An abdominal peritoneal data monitoring method, system and storage medium

Abstract

The application provides a peritoneal dialysis data monitoring method, system and storage medium, which comprises the following steps: acquiring and normalizing time sequence physiological index data of a peritoneal dialysis patient; constructing a time index decay factor weighted covariance matrix, calculating the variance contribution rate and time stability coefficient of principal components; obtaining a comprehensive score by weighted summation of the two; selecting the first N principal components to form a feature space, projecting data to obtain reduced dimension time sequence features; initializing an incremental ELM network, training with reduced dimension data, setting the input weight of a new node by weighting the comprehensive score of the first N principal components, calculating the node activity contribution degree and removing nodes below the dynamic threshold; monitoring using the trained network, increasing the time index decay factor when the sliding window variance of the prediction error exceeds the preset multiple of the historical baseline variance, triggering covariance matrix reconstruction and network node adjustment.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to a method, system and storage medium for peritoneal dialysis data monitoring. Background Technology

[0002] Peritoneal dialysis is one of the renal replacement therapies for patients with end-stage renal disease. Continuous monitoring of various physiological indicators can prevent complications such as peritonitis and overload. Patients undergoing peritoneal dialysis generate a large amount of multidimensional time-series physiological data, such as weight, blood pressure, ultrafiltration volume, and sodium concentration in the dialysate. Principal component analysis (PCA) is a commonly used dimensionality reduction method to extract features from this high-dimensional data. While PCA selects principal components based on variance contribution rates, maximizing the preservation of variance information in the original data, it may overlook the stability of features over time. In dynamically changing physiological systems, some high-variance components caused by noise or transient perturbations may be incorrectly selected as important features, while some key physiological patterns with low variance but stable persistence may be ignored, affecting the accuracy and reliability of subsequent monitoring models. To adapt to the streaming characteristics of time-series data, Incremental Extreme Learning Machine (I-ELM) is proposed, which can add hidden layer nodes based on new data, enabling online learning of the network. The input weights for newly added nodes are typically randomly generated, lacking prior knowledge guidance, leading to low learning efficiency and the potential for invalid input nodes. As data accumulates, the network structure continuously expands, increasing computational complexity and potentially causing overfitting. Monitoring methods are mostly static or incremental learning models, which struggle to reliably handle dynamic changes in patients' physiological states. When changes in patient state cause a decline in model predictive performance, there is a lack of a closed-loop adjustment mechanism to update feature extraction and model parameters, allowing the model to readjust to the patient's latest condition. Summary of the Invention

[0003] This invention proposes a method for monitoring peritoneal dialysis data, comprising the following steps:

[0004] Acquire and normalize time-series physiological data of peritoneal dialysis patients; A covariance matrix weighted by a time exponential decay factor is constructed, the variance contribution rate of each principal component is calculated, and the time stability coefficient is determined based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window. The variance contribution rate and the time stability coefficient are weighted and summed to obtain a comprehensive score. The top N principal components with the highest comprehensive scores are selected to form a feature space, and the normalized data is projected onto the feature space to obtain dimensionality-reduced time series feature data. Initialize the incremental ELM network and train it using the dimensionality-reduced temporal feature data. During training, the input weights of newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score. The activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value. Nodes whose activity contribution is consistently lower than the pruning threshold are removed. The trained network is used for monitoring. When the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the network hidden layer nodes are triggered.

[0005] Optionally, the step of constructing a covariance matrix weighted by a time-exponential decay factor, calculating the variance contribution rate of each principal component, and determining the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window includes: At each time step t, through the recursive formula Update covariance matrix ; Given a continuous time window size L of 20 time steps, for the i-th principal component, the time stability coefficient is... By calculating the feature vector of the principal component at the current time step The feature vectors of the past L time steps The mean cosine similarity is obtained using the formula: ;in, It is the time-exponential decay factor.

