of peaks and troughs, construction of CNN-LSTM models, and support of online learning mechanisms.
It effectively improves the accuracy and real-time nature of hypotension risk prediction, reduces the impact of noise and outliers on model performance, and improves the reliability and accuracy of prediction results.
Smart Images
Figure CN120236753A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of data prediction and relates to a real-time hypotensive risk prediction method based on a composite neural network. Background Art
[0002] Hemodialysis, commonly known as "blood washing", is a process of drawing the blood of end-stage renal disease patients out of the body, passing it through a dialyzer to remove metabolic wastes and impurities in the blood, and then returning the purified blood back into the body, which has a good effect on the treatment of patients' kidney diseases. Hypotension is one of the manifestations of unstable blood system during dialysis, a common complication in hemodialysis, with an incidence rate ranging from 5% to 40%. Clinically, there are some emergency measures such as injecting normal saline, reducing the temperature of dialysis fluid and ultrafiltration rate, etc., but these emergency measures will have certain side effects. When hypotension is severe, dialysis has to be terminated, and the failure to reach the expected ultrafiltration volume in a single treatment will have a great impact on the cardiopulmonary function and survival rate of patients. The occurrence of hypotension will not only affect the dialysis adequacy, but also lead to the formation of thrombus in the blood vessel access of hemodialysis patients, mesenteric ischemia and cardiovascular diseases. Therefore, predicting the probability of hypotension during dialysis has important research significance and application value.
[0003] Traditional hypotensive risk prediction methods often rely on doctors' experience and simple physiological index monitoring, and it is difficult to achieve accurate real-time prediction. With the development of machine learning and deep learning technologies, it has become possible to use historical physiological data for risk prediction. However, existing methods still have deficiencies in feature selection, model training and real-time update, etc.
[0004] The present invention aims to provide a real-time hypotensive risk prediction method based on a composite neural network, which can achieve high-precision prediction of hypotensive risk by constructing a first prediction model and a second prediction model and combining multiple optimization strategies. Summary of the Invention
[0005] The object of the present invention is to propose a real-time hypotensive risk prediction method based on a composite neural network for the above problems existing in the prior art.
[0006] The object of the present invention can be achieved by the following technical solutions: A real-time hypotensive risk prediction method based on a composite neural network, comprising the steps of:
[0007] Construct a first prediction model, which is trained by the historical physiological data of patients and the resource sample data corresponding to the historical physiological data;
[0008] Based on the first prediction model, determine the target historical physiological sample data and the first training parameters corresponding to the target historical physiological sample data;
[0009] Determine the first target parameter corresponding to the first prediction model based on the first training parameter;
[0010] Construct a second prediction model based on the target historical physiological sample data and the first training parameter, where the second prediction model is a CNN-LSTM model;
[0011] Determine the second target parameter corresponding to the second prediction model;
[0012] Adjust the second prediction model based on the first target parameter and the second target parameter to obtain a final hypotensive risk prediction model.
[0013] In the above real-time hypotensive risk prediction method based on a composite neural network, before constructing the first prediction model, there is also a step of preprocessing historical physiological data, and the preprocessing includes:
[0014] Delete the clinical data of target objects with an age less than the preset age and the clinical data of target objects with fewer hemodialysis times than the preset number of times;
[0015] Divide the target clinical data into time-series data and non-time-series data;
[0016] Use the interpolation method to complete the missing part of the time-series data, and use the Gaussian distribution filling method or the mean filling method to complete the missing part of the non-time-series data, and delete the invalid data after completion from the target clinical data;
[0017] Correct the time-series data or non-time-series data with abnormal data types and abnormal numerical values, and delete the invalid data after correction from the target clinical data.
[0018] In the above real-time hypotensive risk prediction method based on a composite neural network, the determining the target historical physiological sample data and the first training parameter corresponding to the target historical physiological sample data based on the first prediction model includes:
[0019] Determine at least one peak and at least one trough of the first prediction model;
[0020] Use the historical physiological sample data corresponding to the at least one peak and the at least one trough as the target historical physiological sample data;
[0021] Input the target historical physiological sample data into the first pre-designed calculation formula respectively, and obtain the first training parameter through the first pre-designed calculation formula.
