A method and system for intelligent prediction and prevention of pressure injuries

Abstract

The present invention relates to an intelligent prediction and prevention method and system for pressure injuries, including determining target comprehensive monitoring data of a patient, the target comprehensive monitoring data including microenvironmental data indicating body surface temperature, body surface humidity, and pressure on a pressured part, as well as individual characteristic data indicating the patient's gender, age, and whether or not they suffer from diabetes; analyzing the comprehensive monitoring data using a machine learning model based on ensemble learning to obtain a risk score, the machine learning model including multiple basic learners, each of which is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight based on the score of a performance evaluation indicator. The prediction results of the ensemble model are obtained by performing a weighted average combination based on the prediction results of all the basic learners; and quantitatively evaluating the risk of pressure injuries at the pressured part of the patient based on the risk score to generate a risk report.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a method and system for intelligently predicting and preventing pressure injuries. Background Art

[0002] Pressure injuries, also known as pressure ulcers, are a common complication caused by prolonged pressure on localized tissue, leading to circulatory impairment and nutritional deficiencies, resulting in localized tissue ulceration and necrosis. These injuries not only increase patient suffering and prolong hospitalizations, but can also lead to serious consequences such as infection and sepsis, and can even threaten the patient's life. During surgery, where patients are in a specific position and often unable to move independently, pressure injuries are a high-risk period.

[0003] Traditional pressure injury risk assessment relies primarily on the experience and intuition of medical staff, which is subject to significant subjectivity and makes it difficult to ensure accurate and consistent assessments. With the development of machine learning, researchers have begun exploring the use of machine learning algorithms to predict and assess pressure injury risk. For example, prediction models are constructed using algorithms such as support vector machines and random forests. While this approach can learn the complex relationships between risk and multiple factors based on large amounts of historical data, improving prediction accuracy, single algorithms often have limitations, such as overfitting, weak generalization, or poor adaptability to new data. Therefore, it is necessary to combine multiple machine learning algorithms for pressure injury risk assessment from an ensemble learning perspective to improve the accuracy of pressure injury prediction. Summary of the Invention

[0004] In order to improve the accuracy of pressure injury prediction, the present invention aims to provide a method and system for intelligent pressure injury prediction and prevention. The technical solutions adopted are as follows:

[0005] In a first aspect, the present application discloses a method for intelligent prediction and prevention of pressure injuries, comprising:

[0006] S1. Determining target comprehensive monitoring data for the patient, wherein the target comprehensive monitoring data includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on a pressure-receiving part, as well as individual characteristic data indicating the patient's gender, age, and whether or not the patient has diabetes;

[0007] S2. Analyzing the target comprehensive monitoring data using an ensemble learning-based machine learning model to obtain a risk score, wherein the machine learning model includes multiple base learners, each of which is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight based on the score of the performance evaluation indicator. The prediction results of the ensemble model are obtained by performing a weighted average combination based on the prediction results of all the base learners;

[0008] S3. Based on the risk score, a quantitative assessment is performed on the risk of pressure injury at the patient's compressed area, and a risk report is generated.

[0009] Furthermore, in step S1, determining the target comprehensive monitoring data of the patient includes:

[0010] S11. Obtain the patient's initial comprehensive monitoring data;

[0011] S12. Based on the initial comprehensive monitoring data, remove invalid data through multi-dimensional verification to obtain valid monitoring data, wherein the multi-dimensional verification includes a numerical range verification and a time continuity verification for the microenvironment data, and a data consistency verification and a logical rationality verification for the individual feature data;

[0012] S13. Based on the valid monitoring data, performing missing value checking and filling processing through a missing value prediction and filling model to obtain reliable monitoring data;

[0013] S14. Based on the reliable monitoring data, perform standardization processing on numerical data and categorical data to obtain target comprehensive monitoring data.

[0014] Furthermore, in step S13, the missing value prediction and filling model is composed of a mask-based adaptive convolutional neural network, a long short-term memory network, and a self-attention and cross-attention mechanism, wherein:

[0015] The mask-based adaptive convolutional neural network serves as a feature extractor and includes a mask estimation module, multiple parallel adaptive convolutional layers, and a feature fusion and output module, wherein: the mask estimation module is used to generate a mask matrix corresponding to the input data based on context information, and the mask matrix marks the estimated positions of missing values ​​in the data; each of the adaptive convolutional layers is used to dynamically adjust the weight of the convolution kernel based on the mask matrix, combining the distribution characteristics of the estimated positions of missing values ​​and the characteristics of the convolution kernel itself, so that the convolution kernel can reduce the weight of the estimated positions of missing values ​​marked by the mask matrix when processing these positions, thereby avoiding the negative impact of missing values ​​on feature extraction; the feature fusion and output module is used to fuse the features extracted from each adaptive convolutional layer and generate a final feature representation;

[0016] The long short-term memory network acts as a temporal dependency capturer and feature fusion processor, and is used to process time series data using its memory capacity and temporal analysis capabilities, further capturing temporal dependencies in the data, and integrating feature attributes associated with each time series monitoring value;

[0017] The self-attention and cross-attention mechanisms serve as information focusers and fusers. On the one hand, they allow the model to calculate the correlation between each point in time and other points based on the contextual information of the entire time series when processing the data of each time point, so that the model can more accurately identify which time points’ data are most critical for predicting missing values, thereby enhancing the robustness of the long short-term memory network in capturing temporal dependencies. On the other hand, they allow the model to simultaneously focus on the feature attribute information associated with each monitoring value when processing time series data, and help the long short-term memory network better integrate information between different data sources by calculating the correlation between elements of time series data and feature attribute data, thereby generating more comprehensive and accurate prediction results.