[0006] Optionally, the step of weighted summing of the variance contribution rate and the time stability coefficient to obtain a comprehensive score, and selecting the top N principal components of the comprehensive score to form a feature space, includes: For the i-th principal component, the variance contribution rate is... The time stability coefficient is Overall score Through formula Calculated; Sort all principal components from highest to lowest based on their overall scores, and select the top N=20 principal components to form the feature space.

[0007] Optionally, the acquisition and normalization of time-series physiological index data of peritoneal dialysis patients includes: Obtain the patient's systolic blood pressure, diastolic blood pressure, weight, ultrafiltration volume, and peritoneal dialysis fluid glucose concentration data; For each physiological indicator data point acquired, the Z-score standardization method was used, and the result was calculated using the formula... Normalization is performed, where x is the original data point. The historical average of the indicator. This represents the historical standard deviation.

[0008] Optionally, the input weights of the newly added hidden layer nodes are set by a weighted linear combination of the first N principal components based on their comprehensive scores, including: The combined score of the top N principal components

[0009] Normalization is performed to obtain the weight coefficients. The normalization formula is ; The input weight vector of the newly added hidden layer node Through formula The calculation yielded, where Let be the eigenvector of the i-th principal component; The bias of the newly added node Randomly selected from a uniform distribution in the range [-0.5, 0.5].

[0010] Optionally, the step of calculating the activity contribution based on the product of the absolute value of the node's output weight and the recent average activation value, and removing nodes whose activity contribution is consistently below the pruning threshold, includes: Set the recent time window to the past 100 samples, and calculate the average activation value of the j-th hidden layer node within the window. ; The activity contribution of the node Through formula The calculation yielded, where The output weights of the nodes; Pruning threshold Set as the contribution of all node activity 15% of the average; If the activity contribution of a certain node If a node is below the pruning threshold for 300 consecutive sample points, it is removed from the network.

[0011] Optionally, the monitoring using the trained network, when the sliding window variance of the error between the predicted and actual values ​​exceeds a preset multiple of the historical baseline variance, increases the time exponential decay factor, including: Set the sliding window size to 150 data points and calculate the variance of the monitoring error within the current window. ; Set historical baseline variance The error variance of the first 3000 data points during stable system operation; When the condition is met At that time, the time exponential decay factor will be used. The value is increased by 0.005, and the upper limit is set to 0.99.

[0012] Furthermore, the present invention also relates to a peritoneal dialysis data monitoring system, comprising the following modules: The acquisition module is used to acquire and normalize the time-series physiological index data of peritoneal dialysis patients. The module is used to construct a covariance matrix weighted by a time exponential decay factor, calculate the variance contribution rate of each principal component, and determine the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window; the variance contribution rate and the time stability coefficient are weighted and summed to obtain a comprehensive score; the top N principal components with the highest comprehensive scores are selected to form a feature space; and the normalized data is projected onto the feature space to obtain dimensionality-reduced time series feature data. The calculation module is used to initialize the incremental ELM network and train it using the dimensionality-reduced temporal feature data. During the training process, the input weights of the newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score. The activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value, and nodes whose activity contribution is consistently lower than the pruning threshold are removed. The monitoring module is used to monitor the network after training. When the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the hidden layer nodes of the network are triggered.

[0013] Preferably, the step of constructing a covariance matrix weighted by a time-exponential decay factor, calculating the variance contribution rate of each principal component, and determining the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window includes: At each time step t, through the recursive formula Update covariance matrix ; Given a continuous time window size L of 20 time steps, for the i-th principal component, the time stability coefficient is... By calculating the feature vector of the principal component at the current time step The feature vectors of the past L time steps The mean cosine similarity is obtained using the formula: ;in, It is the time-exponential decay factor.

[0014] Preferably, the step of weighted summing of the variance contribution rate and the time stability coefficient to obtain a comprehensive score, and selecting the top N principal components of the comprehensive score to form a feature space, includes: For the i-th principal component, the variance contribution rate is... The time stability coefficient is Overall score Through formula Calculated; Sort all principal components from highest to lowest based on their overall scores, and select the top N=20 principal components to form the feature space.