[0022] In the above real-time hypotensive risk prediction method based on a composite neural network, the first pre-designed calculation formula is:
[0023]
[0024] where T is the number of trees, and Δ i is the splitting gain caused by a certain feature in the i-th tree, and N is the number of samples.
[0025] In the above real-time hypotensive risk prediction method based on a composite neural network, determining the first target parameter corresponding to the first prediction model based on the first training parameter includes:
[0026] Determining a first target value from the first training parameter based on a second pre-designed calculation formula, where the first target value is used to indicate the level of robustness;
[0027] Determining the first target parameter corresponding to the first target value based on the first prediction model.
[0028] In the above real-time hypotensive risk prediction method based on a composite neural network, the robustness evaluation formula is:
[0029] where is the loss between the original predicted value and the true value, is the loss between the predicted value after adding noise and the true value.
[0030] In the above real-time hypotensive risk prediction method based on a composite neural network, constructing a second prediction model based on the target historical physiological sample data and the first training parameter includes:
[0031] Combining the target historical physiological sample data to obtain a first set;
[0032] Combining the first training parameter to obtain a second set;
[0033] Fitting to obtain the mapping relationship between the first set and the second set;
[0034] Constructing a second prediction model based on the target historical physiological sample data, the first training parameter, and the mapping relationship, where the second prediction model is a CNN-LSTM model, and the CNN-LSTM model includes:
[0035] A convolutional layer for extracting the spatial features of time series data;
[0036] An LSTM layer for capturing the long-term dependency relationships of time series data;
[0037] A fully connected layer for outputting the final prediction result.
[0038] In the above real-time hypotension risk prediction method based on a composite neural network, the convolutional layer uses 64 filters with a kernel size of 3 and a ReLU activation function; the LSTM layer uses 50 LSTM units with a ReLU activation function; the fully connected layer uses a Sigmoid activation function to output a single probability value.
[0039] In the above real-time hypotension risk prediction method based on a composite neural network, determining the second target parameter corresponding to the second prediction model includes:
[0040] Based on the second prediction model, taking the position corresponding to the maximum value as the second target parameter;
[0041] During the training process, K-fold cross-validation is used, where K is 5 or 10.
[0042] In the above real-time hypotension risk prediction method based on a composite neural network, the final hypotension risk prediction model supports an online learning mechanism, can update model parameters in real time according to newly collected historical physiological data, and the historical physiological data includes multimodal data, and the multimodal data includes but is not limited to electrocardiogram, blood oxygen saturation, and body temperature.
[0043] Compared with the prior art, the beneficial effects of the present invention are:
[0044] 1. By constructing two prediction models and combining multiple optimization strategies, the present invention can effectively improve the accuracy and real-time performance of hypotension risk prediction, thereby better supporting clinical decision-making.
[0045] 2. By performing detailed preprocessing on historical physiological data, the quality and consistency of the data are ensured, the influence of noise and outliers on the model performance is reduced, and the reliability and accuracy of the prediction results are improved.
[0046] 3. By identifying peaks and valleys to select key historical physiological sample data, it helps to capture important feature points, thereby improving the robustness and prediction ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 is the data preprocessing flowchart of the present invention.
[0048] Figure 2 is the construction and training flowchart of the first prediction model of the present invention.
[0049] Figure 3 is the schematic diagram of the selection of target historical physiological sample data and the calculation of feature importance of the present invention.
[0050] Figure 4Schematic diagram of the CNN-LSTM model structure of the present invention.
[0051] Figure 5 Schematic diagram of K-fold cross-validation of the present invention.
[0052] Figure 6 Schematic diagram of the online learning mechanism of the present invention.
