A machine learning-based prediction system for hypotension during hemodialysis

Abstract

The present invention discloses a system for predicting hypotension during hemodialysis based on machine learning, comprising: an information entry module, a patient clustering module, a feature screening module, and a risk prediction module. The risk prediction module completes the construction of the prediction system by constructing a static data risk prediction model and a dynamic data risk prediction model; the system achieves a technology for predicting the risk of hypotension in advance when the patient undergoes hemodialysis, solving the problem that the occurrence of hypotension cannot be predicted in advance during hemodialysis treatment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medical technology, and in particular to a machine learning-based system for predicting hypotension during hemodialysis. Background Art

[0002] Maintenance hemodialysis is the primary renal replacement therapy for patients with end-stage renal disease (ESRD). However, many hemodialysis patients experience intradialytic hypotension during dialysis, and its incidence fluctuates widely due to the lack of a standardized definition. It is important to note that the occurrence of hypotension is closely associated with adverse outcomes in maintenance hemodialysis patients, including an increased risk of cardiovascular and cerebrovascular disease, accelerated loss of residual renal function, inadequate dialysis dose, and vascular access thrombosis. Therefore, accurate prediction and timely intervention of intradialytic hypotension events are crucial for the management of maintenance hemodialysis patients. Despite this, due to the complex etiology of intradialytic hypotension, clinical prediction relies primarily on empirical experience and subjective judgment, which carries certain uncertainties and limitations. Traditional statistical methods, such as logistic regression, have been used in medical research to analyze and develop predictive models. In recent years, the rapid development of big data and artificial intelligence (AI) technologies in the medical field has provided new opportunities for improving the adverse complications of patients with ESRD. The development of hypotension prediction models based on machine learning and deep learning has become a research hotspot in the hemodialysis field. Compared with traditional statistical methods, they can better explore potential patterns and regularities in clinical data and provide more reliable support for clinical decision-making.

[0003] However, existing prediction models also have some issues. For example, they only consider the degree of blood pressure drop while ignoring machine parameter predictions, and fail to comprehensively consider factors such as medication use, accompanying clinical symptoms, and the need for medical intervention. More importantly, most of these methods directly model the data of all patients, ignoring individual differences and heterogeneity among patients. This limits the model's generalization ability and clinical application effectiveness. Summary of the Invention

[0004] Existing technologies focus solely on the degree of blood pressure drop while ignoring machine parameter predictions and failing to comprehensively consider factors such as medication use, accompanying clinical symptoms, and the need for medical intervention. Most of these methods directly model the data of all patients, ignoring individual differences and heterogeneity among patients. This limits the model's generalizability and clinical effectiveness.

[0005] In order to solve the above problems, the present invention provides a machine learning-based hemodialysis hypotension prediction system, comprising:

[0006] M1: Information entry module; M2: Patient clustering module; M3: Feature screening module; M4: Risk prediction module.

[0007] In a preferred embodiment, M1: information entry module, including: entering the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis;

[0008] M2: Patient clustering module, including:

[0009] M2-1: Set the K-means clustering model parameter K, where K represents the number of clusters given in advance. Enter the patient's basic information as input features. Each patient is considered as , K-means algorithm clusters samples into k clusters;

[0010] M2-2: Randomly select k cluster centroids ;

[0011] M2-3: Repeat the following process until convergence;

[0012] M2-3-1: For each example i, calculate the class it should belong to. The formula is as follows:

[0013]

[0014] M2-3-2: For each class j, recalculate the centroid of the class using the following formula:

[0015]

[0016] Where, Indicates the class closest to sample i among the k classes, The centroid represents the guess of the center point of the samples belonging to the same class. Iterations M2-3 are repeated until the centroid remains unchanged or the moving distance of the centroid between two iterations is less than the preset threshold. The threshold is set to 0.001 in the experiment.