[0015] Preferably, the acquisition and normalization of time-series physiological index data of peritoneal dialysis patients includes: Obtain the patient's systolic blood pressure, diastolic blood pressure, weight, ultrafiltration volume, and peritoneal dialysis fluid glucose concentration data; For each physiological indicator data point acquired, the Z-score standardization method was used, and the result was calculated using the formula... Normalization is performed, where x is the original data point. The historical average of the indicator. This represents the historical standard deviation.

[0016] Preferably, the input weights of the newly added hidden layer nodes are set by a weighted linear combination of the first N principal components based on their comprehensive scores, including: The combined score of the top N principal components

[0017] Normalization is performed to obtain the weight coefficients. The normalization formula is ; The input weight vector of the newly added hidden layer node Through formula The calculation yielded, where Let be the eigenvector of the i-th principal component; The bias of the newly added node Randomly selected from a uniform distribution in the range [-0.5, 0.5].

[0018] Preferably, the step of calculating the activity contribution based on the product of the absolute value of the node's output weight and the recent average activation value, and removing nodes whose activity contribution is consistently below the pruning threshold, includes: Set the recent time window to the past 100 samples, and calculate the average activation value of the j-th hidden layer node within the window. ; The activity contribution of the node Through formula The calculation yielded, where The output weights of the nodes; Pruning threshold Set as the contribution of all node activity 15% of the average; If the activity contribution of a certain node If a node is below the pruning threshold for 300 consecutive sample points, it is removed from the network.

[0019] Preferably, the monitoring using the trained network involves increasing the time exponential decay factor when the sliding window variance of the error between the predicted and actual values ​​exceeds a preset multiple of the historical baseline variance, including: Set the sliding window size to 150 data points and calculate the variance of the monitoring error within the current window. ; Set historical baseline variance The error variance of the first 3000 data points during stable system operation; When the condition is met At that time, the time exponential decay factor will be used. The value is increased by 0.005, and the upper limit is set to 0.99.

[0020] This invention, by inputting time stability assessment into principal component analysis and combining stability with variance contribution rate as the basis for feature selection, can identify features reflecting the patient's true physiological state, avoiding interference from noise or transient perturbations in feature extraction, and improving the quality and reliability of the feature space. In the construction of the incremental learning network, the selected features are used to guide the weight setting of newly added hidden layer nodes, accelerating the model's convergence speed and improving training efficiency. A contribution evaluation mechanism can eliminate nodes with weak effects in the network, maintaining the simplicity of the network structure and preventing overfitting. When the monitoring error increases, it can trigger the reconstruction of the feature space and network structure, enabling the model to keep up with changes in the patient's physiological state, thereby ensuring accuracy and effectiveness in long-term monitoring. Attached Figure Description

[0021] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram of Z-score standardization for physiological indicator data; Figure 3 This is a schematic diagram of principal component selection based on the comprehensive score. Detailed Implementation

[0022] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0023] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0024] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0025] In the first embodiment, the present invention proposes a method for monitoring peritoneal dialysis data, such as... Figure 1 This includes the following steps: S1, Obtain and normalize the time-series physiological data of peritoneal dialysis patients; Multiple physiological indicators, including patient weight, systolic blood pressure, diastolic blood pressure, ultrafiltration volume, and sodium concentration in the transdermal solution, were continuously collected to form a raw data matrix with time as the row and each indicator as the column. For each column of indicator data in the matrix, a min-max normalization method was applied. That is, for any data point, the normalized value is the value of that data point minus the minimum value of the indicator sequence, and then divided by the difference between the maximum and minimum values ​​of the indicator sequence, so that all indicator data are mapped to the interval between 0 and 1.