[0053] Figure 7 Schematic diagram of multi-modal data fusion of the present invention. Detailed implementation manners
[0054] The following are specific embodiments of the present invention and, in conjunction with the accompanying drawings, further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments.
[0055] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement conditions between components in a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.
[0056] In addition, in the present invention, descriptions such as "first", "second", "one", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.
[0057] In the present invention, unless otherwise clearly defined and limited, terms such as "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and can be the internal connection of two components or the interaction relationship between two components, unless otherwise clearly limited. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0058] In addition, the technical solutions between various embodiments of the present invention can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.
[0059] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.
[0060] As Figures 1 - 7 shown, a real-time hypotension risk prediction method based on a composite neural network includes the steps of:
[0061] Construct a first prediction model, which is trained by the historical physiological data of patients and the resource sample data corresponding to the historical physiological data;
[0062] Based on the first prediction model, determine the target historical physiological sample data and the first training parameters corresponding to the target historical physiological sample data;
[0063] Based on the first training parameters, determine the first target parameters corresponding to the first prediction model;
[0064] Based on the target historical physiological sample data and the first training parameters, construct a second prediction model, which is a CNN-LSTM model;
[0065] Determine the second target parameters corresponding to the second prediction model;
[0066] Based on the first target parameters and the second target parameters, adjust the second prediction model to obtain the final hypotension risk prediction model.
[0067] In this embodiment, the present invention can effectively improve the accuracy and real-time performance of hypotension risk prediction by constructing two prediction models and combining various optimization strategies, thereby better supporting clinical decision-making.
[0068] As Figures 1 - 7 shown, on the basis of the above embodiment, before constructing the first prediction model, it further includes the step of preprocessing the historical physiological data, and the preprocessing includes:
[0069] Delete the clinical data of target objects with an age less than the preset age and the clinical data of target objects with fewer hemodialysis times than the preset number of times;
[0070] Divide the target clinical data into time-series data and non-time-series data;
[0071] Use the interpolation method to complete the missing part of the time-series data, and use the Gaussian distribution filling method or the mean filling method to complete the missing part of the non-time-series data, and delete the invalid data after completion from the target clinical data;
[0072] Correct the time-series data or non-time-series data of data type exceptions and numerical exceptions, and delete the invalid data after correction from the target clinical data.
[0073] In this embodiment, by performing detailed preprocessing on historical physiological data, the quality and consistency of the data are ensured, the influence of noise and outliers on the model performance is reduced, and the reliability and accuracy of the prediction results are improved.
[0074] As Figures 1 - 7 shown, based on the above embodiment, determining the target historical physiological sample data and the first training parameters corresponding to the target historical physiological sample data based on the first prediction model includes:
[0075] Determine at least one peak and at least one trough of the first prediction model;
[0076] Use the historical physiological sample data corresponding to the at least one peak and the at least one trough as the target historical physiological sample data;
[0077] Input the target historical physiological sample data into a first pre-designed calculation formula respectively, and obtain the first training parameters through the first pre-designed calculation formula.
[0078] In this embodiment, by identifying peaks and troughs to select key historical physiological sample data, it helps to capture important feature points, thereby improving the robustness and prediction ability of the model.
[0079] As Figures 1 - 7 shown, based on the above embodiment, the first pre-designed calculation formula is:
[0080]
[0081] where T is the number of trees, Δ i is the split gain caused by a certain feature in the i-th tree, and N is the number of samples.
[0082] In this embodiment, this calculation formula can quantify the importance of each feature, help select the most influential features, thereby improving the performance and interpretability of the model.
[0083] As Figures 1 - 7 shown, based on the above embodiment, determining the first target parameter corresponding to the first prediction model based on the first training parameters includes:
[0084] Based on a second pre-designed calculation formula, determine a first target value from the first training parameters, and the first target value is used to indicate the level of robustness;
[0085] Based on the first prediction model, determine the first target parameter corresponding to the first target value.
[0086] In this embodiment, by evaluating the robustness, it can be ensured that the model can still maintain a high prediction performance when facing different noises and changing data, enhancing the stability and reliability of the model.