[0017] M3: Feature screening module, including: feature screening based on the basic information of the patient's calculator and pre-dialysis measurement data to obtain feature types for hypotension risk prediction;

[0018] M3-1: Based on the basic information of multiple patients and pre-dialysis measurement data as feature input, static features are screened using L1 regularization and random forest for calculating feature importance;

[0019] Lasso promotes the sparseness of some feature coefficients by introducing L1 regularization. Lasso penalizes the regression coefficients according to their absolute values ​​and forces some coefficients to be zero, ultimately retaining the most useful features. The objective function of Lasso regression is as follows:

[0020]

[0021] In the formula, λ is the regularization parameter, β is the regression coefficient, and β0 represents the intercept of the regression line on the y-axis. is the L1 norm penalty term;

[0022] Random forest compares the contribution of each feature to each tree in the forest, where the contribution is measured using the Gini index as an evaluation indicator. The variable importance score is represented by VIM, and the Gini index is represented by GI. Suppose there are J features X1, X2, X3, ..., XJ, I decision trees, and C categories. Now we need to calculate the Gini index score of each feature Xj. , the calculation formula of the Gini index of the i-th tree node q is:

[0023]

[0024] The conclusions drawn from the two feature selection methods were compared with the experimental groups of the two feature selection methods. Finally, the random forest was used to select the top 10 contributing features, and the Lasso algorithm was used to determine whether there were other features that needed to be introduced. The specific method was as follows: the absolute value of the regression coefficient β was penalized according to the size of the regularization parameter λ in the Lasso formula, and some coefficients β were forced to be zero. In the experiment, the value of λ ranged from 10-5 to 102, and each λ value was traversed and the Lasso model was trained. Cross-validation was used to select the optimal λ value. The maximum number of iterations was 10,000. Finally, the optimal λ value was returned, and the most useful features were retained.

[0025] M4: Risk prediction module, including: using the hemodialysis hypotension prediction model to predict the patient's risk of hypotension events during this hemodialysis based on the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis;

[0026] M4-1: Construct a static data risk prediction model;

[0027] Use logistic regression to build a linear model, that is, linearly combine input features. The output of the linear model is mapped to a probability value. Use the Sigmoid function to convert the output value of the linear model into a probability value between 0 and 1. The formula is as follows:

[0028]

[0029] Where, Represents the function of the sample point x about θ, θ represents a vector, and g represents the logical function. In the binary classification problem, the goal of the support vector machine is to find a hyperplane that can separate the two types of samples. Assume that the formula of a hyperplane H is:

[0030]

[0031] Where w is the weight and b is a constant. By dividing by this linear equation, there are two hyperplanes H1 and H2 parallel to H:

[0032]

[0033] The hyperplane H classification satisfies the following constraints:

[0034]

[0035] The support vector machine finds the best separating hyperplane by maximizing the distance between the hyperplane and the training sample closest to it;

[0036] M4-2: Constructing a dynamic data risk prediction model;

[0037] The long short-term memory neural network model is a variant of the recurrent neural network structure. It processes sequences of arbitrary length by cyclically using the same network neurons. It introduces a gating mechanism, including an input gate, a forget gate, and an output gate, which passes the previous feature of the sequence to the next as input, thereby capturing the complete feature information before and after the sequence. The equations for the input gate, forget gate, and output gate are as follows:

[0038]

[0039] Among them, w is the weight, x is the input, b is a constant, and h represents the hidden state of the current input; for each unit, its input and output are as follows:

[0040]

[0041] in, is the cell state at timestamp (t), is the candidate for the cell state at timestamp (t); BiLSTM is an extension of LSTM that processes both past and future information at each time step. It includes two hidden layers that process forward and reverse sequence data respectively, and finally merges the information in the two directions to obtain a more comprehensive sequence representation;

[0042] Calculate the forward hidden sequence , then calculate the background sequence , and get an output: This is the predicted risk of the model output;

[0043]

[0044] GRU is a variant of LSTM. It simplifies the gating mechanism of LSTM, merges the forget gate and input gate in LSTM into an update gate, and introduces a reset gate to control the flow from the current input to the hidden state. The formula of the update gate is:

[0045]

[0046] The formula for resetting the gate is:

[0047]

[0048] Among them, W is the weight matrix and U is the unit matrix;

[0049] M4-3: Calculate bias values ​​and weights;

[0050] enter ,in, is the preliminary prediction result matrix of the static data risk prediction model and the dynamic data risk prediction model, Given feature weights to determine whether hypotension actually occurs And the bias value b, referring to the Bernoulli distribution, the probability of the sample belonging to each category is expressed as:

[0051]

[0052] The feature weights and bias values ​​in each cluster are collected separately. After obtaining the preliminary prediction results, the weights and bias values ​​are added to perform weighted integration on the prediction results, and the final prediction results are output. The weights and bias values ​​are output through logistic regression.

[0053] In the preferred method, the basic information of the patient entered includes: age, gender, dialysis age, and BMI, a total of 4 data.

[0054] In a preferred manner, the measurement data before dialysis include: height measured on the day, weight measured on the day, percentage of weight loss during dialysis, percentage of weight gain compared with the last dialysis, weight gain, weight loss during dialysis, heart rate before dialysis, diastolic blood pressure before dialysis, systolic blood pressure before dialysis, pulse pressure difference, and accurate ultrafiltration rate, a total of 11 data.

[0055] In a preferred embodiment, the sensor machine data during dialysis include: arterial pressure AP, venous pressure VP, real-time blood flow, average blood flow, dialysate flow, temperature and dialysate conductivity, a total of 7 data.

[0056] In the preferred embodiment, step M2-3-1 calculates the distance from each patient data to each of the k centroids, and then selects the cluster with the closest distance as ; Step M2-3-2 For each cluster, recalculate its centroid .

[0057] Beneficial effects of the present invention: By constructing a cardiovascular event prediction system for hemodialysis patients based on machine learning, the present invention has achieved the technology of predicting the risk of hypotension in advance when the patient is undergoing hemodialysis, thus solving the problem that the occurrence of hypotension cannot be predicted in advance during hemodialysis treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the system framework of the machine learning-based cardiovascular event prediction system for hemodialysis patients;

[0059] Figure 2 Schematic diagram of the experimental data acquisition method of the present invention;

[0060] Figure 3 This is a schematic diagram of the modeling process of the hypotensive event prediction model of the present invention. DETAILED DESCRIPTION

[0061] Example 1:

[0062] A machine learning-based system for predicting hypotension during hemodialysis, comprising:

[0063] M1: Information entry module, including: entering the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis;

[0064] M2: Patient clustering module, including:

[0065] M2-1: Set the K-means clustering model parameter K, where K represents the number of clusters given in advance. Enter the patient's basic information as input features. Each patient is considered as , K-means algorithm clusters samples into k clusters;

[0066] M2-2: Randomly select k cluster centroids ;

[0067] M2-3: Repeat the following process until convergence;

[0068] M2-3-1: For each example i, calculate the class it should belong to. The formula is as follows:

[0069]

[0070] M2-3-2: For each class j, recalculate the centroid of the class using the following formula:

[0071]

[0072] Where, Indicates the class closest to sample i among the k classes, The centroid represents the guess of the center point of the samples belonging to the same class. Iterations M2-3 are repeated until the centroid remains unchanged or the moving distance of the centroid between two iterations is less than the preset threshold. The threshold is set to 0.001 in the experiment.