[0026] In an optional embodiment, the acquisition and normalization of time-series physiological data of peritoneal dialysis patients includes: Obtain the patient's systolic blood pressure, diastolic blood pressure, weight, ultrafiltration volume, and peritoneal dialysis fluid glucose concentration data; For each physiological indicator data point acquired, the Z-score standardization method was used, and the result was calculated using the formula... Normalization is performed, where x is the original data point. The historical average of the indicator. This represents the historical standard deviation.

[0027] Specifically, the first step is data collection, which involves continuously monitoring and recording the patient's various physiological indicators. For example, at a certain moment, a set of raw data is collected: systolic blood pressure is 140 mmHg, and weight is 72 kg. To perform multivariate analysis, it is necessary to eliminate the influence of units and dimensions between different indicators.

[0028] The second step is to calculate historical statistics, which requires using the patient's long-term historical monitoring data to calculate the mean of each indicator. and standard deviation Assuming that, based on historical data, the mean systolic blood pressure is 125 mmHg with a standard deviation of 10 mmHg; and the mean weight is 70 kg with a standard deviation of 1.5 kg, the statistics represent the baseline level and normal fluctuation range of the patient's physiological state.

[0029] The third step is to perform normalization, substituting the currently collected raw data points into the Z-score standardization formula, such as... Figure 2 For systolic blood pressure, the normalized value is 1.5. For body weight, the normalized value is approximately 1.33. This process is applied to all collected physiological indicators, transforming the raw data vector into a standardized data vector, such as 1.5,...,1.33,..., where each element of the vector represents the standard deviation of the corresponding indicator from its historical mean, making different indicators comparable.

[0030] S2, construct a covariance matrix weighted by a time exponential decay factor, calculate the variance contribution rate of each principal component, and determine the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window; weight the variance contribution rate and the time stability coefficient to obtain a comprehensive score, select the top N principal components with the comprehensive scores to form a feature space, and project the normalized data into the feature space to obtain dimensionality-reduced time series feature data; Each data vector at a given time point is assigned a time decay weight, with data closer to the current time point receiving a greater weight. The weight is calculated as e^(-λ) multiplied by the power of the time difference. A weighted mean vector is then calculated based on this weight, leading to a weighted covariance matrix. This covariance matrix is ​​then subjected to eigenvalue decomposition to obtain a series of eigenvalues ​​and their corresponding eigenvectors, i.e., principal components. In one embodiment, the variance contribution rate of each principal component is the sum of its corresponding eigenvalue and all eigenvalues. The entire time series data is divided into multiple consecutive and non-overlapping time windows. Principal components are calculated independently within each window. For principal components with the same index, the cosine of the angle between the eigenvectors of the principal component in adjacent windows is calculated, and the cosine values ​​of all adjacent windows are averaged to obtain the time stability coefficient of the principal component.

[0031] Assign a weighting factor, such as 0.6, to the variance contribution rate and the time stability coefficient, respectively. Multiply the variance contribution rate of each principal component by 0.6, and add the time stability coefficient multiplied by 0.4 to obtain the overall score of that principal component. Figure 3 Sort all principal components by their comprehensive scores from highest to lowest, and select the top N principal components whose cumulative comprehensive scores reach a preset threshold, such as 95%. Use the eigenvectors corresponding to these N principal components as column vectors to form a projection matrix. Multiply the normalized original data matrix by this projection matrix to obtain a dimensionality-reduced time-series feature data matrix with N rows representing time points and N columns.

[0032] In an optional embodiment, the step of constructing a covariance matrix weighted by a time-exponential decay factor, calculating the variance contribution rate of each principal component, and determining the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window includes: At each time step t, through the recursive formula Update covariance matrix ; Given a continuous time window size L of 20 time steps, for the i-th principal component, the time stability coefficient is... By calculating the feature vector of the principal component at the current time step The feature vectors of the past L time steps The mean cosine similarity is obtained using the formula: ;in, It is the time-exponential decay factor.