[0018] Furthermore, in step S2, the basic learner includes at least one of a decision tree, a support vector machine, a random forest, and a neural network.

[0019] Furthermore, for each base learner, in step S2, the performance of each base learner is evaluated based on a method combining cross-validation and feature importance evaluation, including:

[0020] S21. Perform feature selection based on statistical analysis and correlation selection to obtain a preliminarily screened feature set;

[0021] S22. Performing a feature importance evaluation on each feature in the feature set after the preliminary screening using a Lasso regression algorithm to obtain a feature importance evaluation result;

[0022] S23. Based on the feature importance evaluation results, the most important target feature subset is screened from the feature set after the preliminary screening by sorting and thresholding methods;

[0023] S24. According to the size of the target feature subset, a K-fold cross-validation method is selected to train and validate the base learner, wherein, in each cross-validation cycle, K-1 pieces of training data are used for training, and the remaining 1 piece of data is used for validation;

[0024] S25. Repeat the cross-validation process K times to obtain K independent performance evaluation results, and obtain a robust estimate of the performance of the base learner based on the mean and standard deviation.

[0025] Furthermore, in step S2, each basic learner is assigned a corresponding weight according to the score of the performance evaluation index. After the prediction results of all basic learners are combined by weighted average, the prediction results of the integrated model can be obtained, including:

[0026] S26. Based on the performance evaluation index score of each base learner in the cross-validation process, the weight of each learner is calculated using a multi-objective optimization technique;

[0027] S27. For each sample to be predicted, use all basic learners to predict it and obtain their respective prediction results;

[0028] S28. According to the determined weight of each learner, a weighted average is performed based on the prediction results of each learner to obtain an integrated prediction result.

[0029] Furthermore, in step S27, after obtaining the prediction result, the method further includes:

[0030] S271. Calculate the prediction accuracy and stability of each learner on the test sample;

[0031] S272: When it is determined that the accuracy and / or stability of the corresponding learner is greater than a preset threshold, the weight of the learner is increased according to the set weight adjustment coefficient; otherwise, the weight of the learner is decreased;

[0032] S273. When it is determined that the accuracy of the corresponding learner drops below a preset threshold and / or the stability fluctuation exceeds a preset range, the weight of the learner is recalculated and adjusted according to the evaluation result to ensure the prediction performance and stability of the integrated model.

[0033] Furthermore, in step S3, the risk of pressure injury at the patient's compressed part is quantitatively assessed based on the risk score, and a risk report is generated, including:

[0034] S31. Classify the patient's risk score into different risk levels based on a preset risk score threshold range, and generate preventive measures recommendations corresponding to the risk level in conjunction with the intelligent preventive measures engine;

[0035] S32. Obtain the patient's basic information, the risk score of the pressure site, the risk level, the recommended preventive measures for the risk level, and the potential risk factors, and generate a risk report through an interactive report template generation system.

[0036] In a second aspect, the present application discloses an intelligent pressure injury prediction and prevention system, which includes a data acquisition module, a risk scoring module, and a risk quantification assessment module, wherein:

[0037] The data acquisition module is used to determine the target comprehensive monitoring data of the patient, wherein the target comprehensive monitoring data includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on the pressure-receiving part, and individual characteristic data indicating the patient's gender, age, and whether or not the patient has diabetes;

[0038] The risk scoring module is configured to analyze the target comprehensive monitoring data using a machine learning model based on ensemble learning to obtain a risk score, wherein the machine learning model includes multiple basic learners, each of which is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight based on the score of the performance evaluation indicator. The prediction results of the ensemble model can be obtained by performing a weighted average combination based on the prediction results of all the basic learners;

[0039] The risk quantification assessment module is used to quantitatively assess the risk of pressure injury at the patient's compressed area based on the risk score and generate a risk report.

[0040] In a third aspect, the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction and prevention method for pressure injuries.

[0041] The present invention has the following beneficial effects:

[0042] 1) By determining the patient's target comprehensive monitoring data, including microenvironmental data (body surface temperature, body surface humidity, pressure at the compressed area) and individual characteristic data (gender, age, diabetes status), comprehensive monitoring of the patient's risk of pressure injury at the compressed area can be achieved;

[0043] 2) Analyzing using an ensemble-based machine learning model can fully leverage the strengths of multiple base learners, combining their predictions through weighted averaging to improve overall prediction accuracy and stability. Furthermore, the base learners are pre-evaluated using a combination of cross-validation and feature importance assessment, with weights assigned based on the performance evaluation metric scores. This ensures that the strengths and weaknesses of the learners are considered during model construction, further improving prediction performance.

[0044] 3) By quantitatively assessing the risk of pressure injuries in patients' compressed areas, it can provide medical staff with clear risk indicators, helping them to take timely intervention measures to reduce the risk of pressure injuries in patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1A flowchart of a method for intelligent prediction and prevention of pressure injuries provided by one embodiment of the present invention;

[0047] Figure 2 This is a system structure diagram of a pressure injury intelligent prediction and prevention system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0048] To further illustrate the technical means and effectiveness of the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of a method and system for intelligently predicting and preventing pressure injuries according to the present invention. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0049] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0050] The specific scheme of the intelligent prediction and prevention method and system for pressure injuries provided by the present invention is described in detail below with reference to the accompanying drawings.