[0073] M3: Feature screening module, including: feature screening based on the basic information of the patient's calculator and pre-dialysis measurement data to obtain feature types for hypotension risk prediction;

[0074] M3-1: Based on the basic information of multiple patients and pre-dialysis measurement data as feature input, static features are screened using L1 regularization and random forest for calculating feature importance;

[0075] Lasso promotes the sparseness of some feature coefficients by introducing L1 regularization. Lasso penalizes the regression coefficients according to their absolute values ​​and forces some coefficients to be zero, ultimately retaining the most useful features. The objective function of Lasso regression is as follows:

[0076]

[0077] In the formula, λ is the regularization parameter, β is the regression coefficient, and β0 represents the intercept of the regression line on the y-axis. is the L1 norm penalty term;

[0078] Random forest compares the contribution of each feature to each tree in the forest, where the contribution is measured using the Gini index as an evaluation indicator. The variable importance score is represented by VIM, and the Gini index is represented by GI. Suppose there are J features X1, X2, X3, ..., XJ, I decision trees, and C categories. Now we need to calculate the Gini index score of each feature Xj. , the calculation formula of the Gini index of the i-th tree node q is:

[0079]

[0080] The conclusions drawn from the two feature selection methods were compared with the experimental groups of the two feature selection methods. Finally, the random forest was used to select the top 10 contributing features, and the Lasso algorithm was used to determine whether there were other features that needed to be introduced. The specific method was as follows: the absolute value of the regression coefficient β was penalized according to the size of the regularization parameter λ in the Lasso formula, and some coefficients β were forced to be zero. In the experiment, the value of λ ranged from 10-5 to 102, and each λ value was traversed and the Lasso model was trained. Cross-validation was used to select the optimal λ value. The maximum number of iterations was 10,000. Finally, the optimal λ value was returned, and the most useful features were retained.

[0081] M4: Risk prediction module, including: using the hemodialysis hypotension prediction model to predict the patient's risk of hypotension events during this hemodialysis based on the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis;

[0082] M4-1: Construct a static data risk prediction model;

[0083] Use logistic regression to build a linear model, that is, linearly combine input features. The output of the linear model is mapped to a probability value. Use the Sigmoid function to convert the output value of the linear model into a probability value between 0 and 1. The formula is as follows:

[0084]

[0085] Where, Represents the function of the sample point x about θ, θ represents a vector, and g represents the logical function. In the binary classification problem, the goal of the support vector machine is to find a hyperplane that can separate the two types of samples. Assume that the formula of a hyperplane H is:

[0086]

[0087] Where w is the weight and b is a constant. By dividing by this linear equation, there are two hyperplanes H1 and H2 parallel to H:

[0088]

[0089] The hyperplane H classification satisfies the following constraints:

[0090]

[0091] The support vector machine finds the best separating hyperplane by maximizing the distance between the hyperplane and the training sample closest to it;

[0092] M4-2: Constructing a dynamic data risk prediction model;

[0093] The long short-term memory neural network model is a variant of the recurrent neural network structure. It processes sequences of arbitrary length by cyclically using the same network neurons. It introduces a gating mechanism, including an input gate, a forget gate, and an output gate, which passes the previous feature of the sequence to the next as input, thereby capturing the complete feature information before and after the sequence. The equations for the input gate, forget gate, and output gate are as follows:

[0094]

[0095] Among them, w is the weight, x is the input, b is a constant, and h represents the hidden state of the current input; for each unit, its input and output are as follows:

[0096]

[0097] in, is the cell state at timestamp (t), is the candidate for the cell state at timestamp (t); BiLSTM is an extension of LSTM that processes both past and future information at each time step. It includes two hidden layers that process forward and reverse sequence data respectively, and finally merges the information in the two directions to obtain a more comprehensive sequence representation;

[0098] Calculate the forward hidden sequence , then calculate the background sequence , and get an output: This is the predicted risk of the model output;

[0099]

[0100] GRU is a variant of LSTM. It simplifies the gating mechanism of LSTM, merges the forget gate and input gate in LSTM into an update gate, and introduces a reset gate to control the flow from the current input to the hidden state. The formula of the update gate is:

[0101]

[0102] The formula for resetting the gate is:

[0103]

[0104] Among them, W is the weight matrix and U is the unit matrix;

[0105] M4-3: Calculate bias values ​​and weights;

[0106] enter ,in, is the preliminary prediction result matrix of the static data risk prediction model and the dynamic data risk prediction model, Given feature weights to determine whether hypotension actually occurs And the bias value b, referring to the Bernoulli distribution, the probability of the sample belonging to each category is expressed as:

[0107]

[0108] The feature weights and bias values ​​in each cluster are collected separately. After obtaining the preliminary prediction results, the weights and bias values ​​are added to perform weighted integration on the prediction results, and the final prediction results are output. The weights and bias values ​​are output through logistic regression.