[0087] As Figures 1 - 7 shown, based on the above embodiment, the robustness evaluation formula is:
[0088] where is the loss between the original predicted value and the true value, is the loss between the predicted value after adding noise and the true value.
[0089] In this embodiment, this robustness evaluation formula can quantify the performance of the model in a noisy environment, help select more stable model parameters, and further improve the generalization ability and reliability of the model.
[0090] As Figures 1 - 7 shown, based on the above embodiment, constructing the second prediction model based on the target historical physiological sample data and the first training parameter includes:
[0091] Combine the target historical physiological sample data to obtain a first set;
[0092] Combine the first training parameter to obtain a second set;
[0093] Fit to obtain the mapping relationship between the first set and the second set;
[0094] Based on the target historical physiological sample data, the first training parameter, and the mapping relationship, construct a second prediction model, the second prediction model is a CNN-LSTM model, and the CNN-LSTM model includes:
[0095] A convolutional layer for extracting the spatial features of time series data;
[0096] An LSTM layer for capturing the long-term dependence relationship of time series data;
[0097] A fully connected layer for outputting the final prediction result.
[0098] In this embodiment, by constructing a CNN-LSTM model and combining the advantages of the convolutional layer and the LSTM layer, it can effectively extract the spatial features and long-term dependence relationship of time series data, significantly improving the accuracy of hypotensive risk prediction.
[0099] AsFigures 1 - 7 As shown, based on the above embodiments, the convolutional layer uses 64 filters, the kernel size is 3, and the activation function is ReLU; the LSTM layer uses 50 LSTM units, and the activation function is ReLU; the fully connected layer uses the Sigmoid activation function to output a single probability value.
[0100] In this embodiment, the specific configurations of the convolutional layer and the LSTM layer can optimize the performance of the model, ensure good performance when processing complex time series data, and at the same time, the Sigmoid activation function can map the prediction results to the interval [0,1], which is convenient for interpretation.
[0101] As Figures 1 - 7 shown, based on the above embodiments, determining the second target parameter corresponding to the second prediction model includes:
[0102] Based on the second prediction model, taking the position corresponding to the maximum value as the second target parameter;
[0103] During the training process, K-fold cross-validation is used, where K is 5 or 10.
[0104] In this embodiment, through K-fold cross-validation, the performance of the model can be more comprehensively evaluated on a limited data set, and the best parameter combination can be selected, thereby improving the generalization ability and stability of the model.
[0105] As Figures 1 - 7 shown, based on the above embodiments, the final low blood pressure risk prediction model supports an online learning mechanism, can update the model parameters in real time according to newly collected historical physiological data, and the historical physiological data includes multi-modal data, and the multi-modal data includes but is not limited to electrocardiogram, blood oxygen saturation, and body temperature.
[0106] In this embodiment, supporting the online learning mechanism enables the model to adapt to new data changes and maintain the latest prediction ability, while the multi-modal data fusion provides a richer information source, improving the comprehensive judgment ability and prediction accuracy of the model.
[0107] As Figures 1 - 7 shown, the overall work process is as follows
[0108] 1. Data preprocessing
[0109] Before constructing the first prediction model, it is necessary to preprocess the historical physiological data to ensure the quality and consistency of the data. The preprocessing steps include:
[0110] Data cleaning: Remove duplicate records, invalid data, and obviously incorrect data from the original data. For example, delete records where the patient's age is less than the preset age or the number of dialysis sessions is less than the preset number. For hemodialysis patients, data of underage patients (such as those under 18 years old) and non-maintenance hemodialysis patients (such as those with less than 3 dialysis sessions) are usually excluded.
[0111] Missing value imputation: For missing data, use interpolation methods to complete time-series data and use Gaussian distribution filling method or mean filling method to complete non-time-series data. If the above methods are ineffective, delete the data. Specifically, for time-series data (such as blood pressure, heart rate, etc.), linear interpolation can be used; for non-time-series data (such as hemoglobin concentration, electrolyte levels, etc.), mean filling method or Gaussian distribution filling method can be used.