[0051] See also Figure 1 , which shows a method flow chart of a method for intelligent prediction and prevention of pressure injuries provided by one embodiment of the present invention, the method comprising:

[0052] Step S1, determining the target comprehensive monitoring data of the patient, wherein the target comprehensive monitoring data includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on the pressure-receiving part, and individual characteristic data indicating the patient's gender, age, and whether or not the patient has diabetes.

[0053] Step S2: Based on the target comprehensive monitoring data, a machine learning model based on ensemble learning is used for analysis to obtain a risk score, wherein the machine learning model includes multiple basic learners, each basic learner is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and a corresponding weight is assigned according to the score of the performance evaluation index. After the prediction results of all basic learners are weighted averaged and combined, the prediction result of the ensemble model can be obtained.

[0054] Step S3: quantitatively assessing the risk of pressure injury at the patient's compressed area based on the risk score, and generating a risk report.

[0055] As can be seen from the above, the intelligent prediction and prevention method for pressure injuries disclosed in this application can achieve comprehensive monitoring of the risk of pressure injuries in the patient's pressure area by determining the patient's target comprehensive monitoring data, including microenvironmental data (body surface temperature, body surface humidity, pressure on the pressure area) and individual characteristic data (gender, age, diabetes status); the use of a machine learning model based on ensemble learning for analysis can fully utilize the advantages of multiple basic learners and combine the prediction results of each learner by weighted average, thereby improving the accuracy and stability of the overall prediction. Moreover, the performance of the basic learners is pre-evaluated based on a method combining cross-validation and feature importance evaluation, and weights are assigned according to the scores of the performance evaluation indicators, ensuring that the advantages and disadvantages of the learners are taken into account when the model is constructed, further improving the prediction performance; by quantitatively evaluating the risk of pressure injuries in the patient's pressure area, it can provide medical staff with clear risk indicators, helping them to take timely measures to intervene and reduce the risk of pressure injuries in patients.

[0056] In one embodiment, in step S1, determining the target comprehensive monitoring data of the patient includes:

[0057] Step S11: Acquire the patient's initial comprehensive monitoring data.

[0058] Step S12, based on the initial comprehensive monitoring data, invalid data is removed through multi-dimensional verification to obtain valid monitoring data, wherein the multi-dimensional verification includes a numerical range verification and a time continuity verification for the microenvironment data, a data consistency verification and a logical rationality verification for the individual feature data.

[0059] Among them, the numerical range verification and time continuity verification of microenvironmental data include: 1) checking whether the microenvironmental data falls within a reasonable physiological or equipment measurement range, and marking the data outside the range as invalid data; 2) considering the importance of data integrity and quality assessment, the microenvironmental data can be checked for temporal continuity to screen out invalid data.

[0060] Data consistency and logical rationality checks for individual characteristic data include: 1) checking whether individual characteristic data is consistent across different records or data sources. Inconsistent data is marked as invalid; 2) checking whether individual characteristic data is reasonable based on medical common sense and logical rules. For example, age should be a positive number, gender should be a valid option (male, female, etc.), and diabetes status should be consistent with the patient's other medical records.

[0061] In one embodiment, the screened invalid data may be displayed on a visual interface, and after further confirmation by the user, the invalid data may be removed or corrected.

[0062] Step S13: Based on the effective monitoring data, missing value checking and filling processing are performed through a missing value prediction and filling model to obtain reliable monitoring data.

[0063] Specifically, this application constructs training data based on valid monitoring data. The format of the training data is a time series + feature attribute matrix. That is, the training data includes not only the monitoring values ​​arranged in a time series, but also the feature attributes associated with each monitoring value (such as environmental parameters and device status). These feature attributes help the model more accurately predict and fill missing values. The training data is then fed into the missing value prediction and filling model for model training.

[0064] Finally, new, valid monitoring data is processed based on the trained model. During this process, the model first checks for missing values ​​in the data, then uses learned patterns and rules to predict these missing values ​​and recommends appropriate filling methods (i.e., specific filling values ​​or filling strategies). This makes the monitoring data processed by the model more complete and reliable, providing strong support for subsequent data analysis and decision-making.

[0065] Step S14: Based on the reliable monitoring data, standardization processing is performed on the numerical data and the categorical data to obtain target comprehensive monitoring data.

[0066] Specifically, the numerical data of body surface temperature can be normalized according to the normal human body temperature range. For example, by subtracting the lowest value and dividing by the range difference, it can be converted to a value between 0 and 1. The numerical data of body surface humidity can be converted to a relative humidity percentage in the same way. The numerical data of pressure applied to the pressure area can be normalized according to medically recommended ranges to ensure that the data is on a reasonable scale.

[0067] Specifically, for the categorical data of gender, one-hot encoding can be used for standardization, such as "01" for female and "10" for male. For the categorical data of whether or not to have diabetes, considering that it is a binary categorical variable, binary encoding can be used for standardization, such as "0" for no and "1" for yes. For the categorical data of age, label encoding is used for standardization, such as defining the label as 0 for children aged 0-12, 1 for adolescents aged 13-18, 2 for adults aged 19-64, and 3 for seniors aged 65 and above, so as to divide them into different age groups (such as children, adolescents, adults, and the elderly).

