A blood pressure early warning method, device and storage medium

Abstract

The present invention relates to the field of machine learning, and discloses a blood pressure early warning method, device and storage medium, which are specifically: obtaining data for blood pressure prediction; preprocessing the data; selecting based on the preprocessed data features to form a blood pressure prediction data set; constructing multiple blood pressure prediction models, and using the blood pressure prediction data set for training; evaluating the performance of the trained model, and obtaining the output result of the integrated model by soft voting; evaluating the similarity between the output result of the integrated model and the output result of the trained model, and selecting the trained model with the highest similarity; using an explainable AI tool to interpret the decision, and obtaining the input features that affect the final decision; selecting the top P features that affect the final model detection result, and constructing a new prediction set; using the new prediction set to train the blood pressure prediction model, and then obtaining the final prediction result by soft voting. The present invention can make early predictions of blood pressure.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a blood pressure early warning method, device and storage medium. Background Art

[0002] Studies have shown that blood pressure fluctuations during dialysis show a "U"-shaped relationship with patient mortality. This fluctuation will not only affect the blood supply to the patient's heart, brain and other important organs, increase the risk of bleeding, but may also affect the service life of the dialysis patient's arteriovenous fistula. However, maintenance hemodialysis patients often have complex conditions, multiple comorbidities, and significant hemodynamic changes during the dialysis process. Traditional modeling methods are difficult to cope with, and it is necessary to develop a blood pressure prediction model for the dialysis population.

[0003] Currently, there is no blood pressure prediction tool in clinical practice. Most studies still focus on the construction of prediction models for hypotension during dialysis using machine learning, and there is no integrated prediction model covering hypotension and hypertension.

[0004] Secondly, the included influencing factors are mostly dialysis treatment data and test data, and the accuracy and applicability of the prediction model are still not ideal, with a high risk of bias.

[0005] Third, most models cannot give early warning prompts before dialysis. Existing predictive analysis is mainly based on real-time prediction of blood pressure during dialysis. However, this prediction result is effective in preventing adverse events in advance, but its clinical application prospects are not high.

[0006] Finally, the decision-making models currently used are all single models, which are prone to overfitting and poor robustness to data noise and outliers. It is impossible to specifically learn the intrinsic relationship between the complex clinical situation of uremic hemodialysis patients and blood pressure fluctuations. Summary of the invention

[0007] The purpose of the present invention is to propose a blood pressure early warning method, device and storage medium to solve the technical problem that the existing technology cannot provide an effective and accurate blood pressure prediction model for blood pressure prediction.

[0008] A blood pressure early warning method comprises the following steps:

[0009] S1: Obtain relevant data for blood pressure prediction;

[0010] S2: preprocess the relevant data to obtain preprocessed data;

[0011] S3: Perform feature selection based on the preprocessed data to screen important data that affect blood pressure prediction and form a blood pressure prediction data set;

[0012] S4: construct multiple blood pressure prediction models, use the blood pressure prediction data set for training, and obtain the trained models;

[0013] S5: Evaluate the performance of the trained model and obtain the output of the integrated model through soft voting;

[0014] S6: Evaluate the similarity between the output results of the integrated model and the output results of the trained model, and select the trained model with the highest similarity;

[0015] S7: Use explainable AI tools to interpret the prediction results of the most similar trained model and obtain the input features that affect the final decision;

[0016] S8: Select the first P features of the input features that affect the final model detection results and construct a new prediction data set;

[0017] S9: Use the new prediction data set to retrain multiple blood pressure prediction models, and again use soft voting to obtain the final prediction results to complete early blood pressure prediction.

[0018] A storage medium stores instructions and data for implementing a blood pressure early warning method.

[0019] A blood pressure early warning device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a blood pressure early warning method.

[0020] The beneficial effects provided by the present invention are:

[0021] 1. Use different machine learning models as base models to implement the ensemble learning method of soft voting. By combining the prediction results of multiple learning algorithms, the performance of the overall model is improved, the impact of overfitting is reduced, and the prediction accuracy of the model is significantly improved.

[0022] 2. Use the SHAP method of the explainable AI tool to quantify the contribution of features to model decisions, reveal the key clinical features that play a decisive role in the prediction of abnormal blood pressure, and provide medical professionals with a transparent, intuitive and in-depth understanding;

[0023] 3. In-depth data support for multi-factor blood pressure prediction, taking into account multiple dimensions such as demographics, primary disease, dialysis characteristics and laboratory test results;

[0024] 4. Fully consider the impact of hypoproteinemia on total serum calcium. Incorporate a correction calcium formula to identify patients with hypoproteinemia, correct calcium levels, and reduce model prediction error factors;