[0112] Outlier handling: Identify and handle outliers through statistical methods. For example, use the interquartile range (IQR) method to identify and remove outliers. Specifically, calculate the quartiles (Q1 and Q3) of each feature and define the upper and lower limits as
[0113] Q1 - 1.5×IQR and Q3 + 1.5×IQR, and data points outside this range are considered outliers and deleted.
[0114] Standardization: Standardize or normalize numerical data so that different features have the same scale, facilitating subsequent model training. For example, the Z-score standardization method can be used, that is, subtract the mean value of each feature from its value and then divide by its standard deviation.
[0115] 2. Build the first prediction model
[0116] The first prediction model can be an XGBoost model, and its hyperparameters are optimized through grid search. The specific steps are as follows:
[0117] Initialize the model: Select XGBoost as the initial model and set the initial parameters, such as the learning rate, the number of trees, and the maximum depth, etc. For example, the initial parameters can be set as the learning rate is 0.1, the number of trees is 100, and the maximum depth is 6.
[0118] Define the parameter grid: Define a parameter grid for the hyperparameters of the model, including different combinations of hyperparameters. For example, the learning rate can be set to 0.01 and 0.1, the number of trees can be set to 100 and 200, and the maximum depth can be set to 3 and 5.
[0119] Grid Search: Use the cross-validation method to traverse all combinations in the parameter grid to find the optimal hyperparameter configuration. For example, use 5-fold cross-validation (K-Fold Cross Validation) to evaluate the performance of each hyperparameter combination and select the combination with the best performance. During cross-validation, 4 / 5 of the data is used for training each time, and the remaining 1 / 5 of the data is used for validation. Finally, the average performance is taken as the performance of the model.
[0120] 3. Determine the target historical physiological sample data and the first training parameter
[0121] By analyzing the results of the first prediction model, select the historical physiological sample data corresponding to the peaks and valleys as the target samples, and calculate the corresponding training parameters. The specific steps are as follows:
[0122] Peak and valley detection: Identify the positions of peaks and valleys in the model output. These positions usually represent important feature points in the data. For example, the peak and valley can be determined by local maximum and minimum detection algorithms. Specifically, the find_peaks function in the SciPy library can be used to detect peaks and valleys.
[0123] Extract the target sample data: Use the historical physiological sample data corresponding to the peaks and valleys as the target sample data. These data contain important information and help improve the prediction ability of the model.
[0124] Calculate feature importance: Calculate the importance of each feature according to the feature importance score of the model, so as to determine the first training parameter. For example, the feature importance scoring function provided by the XGBoost model can be used to calculate the importance of each feature. Specifically, the model.feature_importances_ attribute can be used to obtain the importance score of each feature.
[0125] 4. Build the second prediction model (CNN-LSTM)
[0126] Based on the target historical physiological sample data and the first training parameter, build the second prediction model - the CNN-LSTM model. The specific steps are as follows:
[0127] Convolutional Layers: Used to extract the spatial features of time series data. The convolutional layer can capture local time patterns and reduce the input dimension. For example, use 64 filters, the kernel size is 3, and the activation function is ReLU. Specifically, the Conv1D layer in the TensorFlow or PyTorch framework can be used to implement the convolution operation.
[0128] LSTM Layers (Long Short-Term Memory Layers): Used to capture long-term dependencies in time series data. LSTM layers can effectively process data with long time intervals and maintain the memorability of information. For example, 50 LSTM units are used, and the activation function is ReLU. Specifically, the LSTM operation can be implemented using the LSTM layer in the TensorFlow or PyTorch framework.
[0129] Fully Connected Layers: Used to output the final prediction result. The fully connected layer integrates the features extracted by the previous layers to generate the final predicted value. For example, a fully connected layer is used, the activation function is Sigmoid, and a single probability value is output. Specifically, the fully connected operation can be implemented using the Dense layer in the TensorFlow or PyTorch framework.