[0068] In the above embodiment, on the one hand, through multi-dimensional verification, invalid data can be effectively identified and removed, such as microenvironmental data that exceeds a reasonable numerical range, data that is discontinuous in time, and inconsistent or logically unreasonable data in individual characteristic data. This ensures that the data relied on by subsequent analysis is accurate and reliable. On the other hand, using the missing value prediction and filling model, missing values ​​in effective monitoring data can be accurately predicted and filled. This reduces the lack of information in the data, improves the integrity and availability of the data, and provides more comprehensive data support for subsequent data analysis. Finally, numerical data and categorical data are standardized so that data from different sources and different scales can be converted into a unified format and standard. This helps to eliminate heterogeneity between data and improve the consistency and comparability of data.

[0069] In one embodiment, in step S13, the missing value prediction and filling model is composed of a mask-based adaptive convolutional neural network, a long short-term memory network, and a self-attention and cross-attention mechanism, wherein:

[0070] The mask-based adaptive convolutional neural network serves as a feature extractor, including a mask estimation module, multiple parallel adaptive convolutional layers, and a feature fusion and output module, wherein: the mask estimation module is used to generate a mask matrix corresponding to the input data based on context information, and the mask matrix marks the estimated positions of missing values ​​in the data; each of the adaptive convolutional layers is used to dynamically adjust the weight of the convolution kernel based on the mask matrix, combining the distribution characteristics of the estimated positions of missing values ​​and the characteristics of the convolution kernel itself, so that when the convolution kernel processes the estimated positions of missing values ​​marked by the mask matrix, it can reduce the weight of these positions, thereby avoiding the negative impact of missing values ​​on feature extraction; the feature fusion and output module is used to fuse the features extracted from each adaptive convolutional layer and generate a final feature representation.

[0071] Specifically, the mask generation module first analyzes contextual information, including but not limited to time series trends and the spatial distribution characteristics of the data. Based on this analysis, it generates a mask matrix corresponding to the input data. This mask matrix has the same shape as the input data, but each element indicates whether the data at the corresponding position is missing. Positions with missing values ​​are marked as 0, and positions with non-missing values ​​are marked as 1.

[0072] Specifically, each adaptive convolution layer first determines the distribution characteristics of the estimated missing value locations based on the mask matrix, such as the density of missing values. The density of missing values ​​can reflect the severity and distribution of missing data. The convolution kernel's weight adjustment strategy is then determined based on its own characteristics, such as its shape and size. During this process, if the density of missing values ​​within a certain position range is high and this position range is within the convolution kernel's coverage, the convolution kernel weight corresponding to this position range can be reduced to reduce the impact of missing values ​​on the convolution operation results. Finally, the convolution kernel with adjusted weights is used to traverse the input data and perform the convolution operation. During the convolution operation, the adjusted convolution kernel will be more adapted to the distribution characteristics of missing values ​​in the input data, thereby more accurately extracting the data's feature information and providing richer and more accurate input for subsequent network layers.

[0073] The long short-term memory network serves as a temporal dependency capturer and feature fusion processor, and is used to utilize its memory capacity and temporal analysis capabilities to process time series data, further capture the temporal dependencies in the data, and integrate the feature attributes associated with each time series monitoring value.

[0074] Specifically, when processing composite data consisting of time-series monitoring values ​​and the characteristic attributes associated with each monitoring value, the LSTM network not only focuses on the changing trends and periodicity of the time series data itself, but also fully considers the characteristic attributes associated with each monitoring value, such as environmental parameters and equipment status. This also means that when predicting missing values, the model comprehensively considers the changing trends and periodicity of the time series data, as well as the characteristic attributes associated with each monitoring value, such as environmental parameters and equipment status.

[0075] The self-attention and cross-attention mechanisms serve as information focusers and fusers. On the one hand, they allow the model to calculate the correlation between each point in time and other points based on the contextual information of the entire time series when processing the data of each time point, so that the model can more accurately identify which time points’ data are most critical for predicting missing values, thereby enhancing the robustness of the long short-term memory network in capturing temporal dependencies. On the other hand, they allow the model to simultaneously focus on the feature attribute information associated with each monitoring value when processing time series data, and help the long short-term memory network better integrate information between different data sources by calculating the correlation between elements of time series data and feature attribute data, thereby generating more comprehensive and accurate prediction results.

[0076] Specifically, when processing data at each time point, the self-attention mechanism model can more accurately identify which time points have strong correlations between data changes and missing values. By focusing on these key time points, the long short-term memory network's robustness in capturing temporal dependencies is further enhanced, enabling the model to more accurately understand the dynamic changes in time series data. When processing time series data, the cross-attention mechanism model can simultaneously focus on the characteristic attribute information associated with each monitoring value. By integrating information across data sources, the model can generate more comprehensive and accurate prediction results.

[0077] In the above embodiment, the introduction of the long short-term memory network enables the model to utilize its memory capacity and time series analysis capabilities to process time series data and effectively capture the temporal dependencies in the data. The introduction of the self-attention mechanism allows the model to calculate correlations based on the contextual information of the entire time series when processing the data at each time point, thereby more accurately identifying key time points and enhancing the robustness of the model in capturing temporal dependencies. The introduction of the cross-attention mechanism promotes information interaction between time series data and feature attribute data, helping the model better integrate information between different data sources and improving the comprehensiveness and accuracy of the prediction results.

[0078] In one embodiment, in step S2, the basic learner includes at least one of a decision tree, a support vector machine, a random forest, and a neural network.

[0079] In one embodiment, for each base learner, in step S2, the performance of each base learner is evaluated based on a method combining cross-validation and feature importance evaluation, including:

[0080] Step S21 : performing feature selection based on statistical analysis and correlation selection to obtain a preliminarily screened feature set.