[0033] The first step is to set the time exponential decay factor when the system starts up. The initial value is 0.98, and the covariance matrix is ​​initialized. It is usually a zero matrix. The data vector at the first time step t=1 For example, when normalized physiological index data arrives, a new covariance matrix is ​​calculated using a recursive formula. The value is The update process continues at each subsequent time step t, ensuring that recent data contributes more to the covariance matrix, thus enabling principal component analysis to adapt to the latest changes in the data.

[0034] The second step is to calculate the time stability coefficient. At any time step t, the updated covariance matrix is... Perform eigenvalue decomposition to obtain a set of eigenvectors. To evaluate the time stability of the first principal component. It is necessary to obtain the feature vectors of the principal components at a total of 21 time steps, including the current and past ones. to .calculate With each feature vector of the past 20 time steps The cosine similarity between the components is used to generate 20 similarity values, such as 0.99, 0.97, etc. The arithmetic mean of these 20 similarity values ​​is then calculated, and the result is the time stability coefficient of the first principal component at time step t. This process is repeated for all principal components.

[0035] In an optional embodiment, the step of weighted summing of the variance contribution rate and the time stability coefficient to obtain a comprehensive score, and selecting the top N principal components of the comprehensive score to form a feature space, includes: For the i-th principal component, the variance contribution rate is... The time stability coefficient is Overall score Through formula Calculated; Sort all principal components from highest to lowest based on their overall scores, and select the top N=20 principal components to form the feature space.

[0036] Specifically, the first step is to calculate the composite score of each principal component. At a certain point in time, suppose three principal components were obtained through principal component analysis. The variance contribution rate of the first principal component... A value of 0.5 indicates that it explains 50% of the data variation, representing a time stability coefficient. A value of 0.9 indicates that the direction changes very little over time. The overall score is 0.62. The variance contribution rate of the second principal component... The time stability coefficient is 0.3. The score was 0.6, and the overall score was 0.39.

[0037] The second step is sorting and selection. All principal components are sorted in descending order based on their calculated composite scores. In the example above, the score of the first principal component (0.62) is higher than that of the second principal component (0.39). Assuming there are a total of 50 principal components, this process will be performed on all 50 principal components, generating a sorted list from highest to lowest score. Based on a preset number N=20, the top 20 principal components are selected from the top of this sorted list. The feature vectors corresponding to these 20 principal components collectively constitute a dimensionality-reduced feature space that represents both the direction of data variation and time stability, used for subsequent modeling.

[0038] S3, Initialize the incremental ELM network and train it using the dimensionality-reduced temporal feature data; During training, the input weights of newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score, the activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value, and nodes whose activity contribution is consistently lower than the pruning threshold are removed. An Extreme Learning Machine (ELM) network with no or only a few hidden layer nodes is constructed. When a new batch of dimensionality-reduced temporal feature data is input, if the network error exceeds the tolerance range, a new hidden layer node is added. The input weight vector of this new node is obtained by weighted summation of the feature vectors of the first N principal components based on their normalized comprehensive scores. After the network training updates the output weights, for each hidden layer node in the network, the average activation value on the most recent batch of data is recorded, and this average activation value is multiplied by the absolute value of the corresponding output weight to obtain the node's activity contribution. A dynamic threshold is set, for example, 10% of the average contribution of all nodes. If the activity contribution of a node is lower than this threshold for several consecutive iterations, the node and its connection weights are removed from the network.

[0039] In an optional embodiment, the input weights of the newly added hidden layer nodes are set by a weighted linear combination of the first N principal components based on their comprehensive scores, including: The combined score of the top N principal components

[0040] Normalization is performed to obtain the weight coefficients. The normalization formula is ; The input weight vector of the newly added hidden layer node Through formula The calculation yielded, where Let be the eigenvector of the i-th principal component; The bias of the newly added node Randomly selected from a uniform distribution in the range [-0.5, 0.5].