[0025] 5. Two models, A and B, are designed for different usage scenarios: Model A integrates more difficult-to-obtain features, aiming to improve prediction accuracy through complex data input; Model B focuses on easy-to-obtain features, and finally integrates the model results, using explainable AI tools to reversely interpret the results, obtain accurate input features, and further improve the model's prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic flow chart of the method of the present invention;

[0027] Figure 2 This is an explanation diagram of the interpretable AI analysis model in an embodiment of the present invention predicting the occurrence of hypotension during the dialysis process of a user;

[0028] Figure 3 It is a summary diagram of the feature importance of the more important features for the low blood pressure prediction category obtained based on the explainable AI analysis in an embodiment of the present invention;

[0029] Figure 4 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0030] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0031] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0032] Please refer to Figure 1 , Figure 1 It is a schematic flow diagram of the method of the present invention;

[0033] The present invention provides a blood pressure early warning method, comprising the following steps:

[0034] S1: Obtain relevant data for blood pressure prediction;

[0035] It should be noted that the present invention extracts relevant data for blood pressure prediction from a relevant hemodialysis management system or an electronic health record system.

[0036] S2: preprocess the relevant data to obtain preprocessed data;

[0037] It should be noted that the preprocessing means in step S2 include: deleting missing values, outliers, performing calcium correction, and normalization coding.

[0038] As an embodiment, the present invention adopts a strict data screening process and deletes samples containing any missing values ​​from the remaining data.

[0039] In the data set, all missing data are considered to occur randomly, and no processing is performed on these missing values, which are directly deleted to keep the original characteristics of the data unaffected.

[0040] Secondly, considering that some patients may have combined hypoproteinemia, in order to more accurately reflect the actual level of total serum calcium, the present invention uses a calcium correction formula, i.e., corrected calcium = measured calcium + (40-measured albumin) × 0.02, to make necessary corrections to the calcium content in samples combined with hypoproteinemia.

[0041] Finally, in order to handle categorical variables such as gender, comorbidities, lifestyle, and converted laboratory data, this study adopted the One-Hot Encoding method to convert these features into binary form, that is, using binary variables of 0 and 1 to represent them, so as to facilitate effective processing during model training.

[0042] S3: Perform feature selection based on the preprocessed data to screen important data that affect blood pressure prediction and form a blood pressure prediction data set;

[0043] It should be noted that in step S3, random forest is used to sort the features in the preprocessed data, and the top N features are selected to construct the first blood pressure prediction data set; M easily accessible features are selected to construct the second blood pressure prediction data set.

[0044] As an example, random forest is used to rank the importance of 125 features.

[0045] Table 1 is a classification table of some features of the data set in the embodiment of the present invention

[0046]

[0047]

[0048] Select the first 51 features to construct model A. Specifically, construct a first blood pressure prediction data set with the first 51 features, and use the first blood pressure prediction data set to train the model;

[0049] It should be noted that the model constructed using the first 51 features is a model with high accuracy in predicting blood pressure.

[0050] In addition, 10 easily available features are screened to construct model B; it should be noted that model B is a prediction model with high ease of use, which increases the applicability range of blood pressure prediction.

[0051] As an embodiment, the first 51 features of the present invention use on-line data (including on-line mean arterial pressure, on-line systolic pressure, on-line diastolic pressure), patient dialysis information (pre-dialysis weight, dialysis age, dialysis mode, preset ultrafiltration volume, etc.), laboratory test data (lactate dehydrogenase, sodium, calcium, etc.) and cardiac ultrasound data (left atrial diameter, ascending aorta diameter, etc.) and medication status (calcium channel blockers, angiotensin II receptor antagonists, etc.); 10 easily accessible features use on-line data, ultrafiltration rate, dialysis mode, gender, etc.;

[0052] S4: construct multiple blood pressure prediction models, use the blood pressure prediction data set for training, and obtain the trained models;

[0053] It should be noted that the multiple blood pressure prediction models include: logistic regression model, random forest model, multi-layer perceptron model and extreme gradient boosting tree model. For the convenience of description, the present invention will be described in the following using C, D, E, and F models.

[0054] In the present invention, the first blood pressure prediction data and the second blood pressure prediction data set are used to train multiple blood pressure prediction models, respectively, to obtain the trained first type of multiple models and the second type of multiple models.

[0055] The present invention adopts the K-fold cross-validation method for training; the K-fold cross-validation method is not the focus of the present invention and will not be described in detail here.