[0081] Specifically, this application will calculate the correlation coefficient (such as the Pearson correlation coefficient) between each feature and the target variable, and sort them according to the absolute value of the correlation coefficient to determine the first feature with a high correlation with the target variable. Then, the correlation between the first feature and other features is calculated to avoid highly redundant features, and the second feature is obtained. After that, after combining statistical tests to verify the significance relationship between the second feature and the target variable, those features that are highly correlated with the target variable and not highly redundant with other features are further screened out to form a preliminarily screened feature set.

[0082] It should be noted that the features here refer to the variables used to predict the target variable. Specifically, these features include microenvironmental data indicating body surface temperature, surface humidity, and pressure at the pressure site, as well as individual characteristic data indicating the patient's gender, age, and whether they have diabetes. These features together constitute the input of the machine learning model, which is used to train the model to predict the target variable. The target variable is the variable predicted by the features, specifically the risk score for pressure ulcers.

[0083] Step S22: performing a feature importance evaluation on each feature in the feature set after the preliminary screening by using a Lasso regression algorithm to obtain a feature importance evaluation result.

[0084] Specifically, the Lasso regression algorithm introduces an L1 regularization term to compress feature coefficients, reducing some unimportant feature coefficients to zero while retaining important ones. After the calculation is complete, the absolute value of the remaining non-zero coefficients is used to assess feature importance. Coefficients with larger absolute values ​​indicate a greater contribution to the model and a greater predictive or explanatory effect on the target variable.

[0085] Step S23 , based on the feature importance evaluation result, the most important target feature subset is screened out from the feature set after the preliminary screening by sorting and thresholding method.

[0086] Specifically, based on the feature coefficients obtained by Lasso regression, the features in the initially screened feature set are ranked by absolute value. A threshold is then set, and features with importance scores above the threshold are selected to form the target feature subset. This ensures that the selected features are both important and relatively few, helping to reduce model complexity and the risk of overfitting.

[0087] Step S24: According to the size of the target feature subset, a K-fold cross-validation method is selected to train and validate the base learner, wherein, in each cross-validation cycle, K-1 pieces of training data are used for training, and the remaining 1 piece of data is used for validation.

[0088] Specifically, you first need to determine the value of K based on the size of the target feature subset (such as common choices are K = 5 or K = 10, which depends on the amount of data and computing resource limitations, as well as the stability of the desired model performance evaluation), and then randomly divide the input data set into K parts. It should be noted that in each cross-validation cycle, K-1 parts of the data will be randomly selected to construct the training set, and the remaining 1 part of the data will be used as the validation set. After that, the training set can be used to train the current basic learner. This step involves parameter optimization and model fitting of the learner. After the training is completed, the validation set is used to evaluate the performance of the model. The indicators for verifying model performance include at least one of accuracy, recall rate, and F1 score.

[0089] Step S25 , repeat the cross-validation process K times to obtain K independent performance evaluation results, and obtain a robust estimate of the performance of the base learner based on the mean and standard deviation.

[0090] Specifically, after each cross-validation cycle, the model's performance evaluation results are recorded. After completing K cycles, the mean and standard deviation of all performance evaluation results are calculated. The mean provides an overall estimate of model performance, while the standard deviation reflects the fluctuation range of the performance evaluation results. These statistics help us gain a more comprehensive understanding of the model's performance and stability.

[0091] In the above embodiment, on the one hand, the preliminary screening based on statistical analysis and correlation selection can remove redundant and noisy features, reduce the number of features, and improve the operating efficiency of the model. On the other hand, through the Lasso regression algorithm and feature importance evaluation, the most important features are further screened out to form a target feature subset. This helps to improve the interpretability and accuracy of the model, because fewer features usually mean less risk of overfitting. Finally, the base learner is trained and verified using the K-fold cross-validation method to ensure the stability and generalization ability of the model on different data sets. Through K independent training and verification processes, a more comprehensive performance evaluation can be obtained, including the mean and standard deviation, so as to more accurately estimate the true performance of the model.

[0092] In one embodiment, in step S2, each basic learner is assigned a corresponding weight according to the score of the performance evaluation index. After performing a weighted average combination based on the prediction results of all basic learners, the prediction result of the integrated model can be obtained, including:

[0093] Step S26, based on the performance evaluation index score of each base learner in the cross-validation process, the weight of each learner is calculated using a multi-objective optimization technique.

[0094] Specifically, this application will construct a multi-objective optimization model, which takes the performance evaluation index score of each basic learner as input and the weight of the learner as output. By solving the multi-objective optimization model, a set of optimal solutions are obtained, which represent the learner weight distribution schemes that balance different performance indicators. Afterwards, according to actual needs (such as giving priority to accuracy or recall rate, etc.), a suitable weight distribution scheme is selected from the optimal solution. Finally, the calculated weights are applied to each basic learner to construct the final integrated learning model. This model can combine the advantages of multiple learners to improve the accuracy and robustness of the prediction.

[0095] Step S27: For each sample to be predicted, use all basic learners to predict it and obtain respective prediction results.

[0096] Specifically, this application will obtain sample data to be predicted from actual application scenarios. Then, for each sample data to be predicted, all base learners will be used to predict it, and their respective prediction results will be obtained. Furthermore, this application will also organize the prediction results of each base learner into a unified form for subsequent weighted average calculation.