[0041] The first step is to calculate the weighting coefficients. Assuming that N equals 3 principal components have been selected, and the overall scores are as follows: , , Calculate the total score, which is 1.8. Then, calculate the weight coefficient of each principal component using the normalization formula. , , The weights reflect the relative importance of each principal component in constructing new knowledge.

[0042] The second step is to set the input weight vector. This involves obtaining the feature vectors corresponding to the three principal components. The input weight vector of the newly added hidden layer node. It is obtained by multiplying the feature vector by its corresponding weight coefficient and summing the results, i.e. The constructed weight vector incorporates the most important and stable data change patterns. A bias value is generated for the new node. For example, a value such as 0.15 can be drawn from the range of -0.5 to 0.5 using a random number generator. The new node and the initialized weights and biases are then added to the hidden layer of the network.

[0043] In an optional embodiment, the step of calculating the activity contribution based on the product of the absolute value of the node's output weight and the recent average activation value, and removing nodes whose activity contribution is consistently below the pruning threshold, includes: Set the recent time window to the past 100 samples, and calculate the average activation value of the j-th hidden layer node within the window. ; The activity contribution of the node Through formula The calculation yielded, where The output weights of the nodes; Pruning threshold Set as the contribution of all node activity 15% of the average; If the activity contribution of a certain node If a node is below the pruning threshold for 300 consecutive sample points, it is removed from the network.

[0044] The first step is to calculate the activity contribution of each hidden layer node. Taking the j-th hidden node in the network as an example, record its activation value in the most recent 100 input samples and calculate the average value. Assuming we get Obtain the output weights of this node. Assuming a value of -0.2, this represents the node's activity contribution. The value is 0.14. This calculation is performed on each hidden layer node in the network one by one.

[0045] The second step is to determine the pruning threshold. Assuming the hidden layer of the network has 50 nodes, after calculating the activity contribution of all nodes, the arithmetic mean of these 50 activity contributions is calculated. If the average is 0.25, then the pruning threshold is... The threshold is set to 0.0375. This threshold will be adjusted as the network status changes.

[0046] The third step is to perform pruning judgment. For each node, the activity contribution is continuously tracked. Taking node j as an example, its activity contribution of 0.14 is higher than the threshold of 0.0375, so it is considered active. If there is another node m, its activity contribution... If the value is 0.02, which is below the threshold, a counter is started. If the activity contribution of node m remains below the threshold for the next 300 consecutive sample points, the node is determined to be a redundant or invalid node and is permanently removed from the network structure, including all input and output connections.

[0047] S4. Using the trained network for monitoring, when the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the network hidden layer nodes are triggered.

[0048] Specifically, the new, dimensionality-reduced data is input into the trained network, which outputs a predicted value for a key indicator, such as ultrafiltration rate. The error between the predicted value and the actual measurement value is calculated. A sliding window containing, for example, the 100 most recent error values ​​is maintained, and the variance of the error within this window is calculated. A baseline value recording the error variance over a longer historical period is maintained. If the error variance of the current sliding window is greater than three times the historical baseline variance, the model performance is considered to have degraded. The value of the time exponential decay factor λ used to calculate the weighted covariance matrix is ​​increased, for example, by 20%. The process returns to step two, and the weighted covariance matrix is ​​recalculated using the increased decay factor. Subsequent feature selection, spatial projection, and network structure adjustment steps are then executed sequentially to complete the closed-loop update of the entire monitoring model.

[0049] In an optional embodiment, the monitoring using the trained network, when the sliding window variance of the error between the predicted and actual values ​​exceeds a preset multiple of the historical baseline variance, increases the time exponential decay factor, including: Set the sliding window size to 150 data points and calculate the variance of the monitoring error within the current window. ; Set historical baseline variance The error variance of the first 3000 data points during stable system operation; When the condition is met At that time, the time exponential decay factor will be used. The value is increased by 0.005, and the upper limit is set to 0.99.