[0130] 5. Determine the second target parameter
[0131] Determine the optimal parameters of the second prediction model through K-Fold Cross Validation. The specific steps are as follows:
[0132] Divide the dataset: Divide the dataset into K subsets. Each time, use K - 1 subsets for training, and the remaining one subset is used for validation. For example, 5-fold cross-validation is used.
[0133] Model training and validation: Repeat the training and validation process K times. Each time, use a different subset as the validation set, and finally take the average performance as the performance of the model. For example, during each training process, use the Adam optimizer and the Binary Cross-Entropy Loss function for model training.
[0134] Determine the optimal parameters: According to the results of cross-validation, select the model parameters with the best performance as the second target parameters. For example, select the model parameters with the best performance on the validation set as the final second target parameters.
[0135] 6. Adjust the second prediction model
[0136] Adjust the second prediction model according to the first target parameters and the second target parameters to obtain the final low blood pressure risk prediction model. The specific steps are as follows:
[0137] Calculate the difference: Calculate the difference between the first target parameters and the second target parameters. For example, the Euclidean distance between two parameter vectors can be calculated.
[0138] Determine whether the difference is within a preset range: If the difference is within the preset range, it is considered that the model training is completed; otherwise, continue to adjust the model parameters. For example, set a threshold, and when the difference is less than the threshold, it is considered that the model training is completed.
[0139] Update the training set: When the difference is not within the preset range, add the new target parameters to the training set and retrain the model until the condition is met. For example, add the second target parameter to the training set and retrain the model until the difference meets the preset range.
[0140] Online learning mechanism
[0141] To enable the model to adapt to new data, an online learning mechanism is supported, that is, the model can update parameters in real time according to newly collected historical physiological data. The specific implementation methods include:
[0142] Data stream processing: Continuously receive new historical physiological data and incorporate it into the model training process. For example, collect new physiological data regularly every day and add it to the training set.
[0143] Incremental learning: Gradually update the model parameters without retraining the entire model to make it adapt to new data changes. For example, use incremental learning algorithms. When new data is received each time, only update some model parameters instead of retraining the entire model. Specifically, online learning algorithms such as SGD (Stochastic Gradient Descent) or AdaGrad can be used to gradually update the model parameters.
[0144] Multimodal data fusion
[0145] Historical physiological data includes but is not limited to electrocardiogram (ECG), blood oxygen saturation (SpO2), and body temperature (Temperature). The fusion of multimodal data helps to improve the generalization ability and accuracy of the model. The specific steps are as follows:
[0146] Data collection: Collect different types of data from multiple sensors or devices, such as ECG, SpO2, and body temperature, etc. For example, use an electrocardiograph to collect ECG data, use a pulse oximeter to collect SpO2 data, and use a thermometer to collect body temperature data.
[0147] Data synchronization: Synchronize different types of data in terms of time to ensure that they are aligned at the same time point. For example, use timestamps to align different types of physiological data. Specifically, the merge_asof function in the Pandas library can be used for time synchronization.
[0148] Feature extraction: Extract meaningful features from each type of physiological data, such as the R-wave position in ECG, the oxygenation level in SpO2, etc. For example, extract the R-wave peak position from ECG data and the oxygenation level from SpO2 data. Specifically, signal processing algorithms such as the Pan-Tompkins algorithm can be used to extract the R-wave position in ECG.
[0149] Data fusion: Fuse the extracted features to form a comprehensive feature vector as the input to the model. For example, concatenate the feature vectors of ECG, SpO2, and body temperature to form a new feature vector as the input to the model. Specifically, the concatenate function in the NumPy library can be used for feature vector concatenation.