[0097] Step S28: performing weighted averaging based on the prediction results of each learner according to the determined weight of each learner to obtain an integrated prediction result.

[0098] Specifically, the present application obtains the weights of the determined base learners based on step S26. For each sample to be tested, a weighted average calculation is performed based on the weights and prediction results of the base learners. This is done by multiplying the prediction results of each learner by their corresponding weights, and then adding the weighted prediction results. Finally, the obtained weighted average is output as the prediction result of the ensemble model, and is interpreted or post-processed as needed.

[0099] In one embodiment, in step S27, after obtaining the prediction result, the method further includes:

[0100] Step S271, calculating the prediction accuracy and stability of each learner on the test sample.

[0101] Step S272: When it is determined that the accuracy and / or stability of the corresponding learner is greater than a preset threshold, the weight of the learner is increased according to the set weight adjustment coefficient; otherwise, the weight of the learner is reduced.

[0102] Step S273: When it is determined that the accuracy of the corresponding learner drops below a preset threshold and / or the stability fluctuation exceeds a preset range, the weight of the learner is recalculated and adjusted according to the evaluation result to ensure the prediction performance and stability of the integrated model.

[0103] It should be noted that based on steps S271 to S273, this application adjusts the weights of the basic learners in real time by monitoring the accuracy and stability of the prediction results. During specific implementation, this application will first calculate the prediction accuracy and stability of each learner on the test sample, and obtain the following results: the accuracy of learner A is 90%, and the stability is 90%; the accuracy of learner B is 85%, and the stability is 80%; the accuracy of learner C is 80%, and the stability is 75%. Based on this evaluation result, you can choose to increase the weight of learner A with higher accuracy and better stability, while reducing the weight of learner C with lower accuracy and poorer stability. The weight adjustment strategy can be determined based on a set weight adjustment coefficient, which is proportional to the relative performance of the learner. For example, a baseline coefficient (e.g., 1.0) can be pre-set. When the accuracy and / or stability of learner A is 10 percentage points higher than the average level (i.e., 90% - 80% = 10% and 90% - 80% = 10%), a proportional coefficient (e.g., 1.1 or higher, the specific value needs to be adjusted according to actual application and effect) can be set, and the weight adjustment coefficient of learner A is equal to the baseline coefficient multiplied by the square of the proportional coefficient to reflect its importance and contribution in the ensemble model. Conversely, for learner C, whose accuracy and stability are both below average, a decreasing proportional coefficient (e.g., 0.9 or lower) can be set, and the weight adjustment coefficient of learner C is equal to the baseline coefficient multiplied by the square of the decreasing proportional coefficient. Through this calculation method, it is possible to ensure that the weight adjustment coefficient is proportional to the relative performance of the learner, thereby achieving reasonable adjustment of the learner weights.

[0104] In one embodiment, in step S3, the risk of pressure injury at the patient's pressure site is quantitatively assessed based on the risk score and a risk report is generated, including:

[0105] Step S31: Classify the patient's risk score into different risk levels according to the preset risk score threshold range, and generate preventive measure recommendations corresponding to the risk level in combination with the intelligent preventive measure engine.

[0106] Specifically, the intelligent preventive measures engine has a built-in set of complete correspondences between risk levels and preventive measures. These preventive measures are formulated based on extensive clinical research, expert consensus and practical experience. When the engine receives the patient's risk score, it will automatically compare the preset threshold range to determine the patient's risk level. Subsequently, the engine will further analyze the patient's individual characteristics (such as age, weight, medical history, type of surgery, etc.) and select the most suitable preventive measures recommendations for the patient from the preventive measures library. These recommendations not only take into account the risk level, but also fully consider the individual differences of patients, thereby ensuring the pertinence and effectiveness of preventive measures.

[0107] Step S32: Obtain the patient's basic information, the risk score of the pressure site, the risk level, the preventive measures recommended for the risk level, and the potential risk factors, and generate a risk report through an interactive report template generation system.

[0108] Specifically, the interactive report template generation system will first automatically collect and integrate key data such as the patient's basic information, risk score, risk level, preventive measures recommendations, and potential risk factors. Then, the preliminary structure of the report will be automatically generated based on the preset report template framework. These template frameworks contain all the information points required for the report and allow users to fine-tune them according to actual needs. During the report generation process, the system will use visualization methods such as charts, lists, and highlights to intuitively display the patient's risk status and preventive measures recommendations, enhancing the readability and comprehensibility of the report. In addition, the system also supports users to edit and customize reports through an interactive interface, such as adjusting the report layout, adding annotations or notes, etc., to meet the personalized needs of different users. Ultimately, the system will generate a complete, professional and easy-to-understand risk report to provide comprehensive decision-making support for medical staff.

[0109] Please refer to Figure 2 The present application discloses an intelligent pressure injury prediction and prevention system, which includes a data acquisition module, a risk scoring module, and a risk quantification assessment module, wherein:

[0110] The data acquisition module is used to determine the patient's target comprehensive monitoring data, which includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on the compressed part, as well as individual characteristic data indicating the patient's gender, age, and whether he or she suffers from diabetes.

[0111] The risk scoring module is used to analyze the target comprehensive monitoring data using a machine learning model based on ensemble learning to obtain a risk score, wherein the machine learning model includes multiple basic learners, each basic learner is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight according to the score of the performance evaluation indicator. The prediction result of the ensemble model can be obtained by performing a weighted average combination based on the prediction results of all basic learners.