[0050] The first step is to establish a historical baseline. In the initial deployment phase, the network operates under relatively stable conditions. The prediction error of the first 3000 data points is collected, representing the difference between the model's predicted values ​​and the actual physiological indicators. The variance of these 3000 error values ​​is calculated to obtain the historical baseline variance. For example, a value of 0.01. This value represents the level of fluctuation in prediction error under normal conditions.

[0051] The second step is to monitor the error variance in real time. During continuous operation, a sliding window containing the 150 most recent prediction errors is maintained. Whenever new data points arrive, the window is updated, and the variance of the 150 errors is recalculated to obtain the current window variance. Suppose that at some point, due to a sudden change in the patient's physiological state, the prediction error increases, and the calculated... It is 0.03.

[0052] The third step is to make judgments and adjustments, comparing the current variance with the baseline variance, and determining the judgment criteria. Is this true? In the example above, the result is true because 0.03 > 0.025. At this point, the data pattern may have changed, requiring a reduction in sensitivity to new data. Therefore, the current time exponential decay factor is adjusted. Assuming the original value was 0.98, and it is increased by 0.005, the updated value is 0.985. If it continues to trigger, It will continue to increase, but the value will not exceed the set upper limit of 0.99.

[0053] In a second embodiment, the present invention also provides a peritoneal dialysis data monitoring system, comprising the following modules: The acquisition module is used to acquire and normalize the time-series physiological index data of peritoneal dialysis patients. The module is used to construct a covariance matrix weighted by a time exponential decay factor, calculate the variance contribution rate of each principal component, and determine the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window; the variance contribution rate and the time stability coefficient are weighted and summed to obtain a comprehensive score; the top N principal components with the highest comprehensive scores are selected to form a feature space; and the normalized data is projected onto the feature space to obtain dimensionality-reduced time series feature data. The calculation module is used to initialize the incremental ELM network and train it using the dimensionality-reduced temporal feature data. During the training process, the input weights of the newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score. The activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value, and nodes whose activity contribution is consistently lower than the pruning threshold are removed. The monitoring module is used to monitor the network after training. When the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the hidden layer nodes of the network are triggered.

[0054] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, compact disc read-only memory (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0055] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for monitoring peritoneal dialysis data, characterized in that, Includes the following steps: Acquire and normalize time-series physiological data of peritoneal dialysis patients; Construct a covariance matrix weighted by a time exponential decay factor, calculate the variance contribution rate of each principal component, and determine the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window. The variance contribution rate and the time stability coefficient are weighted and summed to obtain the comprehensive score. The top N principal components with the comprehensive scores are selected to form the feature space. The normalized data is then projected onto the feature space to obtain the dimensionality-reduced time series feature data. Initialize the incremental ELM network and train it using the reduced-dimensionality temporal feature data; During training, the input weights of newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score. The activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value, and nodes whose activity contribution is consistently lower than the pruning threshold are removed. The trained network is used for monitoring. When the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the network hidden layer nodes are triggered.

2. The method according to claim 1, characterized in that, The construction of a covariance matrix weighted by a time-exponential decay factor, calculation of the variance contribution rate of each principal component, and determination of the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window include: At each time step t, through the recursive formula Update covariance matrix ; Given a continuous time window size L of 20 time steps, for the i-th principal component, the time stability coefficient is... By calculating the feature vector of the principal component at the current time step The feature vectors of the past L time steps The mean cosine similarity is obtained using the formula: ;in, It is the time-exponential decay factor.

3. The method according to claim 2, characterized in that, The process involves weighting and summing the variance contribution rate and the time stability coefficient to obtain a comprehensive score, and then selecting the top N principal components based on the comprehensive score to construct the feature space, including: For the i-th principal component, the variance contribution rate is... The time stability coefficient is Overall score Through formula Calculated; Sort all principal components from highest to lowest based on their overall scores, and select the top N=20 principal components to form the feature space.