[0056] Combined with the previous text, the trained models can also be divided into two categories, namely Class A and Class B. The corresponding trained model can be recorded as C A -F A Model and C B -F B Model;

[0057] Performing performance evaluation on the first type of multiple models and the second type of multiple models, and screening them to obtain the screened first type of multiple models and the second type of multiple models;

[0058] That is, for C A -F A Model and F B -F B The model is evaluated for performance and screened; as an embodiment, the present invention uses the method of calculating the ROC curve or the AUROC index to perform performance evaluation, and the models that do not meet the preset conditions or preset values ​​of the index are eliminated;

[0059] As an example, for example, from C A -G A In the model, C A , D A 、E AModel, from C B -G B In the model, D B 、E B and F B Model;

[0060] S5: Evaluate the performance of the trained model and obtain the output of the integrated model through soft voting;

[0061] The first type of multiple models and the second type of multiple models are respectively subjected to ensemble learning soft voting to obtain the prediction results of the first type of ensemble model and the prediction results of the second type of ensemble model.

[0062] It should be noted that the implementation principle of ensemble learning soft voting is: each model independent The input is predicted by each model (each model here is called a base model) and the prediction probability of each category is output. Soft voting calculates the average prediction probability output by each base model to obtain a comprehensive and robust prediction probability in order to improve the prediction accuracy.

[0063] Specifically, the soft voting method is used according to C A , D A 、E A The output of the model and D B 、E B and F B The output results of the model obtained two optimal output results.

[0064] S6: Evaluate the similarity between the output results of the integrated model and the output results of the trained model, and select the trained model with the highest similarity;

[0065] It should be noted that the similarity between the prediction results of the first type of integrated model and the output results of the first type of multiple models is evaluated, and the first trained model with the highest similarity is selected;

[0066] Specifically, for example, the prediction results of the first type of ensemble model are similar to those of E A The output results of the model are most similar, indicating that E A The model with the highest similarity is the first trained model.

[0067] S7: Use explainable AI tools to interpret the prediction results of the screened model and obtain the input features that affect the prediction results of the screened model;

[0068] As an example, using explainable AI tools to AExplain the model result H and infer the multiple input features that affect the prediction results and their influence degree;

[0069] Specifically, it is to use explainable AI tools to explain the prediction result H, and infer the multiple input features that affect the prediction result H and the degree of their influence;

[0070] S8: Select the first P features of the input features that affect the final model detection results and construct a new prediction data set;

[0071] Filter the first P input features from front to back according to the degree of influence to form a new prediction data set;

[0072] It should be noted that the new prediction data set composed of the P input features selected at this time can make the trained model more accurate.

[0073] S9: Use the new prediction data set to retrain multiple blood pressure prediction models, and again use soft voting to obtain the final prediction results to complete early blood pressure prediction.

[0074] In the present invention, a new prediction data set is used to retrain multiple blood pressure prediction models, and the final prediction result is obtained again through soft voting to complete the early prediction of blood pressure.

[0075] As an example, a new prediction data set is selected to train C A -D A The model is then used to obtain the final prediction result through soft voting.

[0076] As for the second type of model, since it is a feature selected for ease of use, in the present invention, it does not need to be trained with a new data set. The main consideration is ease of use, not accuracy.

[0077] Please refer to Table 2, which is a comparison table of AUROC scores of the ensemble learning model and the corresponding base model in an embodiment of the present invention.

[0078] Table 2 Comparison of AUROC scores between ensemble learning model and corresponding base model

[0079]

[0080] From Table 2, we can see that the integrated model has achieved improved scores in both types of models, significantly improving the prediction accuracy of the model.

[0081] Please refer to Figure 2 , Figure 2 This is an explanatory diagram of the explainable AI analysis model of an embodiment of the present invention predicting the occurrence of hypotension during the dialysis process of a user.

[0082] The prediction results are consistent with the true labels of the samples, among which the user's preset ultrafiltration volume, dialysis age, mean arterial pressure before the machine and other variables have a positive impact on the occurrence of hypotension during dialysis (red blocks), while lactate dehydrogenase and creatine kinase have a negative impact on the risk of hypotension (blue blocks). The size of the block reflects the relative influence of the feature value on the prediction result: the larger the block, the greater the impact on the model decision output. In the subsequent follow-up, the uncontrollable factors were put aside, and by limiting the user's rapid weight gain between dialysis periods, adjusting the preset ultrafiltration, re-evaluating the user's dry weight, and adjusting the patient's antihypertensive drug dosage, the user's blood pressure is currently well controlled, and there is no occurrence of hypertension during dialysis.

[0083] Please refer to Figure 3 , Figure 3 This is a summary diagram of the feature importance of the more important features for the low blood pressure prediction category obtained by the embodiment of the present invention based on the explainable AI analysis. It provides the average impact of each feature on the blood pressure prediction model for hemodialysis patients from a global perspective, where each point represents a different sample, the color represents the actual value of the feature, and the positive feature SHAP value indicates that the feature has a positive contribution to the prediction result, which increases the prediction value. On the contrary, a negative feature SHAP value indicates that the feature has a negative contribution to the prediction result, which reduces the prediction value. The features are sorted according to the average absolute SHAP value of the feature samples, which shows the degree of influence of the feature on the prediction of the corresponding category. It can be observed that features such as on-machine mean arterial pressure, on-machine systolic pressure, on-machine diastolic pressure, and lactate dehydrogenase are more important factors in determining the two prediction categories.