Claims
1. A real-time hypotension risk prediction method based on a composite neural network, characterized in that: Includes steps: Constructing a first prediction model, wherein the first prediction model is obtained by training the patient's historical physiological data and resource sample data corresponding to the historical physiological data; Determine target historical physiological sample data and first training parameters corresponding to the target historical physiological sample data based on the first prediction model; Determining a first target parameter corresponding to the first prediction model based on the first training parameter; Based on the target historical physiological sample data and the first training parameters, construct a second prediction model, wherein the second prediction model is a CNN-LSTM model; Determining a second target parameter corresponding to the second prediction model; Based on the first target parameter and the second target parameter, the second prediction model is adjusted to obtain a final hypotension risk prediction model.
2. The real-time hypotension risk prediction method according to claim 1, characterized in that: Before constructing the first prediction model, the step of preprocessing the historical physiological data is also included, and the preprocessing includes: Delete the clinical data of the target subjects whose age is younger than the preset age and the clinical data of the target subjects who have undergone hemodialysis less than the preset number of times; Divide the target clinical data into time series data and non-time series data; Interpolation method is used to complete the missing time series data, and Gaussian distribution filling method or mean filling method is used to complete the missing non-time series data, and invalid data after completion is deleted from the target clinical data; Correct the time series data or non-time series data with abnormal data types and abnormal values, and delete the corrected invalid data from the target clinical data.
3. The real-time hypotension risk prediction method according to claim 1, characterized in that: The determining, based on the first prediction model, target historical physiological sample data and a first training parameter corresponding to the target historical physiological sample data includes: Determine at least one peak and at least one trough of the first prediction model; use the historical physiological sample data corresponding to the at least one peak and the at least one trough as the target historical physiological sample data; The target historical physiological sample data are respectively input into a first preset calculation formula, and the first training parameter is obtained through the first preset calculation formula.
4. The real-time hypotension risk prediction method according to claim 3, characterized in that: The first preset calculation formula is: Where T is the number of trees, Δ i is the split gain caused by a feature in the i-th tree, and N is the number of samples.
5. The real-time hypotension risk prediction method according to claim 1, characterized in that: The determining, based on the first training parameter, a first target parameter corresponding to the first prediction model includes: Based on a second preset calculation formula, determining a first target value from the first training parameter, the first target value being used to indicate the level of robustness; Based on a first prediction model, the first target parameter corresponding to the first target value is determined.
6. The real-time hypotension risk prediction method according to claim 5, characterized in that: The robustness evaluation formula is: in, is the loss between the original predicted value and the true value, is the loss between the predicted value and the true value after adding noise.
7. The real-time hypotension risk prediction method according to claim 1, characterized in that: The constructing a second prediction model based on the target historical physiological sample data and the first training parameter includes: Combining the target historical physiological sample data to obtain a first set; combining the first training parameters to obtain a second set; Fitting to obtain a mapping relationship between the first set and the second set; Based on the target historical physiological sample data, the first training parameter and the mapping relationship, a second prediction model is constructed, wherein the second prediction model is a CNN-LSTM model, and the CNN-LSTM model includes: Convolutional layers, used to extract spatial features of time series data; LSTM layer, used to capture long-term dependencies of time series data; The fully connected layer is used to output the final prediction result.
8. The real-time hypotension risk prediction method according to claim 7, characterized in that: The convolution layer uses 64 filters, the kernel size is 3, and the activation function is ReLU; the LSTM layer uses 50 LSTM units, and the activation function is ReLU; the fully connected layer uses the Sigmoid activation function and outputs a single probability value.
9. The real-time hypotension risk prediction method according to claim 1, characterized in that: The determining a second target parameter corresponding to the second prediction model includes: Based on the second prediction model, taking the position corresponding to the maximum value as the second target parameter; K-fold cross validation was used during training, where K was 5 or 10.
10. The real-time hypotension risk prediction method according to claim 1, characterized in that: The final hypotension risk prediction model supports an online learning mechanism and can update model parameters in real time according to newly collected historical physiological data, and the historical physiological data includes multimodal data, which includes but is not limited to electrocardiogram, blood oxygen saturation and body temperature.
Citation Information
Cited By
Intraoperative hypotension early warning method based on FGL and multi-modal physiological parameters
CN122392979A