4. The method according to claim 1, characterized in that, The acquisition and normalization of time-series physiological data of peritoneal dialysis patients includes: Obtain the patient's systolic blood pressure, diastolic blood pressure, weight, ultrafiltration volume, and peritoneal dialysis fluid glucose concentration data; For each physiological indicator data point acquired, the Z-score standardization method was used, and the result was calculated using the formula... Normalization is performed, where x is the original data point. The historical average of the indicator. This represents the historical standard deviation.

5. The method according to claim 1, characterized in that, The input weights of the newly added hidden layer nodes are set by a weighted linear combination of the first N principal components based on their comprehensive scores, including: The combined score of the top N principal components Normalization is performed to obtain the weight coefficients. The normalization formula is ; The input weight vector of the newly added hidden layer node Through formula The calculation yielded, where Let be the eigenvector of the i-th principal component; The bias of the newly added node Randomly selected from a uniform distribution in the range [-0.5, 0.5].

6. The method according to claim 1, characterized in that, The step of calculating the activity contribution based on the product of the absolute value of the node's output weight and the recent average activation value, and removing nodes whose activity contribution is consistently below the pruning threshold, includes: Set the recent time window to the past 100 samples, and calculate the average activation value of the j-th hidden layer node within the window. ; The activity contribution of the node Through formula The calculation yielded, where The output weights of the nodes; Pruning threshold Set as the contribution of all node activity 15% of the average; If the activity contribution of a certain node If a node is below the pruning threshold for 300 consecutive sample points, it is removed from the network.

7. The method according to claim 1, characterized in that, The monitoring using the trained network involves increasing the time exponential decay factor when the sliding window variance of the error between the predicted and actual values ​​exceeds a preset multiple of the historical baseline variance. This includes: Set the sliding window size to 150 data points and calculate the variance of the monitoring error within the current window. ; Set historical baseline variance The error variance of the first 3000 data points during stable system operation; When the condition is met At that time, the time exponential decay factor will be used. The value is increased by 0.005, and the upper limit is set to 0.

99.

8. A peritoneal dialysis data monitoring system, characterized in that, Includes the following modules: The acquisition module is used to acquire and normalize the time-series physiological index data of peritoneal dialysis patients. The module is used to construct a covariance matrix weighted by a time exponential decay factor, calculate the variance contribution rate of each principal component, and determine the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window. The variance contribution rate and the time stability coefficient are weighted and summed to obtain the comprehensive score. The top N principal components with the comprehensive scores are selected to form the feature space. The normalized data is then projected onto the feature space to obtain the dimensionality-reduced time series feature data. The computation module is used to initialize the incremental ELM network and train it using the reduced-dimensional temporal feature data; During training, the input weights of newly added hidden layer nodes are set by the weighted linear combination of the first N principal components based on the comprehensive score. The activity contribution is calculated based on the product of the absolute value of the node's output weight and the recent average activation value, and nodes whose activity contribution is consistently lower than the pruning threshold are removed. The monitoring module is used to monitor the network after training. When the sliding window variance of the error between the predicted value and the actual value exceeds a preset multiple of the historical baseline variance, the time exponential decay factor is increased, and the reconstruction of the weighted covariance matrix and the adjustment of the hidden layer nodes of the network are triggered.

9. The system according to claim 8, characterized in that, The construction of a covariance matrix weighted by a time-exponential decay factor, calculation of the variance contribution rate of each principal component, and determination of the time stability coefficient based on the cosine mean of the angle between the projection vectors of the matrix within a continuous time window include: At each time step t, through the recursive formula Update covariance matrix ; Given a continuous time window size L of 20 time steps, for the i-th principal component, the time stability coefficient is... By calculating the feature vector of the principal component at the current time step The feature vectors of the past L time steps The mean cosine similarity is obtained using the formula: ;in, It is the time-exponential decay factor.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images