[0084] See also Figure 4 , Figure 4 It is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a blood pressure early warning device 401, a processor 402 and a storage medium 403.

[0085] A blood pressure early warning device 401: The blood pressure early warning device 401 implements the blood pressure early warning method.

[0086] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the blood pressure early warning method.

[0087] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the blood pressure early warning method.

[0088] The beneficial effects of the present invention are:

[0089] 1. Use different machine learning models as base models to implement the ensemble learning method of soft voting. By combining the prediction results of multiple learning algorithms, the performance of the overall model is improved, the impact of overfitting is reduced, and the prediction accuracy of the model is significantly improved.

[0090] 2. Use the SHAP method of the explainable AI tool to quantify the contribution of features to model decisions, reveal the key clinical features that play a decisive role in the prediction of abnormal blood pressure, and provide medical professionals with a transparent, intuitive and in-depth understanding;

[0091] 3. In-depth data support for multi-factor blood pressure prediction, taking into account multiple dimensions such as demographics, primary disease, dialysis characteristics and laboratory test results;

[0092] 4. Fully consider the impact of hypoproteinemia on total serum calcium. Incorporate a correction calcium formula to identify patients with hypoproteinemia, correct calcium levels, and reduce model prediction error factors;

[0093] 5. Two models, A and B, are designed for different usage scenarios: Model A integrates more difficult-to-obtain features, aiming to improve prediction accuracy through complex data input; Model B focuses on easy-to-obtain features, uses explainable AI tools to reverse-interpret the results, obtains accurate input features, and further improves the model's prediction accuracy.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A blood pressure early warning method, characterized in that: The method comprises the following steps: S1: Obtain relevant data for blood pressure prediction; S2: preprocess the relevant data to obtain preprocessed data; S3: Perform feature selection based on the preprocessed data to screen important data that affect blood pressure prediction and form a blood pressure prediction data set; In step S3, random forest is used to sort the features in the preprocessed data, and the top N features are selected to construct a first blood pressure prediction data set; M easily available features are selected to construct a second blood pressure prediction data set, where N and M are preset values; S4: construct multiple blood pressure prediction models, use the blood pressure prediction data set for training, and obtain the trained models; Using the first blood pressure prediction data and the second blood pressure prediction data set to train multiple blood pressure prediction models respectively, to obtain a first type of multiple models and a second type of multiple models that have been trained; S5: Evaluate the performance of the trained model and obtain the output of the integrated model through soft voting; S6: Evaluate the similarity between the output results of the integrated model and the output results of the trained model, and select the trained model with the highest similarity; S7: Use explainable AI tools to interpret the prediction results of the most similar trained model and obtain the input features that affect the final decision; S8: Select the first P features of the input features that affect the final model detection results and construct a new prediction data set; Steps S5 to S8 are specifically as follows: Performing performance evaluation on the first type of multiple models and the second type of multiple models, and screening them to obtain the screened first type of multiple models and the second type of multiple models; The prediction results of the first type of integrated model and the second type of integrated model are obtained by using an integrated learning soft voting method for the screened first type of multiple models and the second type of multiple models respectively; Evaluate the similarity between the prediction results of the first type of integrated model and the output results of the first type of multiple models, and select the first trained model with the highest similarity; Use explainable AI tools to explain the output results of the first trained model, and infer multiple input features that affect the prediction results and their degree of influence; Filter the first P input features from front to back according to the degree of influence to form a new prediction data set; Use the new prediction data set to retrain multiple blood pressure prediction models, and then use soft voting to obtain the final prediction results to complete the early prediction of blood pressure. S9: Use the new prediction data set to retrain multiple blood pressure prediction models, and again use soft voting to obtain the final prediction results to complete early blood pressure prediction.

2. A blood pressure early warning method as claimed in claim 1, characterized in that: The preprocessing means in step S2 include: deleting missing values, outliers, performing calcium correction, and normalization coding.

3. A blood pressure early warning method as claimed in claim 2, characterized in that: The calcium correction is as follows: a=b+(40-c)×0.02 Wherein, a represents the corrected calcium content in serum; b represents the actual calcium content in serum; and c represents the actual albumin content in serum.

4. A blood pressure early warning method as claimed in claim 2, characterized in that: The normalized encoding is performed using a one-hot encoding method.

5. A blood pressure early warning method as claimed in claim 1, characterized in that: The multiple blood pressure prediction models include: logistic regression model, random forest model, multi-layer perceptron model and extreme gradient boosting tree model.

6. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a blood pressure early warning method as described in any one of claims 1 to 5.

7. A blood pressure early warning device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a blood pressure early warning method as described in any one of claims 1 to 5.

Images