[0112] The risk quantification assessment module is used to quantitatively assess the risk of pressure injury at the patient's compressed area based on the risk score and generate a risk report.

[0113] In one embodiment, the above modules are also used to implement the steps illustrated in any of the above method embodiments, which is not limited in this application.

[0114] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction and prevention method for pressure injuries.

[0115] As can be seen from the above, the intelligent prediction and prevention system for pressure injuries disclosed in this application can achieve comprehensive monitoring of the risk of pressure injuries in patients' pressure areas by determining the patient's target comprehensive monitoring data, including microenvironmental data (surface temperature, surface humidity, pressure on the pressure area) and individual characteristic data (gender, age, diabetes status); the use of a machine learning model based on ensemble learning for analysis can fully utilize the advantages of multiple basic learners and combine the prediction results of each learner by weighted average, thereby improving the accuracy and stability of the overall prediction. Moreover, the performance of the basic learners is pre-evaluated based on a method combining cross-validation and feature importance evaluation, and weights are assigned according to the scores of the performance evaluation indicators, ensuring that the advantages and disadvantages of the learners are taken into account when the model is constructed, further improving the prediction performance; by quantitatively evaluating the risk of pressure injuries in patients' pressure areas, it can provide medical staff with clear risk indicators, helping them to take timely measures to intervene and reduce the risk of pressure injuries in patients.

[0116] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0117] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for intelligent prediction and prevention of pressure injuries, characterized in that: The method comprises: S1. Determining target comprehensive monitoring data for the patient, wherein the target comprehensive monitoring data includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on a pressure-receiving part, as well as individual characteristic data indicating the patient's gender, age, and whether or not the patient has diabetes; S2. Analyzing the target comprehensive monitoring data using an ensemble learning-based machine learning model to obtain a risk score, wherein the machine learning model includes multiple base learners, each of which is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight based on the score of the performance evaluation indicator. The prediction results of the ensemble model are obtained by performing a weighted average combination based on the prediction results of all the base learners; S3. quantitatively assessing the risk of pressure injury at the patient's compressed area based on the risk score, and generating a risk report; In step S1, determining the target comprehensive monitoring data of the patient includes: S11. Obtain the patient's initial comprehensive monitoring data; S12. Based on the initial comprehensive monitoring data, remove invalid data through multi-dimensional verification to obtain valid monitoring data, wherein the multi-dimensional verification includes a numerical range verification and a time continuity verification for the microenvironment data, and a data consistency verification and a logical rationality verification for the individual feature data; S13. Based on the valid monitoring data, performing missing value checking and filling processing through a missing value prediction and filling model to obtain reliable monitoring data; S14. Based on the reliable monitoring data, perform standardization processing on numerical data and categorical data to obtain target comprehensive monitoring data; In step S13, the missing value prediction and filling model is composed of a mask-based adaptive convolutional neural network, a long short-term memory network, and a self-attention and cross-attention mechanism, wherein: The mask-based adaptive convolutional neural network serves as a feature extractor and includes a mask estimation module, multiple parallel adaptive convolutional layers, and a feature fusion and output module, wherein: the mask estimation module is used to generate a mask matrix corresponding to the input data based on context information, and the mask matrix marks the estimated positions of missing values ​​in the data; each of the adaptive convolutional layers is used to dynamically adjust the weight of the convolution kernel based on the mask matrix, combining the distribution characteristics of the estimated positions of missing values ​​and the characteristics of the convolution kernel itself, so that the convolution kernel can reduce the weight of the estimated positions of missing values ​​marked by the mask matrix when processing these positions, thereby avoiding the negative impact of missing values ​​on feature extraction; the feature fusion and output module is used to fuse the features extracted from each adaptive convolutional layer and generate a final feature representation; The long short-term memory network acts as a temporal dependency capturer and feature fusion processor, and is used to process time series data using its memory capacity and temporal analysis capabilities, further capturing temporal dependencies in the data, and integrating feature attributes associated with each time series monitoring value; The self-attention and cross-attention mechanisms serve as information focusers and fusers. On the one hand, they allow the model to calculate the correlation between each point in time and other points based on the contextual information of the entire time series when processing the data of each time point, so that the model can more accurately identify which time points’ data are most critical for predicting missing values, thereby enhancing the robustness of the long short-term memory network in capturing temporal dependencies. On the other hand, they allow the model to simultaneously focus on the feature attribute information associated with each monitoring value when processing time series data, and help the long short-term memory network better integrate information between different data sources by calculating the correlation between elements of time series data and feature attribute data, thereby generating more comprehensive and accurate prediction results.

2. The method according to claim 1, characterized in that In step S2, the basic learner includes at least one of a decision tree, a support vector machine, a random forest, and a neural network.

3. The method according to claim 1, characterized in that For each base learner, in step S2, the performance of each base learner is evaluated based on a method combining cross-validation and feature importance evaluation, including: S21. Perform feature selection based on statistical analysis and correlation selection to obtain a preliminarily screened feature set; S22. Performing a feature importance evaluation on each feature in the feature set after the preliminary screening using a Lasso regression algorithm to obtain a feature importance evaluation result; S23. Based on the feature importance evaluation results, the most important target feature subset is screened from the feature set after the preliminary screening by sorting and thresholding methods; S24. According to the size of the target feature subset, a K-fold cross-validation method is selected to train and validate the base learner, wherein, in each cross-validation cycle, K-1 pieces of training data are used for training, and the remaining 1 piece of data is used for validation; S25. Repeat the cross-validation process K times to obtain K independent performance evaluation results, and obtain a robust estimate of the performance of the base learner based on the mean and standard deviation.