[0109] The patient's basic information entered includes: age, gender, dialysis age, and BMI, a total of 4 data.

[0110] The measurement data before dialysis included: height measured on the day, weight measured on the day, percentage of weight loss during dialysis, percentage of weight gain compared with the last dialysis, weight gain, weight loss during dialysis, heart rate before dialysis, diastolic blood pressure before dialysis, systolic blood pressure before dialysis, pulse pressure difference, and accurate ultrafiltration rate, a total of 11 data.

[0111] The sensor machine data during dialysis include: arterial pressure AP, venous pressure VP, real-time blood flow, average blood flow, dialysate flow, temperature and dialysate conductivity, a total of 7 data.

[0112] Step M2-3-1 calculates the distance from each patient data to each of the k centroids, and then selects the cluster with the closest distance as ; Step M2-3-2 For each cluster, recalculate its centroid .

[0113] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based hemodialysis hypotension prediction system, characterized by: include: M1: information entry module; M2: patient clustering module; M3 : Feature screening module; M4: risk prediction module; M1 : Information entry module, including: entering the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis; M2: Patient clustering module, including: M2-1: Set the K-means clustering model parameter K, where K represents the number of clusters given in advance. Enter the patient's basic information as input features. Each patient is considered as , K-means algorithm clusters samples into k clusters; M2-2: Randomly select k cluster centroids ; M2-3: Repeat the following process until convergence; M2-3-1: For each example i, calculate the class it should belong to. The formula is as follows: M2-3-2: For each class j, recalculate the centroid of the class using the following formula: Where, Indicates the class closest to sample i among the k classes, The centroid represents the guess of the center point of the samples belonging to the same class. Iterations M2-3 are repeated until the centroid remains unchanged or the moving distance of the centroid between two iterations is less than the preset threshold. The threshold is set to 0.001 in the experiment. M3: Feature screening module, including: feature screening based on the basic information of the patient's calculator and pre-dialysis measurement data to obtain feature types for hypotension risk prediction; M3-1: Based on the basic information of multiple patients and pre-dialysis measurement data as feature input, static features are screened using L1 regularization and random forest for calculating feature importance; Lasso promotes the sparseness of some feature coefficients by introducing L1 regularization. Lasso penalizes the regression coefficients according to their absolute values ​​and forces some coefficients to be zero, ultimately retaining the most useful features. The objective function of Lasso regression is as follows: In the formula, λ is the regularization parameter, β is the regression coefficient, and β0 represents the intercept of the regression line on the y-axis. is the L1 norm penalty term; Random forest compares the contribution of each feature to each tree in the forest, where the contribution is measured using the Gini index as an evaluation indicator. The variable importance score is represented by VIM, and the Gini index is represented by GI. Suppose there are J features X1, X2, X3, ..., XJ, I decision trees, and C categories. Now we need to calculate the Gini index score of each feature Xj. , the calculation formula of the Gini index of the i-th tree node q is: The conclusions drawn from the two feature selection methods were compared with the experimental groups of the two feature selection methods. Finally, the random forest was used to select the top 10 contributing features, and the Lasso algorithm was used to determine whether there were other features that needed to be introduced. The specific method was as follows: the absolute value of the regression coefficient β was penalized according to the size of the regularization parameter λ in the Lasso formula, and some coefficients β were forced to be zero. In the experiment, the value of λ ranged from 10-5 to 102, and each λ value was traversed and the Lasso model was trained. Cross-validation was used to select the optimal λ value. The maximum number of iterations was 10,000. Finally, the optimal λ value was returned, and the most useful features were retained. M4: Risk prediction module, including: using the hemodialysis hypotension prediction model to predict the patient's risk of hypotension events during this hemodialysis based on the patient's basic information, pre-dialysis measurement data, and sensor machine