4. The method according to claim 3, characterized in that In step S2, each basic learner is assigned a corresponding weight according to the score of the performance evaluation index. After the prediction results of all basic learners are combined by weighted average, the prediction results of the integrated model can be obtained, including: S26. Based on the performance evaluation index score of each base learner in the cross-validation process, the weight of each learner is calculated using a multi-objective optimization technique; S27. For each sample to be predicted, use all basic learners to predict it and obtain their respective prediction results; S28. According to the determined weight of each learner, a weighted average is performed based on the prediction results of each learner to obtain an integrated prediction result.

5. The method according to claim 4, characterized in that In step S27, after obtaining the prediction result, the method further includes: S271. Calculate the prediction accuracy and stability of each learner on the test sample; S272: When it is determined that the accuracy and / or stability of the corresponding learner is greater than a preset threshold, the weight of the learner is increased according to the set weight adjustment coefficient; otherwise, the weight of the learner is decreased; S273. When it is determined that the accuracy of the corresponding learner drops below a preset threshold and / or the stability fluctuation exceeds a preset range, the weight of the learner is recalculated and adjusted according to the evaluation result to ensure the prediction performance and stability of the integrated model.

6. The method according to any one of claims 1 to 5, characterized in that In step S3, the risk of pressure injury at the patient's compressed part is quantitatively assessed based on the risk score, and a risk report is generated, including: S31. Classify the patient's risk score into different risk levels based on a preset risk score threshold range, and generate preventive measures recommendations corresponding to the risk level in conjunction with the intelligent preventive measures engine; S32. Obtain the patient's basic information, the risk score of the pressure site, the risk level, the recommended preventive measures for the risk level, and the potential risk factors, and generate a risk report through an interactive report template generation system.

7. An intelligent pressure injury prediction and prevention system, characterized in that: The system includes a data acquisition module, a risk scoring module, and a risk quantification assessment module, wherein: The data acquisition module is used to determine the target comprehensive monitoring data of the patient, wherein the target comprehensive monitoring data includes microenvironmental data indicating body surface temperature, body surface humidity, and pressure on the pressure-receiving part, and individual characteristic data indicating the patient's gender, age, and whether or not the patient has diabetes; The risk scoring module is configured to analyze the target comprehensive monitoring data using a machine learning model based on ensemble learning to obtain a risk score, wherein the machine learning model includes multiple basic learners, each of which is pre-performance evaluated based on a method combining cross-validation and feature importance evaluation, and is assigned a corresponding weight based on the score of the performance evaluation indicator. The prediction results of the ensemble model can be obtained by performing a weighted average combination based on the prediction results of all the basic learners; The risk quantification assessment module is used to quantitatively assess the risk of pressure injury at the patient's compressed part based on the risk score and generate a risk report; The specific implementation of the data acquisition module determining the target comprehensive monitoring data of the patient is as follows: Obtain initial comprehensive monitoring data from patients; Based on the initial comprehensive monitoring data, invalid data is removed through multi-dimensional verification to obtain valid monitoring data, wherein the multi-dimensional verification includes a numerical range verification and a time continuity verification for the microenvironment data, a data consistency verification and a logical rationality verification for the individual feature data; Based on the effective monitoring data, missing value checking and filling processing are performed through a missing value prediction and filling model to obtain reliable monitoring data; Based on the reliable monitoring data, standardization processing of numerical data and categorical data is performed to obtain target comprehensive monitoring data; The missing value prediction and filling model is composed of a mask-based adaptive convolutional neural network, a long short-term memory network, and a self-attention and cross-attention mechanism, where: The mask-based adaptive convolutional neural network serves as a feature extractor and includes a mask estimation module, multiple parallel adaptive convolutional layers, and a feature fusion and output module, wherein: the mask estimation module is used to generate a mask matrix corresponding to the input data based on context information, and the mask matrix marks the estimated positions of missing values ​​in the data; each of the adaptive convolutional layers is used to dynamically adjust the weight of the convolution kernel based on the mask matrix, combining the distribution characteristics of the estimated positions of missing values ​​and the characteristics of the convolution kernel itself, so that the convolution kernel can reduce the weight of the estimated positions of missing values ​​marked by the mask matrix when processing these positions, thereby avoiding the negative impact of missing values ​​on feature extraction; the feature fusion and output module is used to fuse the features extracted from each adaptive convolutional layer and generate a final feature representation; The long short-term memory network acts as a temporal dependency capturer and feature fusion processor, and is used to process time series data using its memory capacity and temporal analysis capabilities, further capturing temporal dependencies in the data, and integrating feature attributes associated with each time series monitoring value; The self-attention and cross-attention mechanisms serve as information focusers and fusers. On the one hand, they allow the model to calculate the correlation between each point in time and other points based on the contextual information of the entire time series when processing the data of each time point, so that the model can more accurately identify which time points’ data are most critical for predicting missing values, thereby enhancing the robustness of the long short-term memory network in capturing temporal dependencies. On the other hand, they allow the model to simultaneously focus on the feature attribute information associated with each monitoring value when processing time series data, and help the long short-term memory network better integrate information between different data sources by calculating the correlation between elements of time series data and feature attribute data, thereby generating more comprehensive and accurate prediction results.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for intelligent prediction and prevention of pressure injuries according to any one of claims 1 to 6 is implemented.

Images