data during dialysis; M4-1: Construct a static data risk prediction model; Use logistic regression to build a linear model, that is, linearly combine input features. The output of the linear model is mapped to a probability value. Use the Sigmoid function to convert the output value of the linear model into a probability value between 0 and 1. The formula is as follows: Where, Represents the function of the sample point x about θ, θ represents a vector, and g represents the logical function. In the binary classification problem, the goal of the support vector machine is to find a hyperplane that can separate the two types of samples. Assume that the formula of a hyperplane H is: Where w is the weight and b is a constant. By dividing by this linear equation, there are two hyperplanes H1 and H2 parallel to H: The hyperplane H classification satisfies the following constraints: The support vector machine finds the best separating hyperplane by maximizing the distance between the hyperplane and the training sample closest to it; M4-2: Constructing a dynamic data risk prediction model; The long short-term memory neural network model is a variant of the recurrent neural network structure. It processes sequences of arbitrary length by cyclically using the same network neurons. It introduces a gating mechanism, including an input gate, a forget gate, and an output gate, which passes the previous feature of the sequence to the next as input, thereby capturing the complete feature information before and after the sequence. The equations for the input gate, forget gate, and output gate are as follows: Among them, w is the weight, x is the input, b is a constant, and h represents the hidden state of the current input; for each unit, its input and output are as follows: in, is the cell state at timestamp t, is a candidate for the cell state at timestamp t; BiLSTM is an extension of LSTM that processes both past and future information at each time step. It includes two hidden layers that process forward and reverse sequence data respectively, and finally merges the information in the two directions to obtain a more comprehensive sequence representation; Calculate the forward hidden sequence , then calculate the background sequence , and get an output: This is the predicted risk of the model output; GRU is a variant of LSTM. It simplifies the gating mechanism of LSTM, merges the forget gate and input gate in LSTM into an update gate, and introduces a reset gate to control the flow from the current input to the hidden state. The formula of the update gate is: The formula for resetting the gate is: Among them, W is the weight matrix and U is the unit matrix; M4-3: Calculate bias values ​​and weights; enter ,in, is the preliminary prediction result matrix of the static data risk prediction model and the dynamic data risk prediction model, Given feature weights to determine whether hypotension actually occurs And the bias value b, referring to the Bernoulli distribution, the probability of the sample belonging to each category is expressed as: The feature weights and bias values ​​in each cluster are collected separately. After obtaining the preliminary prediction results, the weights and bias values ​​are added to perform weighted integration on the prediction results, and the final prediction results are output. The weights and bias values ​​are output through logistic regression.

2. The machine learning-based hemodialysis hypotension prediction system according to claim 1, characterized in that: The patient's basic information entered includes: age, gender, dialysis age, and BMI, a total of 4 data.

3. The machine learning-based hemodialysis hypotension prediction system according to claim 1, characterized in that: The measurement data before dialysis included: height measured on the day, weight measured on the day, percentage of weight loss during dialysis, percentage of weight gain compared with the last dialysis, weight gain, weight loss during dialysis, heart rate before dialysis, diastolic blood pressure before dialysis, systolic blood pressure before dialysis, pulse pressure difference, and accurate ultrafiltration rate, a total of 11 data.

4. The machine learning-based hemodialysis hypotension prediction system according to claim 1, characterized in that: The sensor machine data during dialysis include: arterial pressure AP, venous pressure VP, real-time blood flow, average blood flow, dialysate flow, temperature and dialysate conductivity, a total of 7 data.

5. The machine learning-based hemodialysis hypotension prediction system according to claim 1, characterized in that: Step M2-3-1 calculates the distance from each patient data to each of the k centroids, and then selects the cluster with the closest distance as ; Step M2-3-2 For each cluster, recalculate its centroid .

Images