A dry body weight dynamic prediction method and system based on uric acid kinetics and machine learning

Abstract

The application discloses a dry body weight dynamic prediction method and system based on uric acid kinetics and machine learning. The method comprises the following steps: acquiring blood test data, dialysis treatment data and historical data of a patient in a current dialysis period, and dividing the data into a training set and a test set; calculating a static dry body weight reference value by using a uric acid kinetics model according to the blood test data and the dialysis treatment data; constructing an initial time series machine learning model, training the model by using the training set to obtain a trained time series machine learning model; inputting the static dry body weight reference value and the test set into the trained time series machine learning model to obtain a dynamic dry body weight recommended value and output the dynamic dry body weight recommended value, and outputting the static dry body weight reference value as a reference. The application realizes dynamic prediction and individualized iterative update of the dry body weight by deep fusion of the uric acid kinetics mechanism and the machine learning time series prediction, and has high prediction accuracy without the need of special equipment.

Description

Technical Field

[0001] This invention relates to the field of volume management for hemodialysis patients, specifically to a method and system for dynamic prediction of dry weight based on uric acid kinetics and machine learning. Background Technology

[0002] Dry weight (DW) refers to the ideal body weight of a hemodialysis patient when there is neither water retention nor dehydration, and it is the core basis for developing a dialysis ultrafiltration plan. Accurate assessment of dry weight is of great significance for preventing intradialysis hypotension, heart failure, and long-term cardiovascular complications.

[0003] Current dry weight assessment techniques are mainly divided into three categories: The first category is based on clinical experience and physical examination, but relies on the doctor's subjective judgment (such as edema, blood pressure, jugular venous distension, etc.), resulting in poor accuracy and real-time performance, and failing to identify hidden volume overload. The second category is bioelectrical impedance analysis (BIA), which requires specialized equipment and is easily affected by factors such as edema and hypoalbuminemia. The third category is an assessment method based on uric acid kinetic models, which calculates the uric acid distribution volume by changing the blood uric acid concentration before and after dialysis, and then estimates dry weight. However, this method does not require specialized equipment, and the output is only an instantaneous estimate of dry weight for the current dialysis cycle, lacking the ability to track dynamic changes in the patient's volume status.

[0004] In clinical practice, a patient's volume status is affected by various factors such as diet, water intake, residual renal function, and cardiac function, exhibiting significant temporal fluctuations. Static assessment results often lag behind actual changes in the patient's condition, failing to provide physicians with forward-looking support for ultrafiltration decision-making.

[0005] In addition, existing technologies do not fully utilize multi-dimensional time-series clinical information such as blood pressure and ultrafiltration rate, nor do they deeply integrate physiological mechanism models with data-driven models, resulting in insufficient personalization and dynamic adaptability of assessment results. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of the prior art by proposing a dynamic prediction method for dry weight based on uric acid kinetics and machine learning, comprising the following steps: S1. Obtain the patient's blood test data, dialysis treatment data, and historical data from previous dialysis cycles in the current dialysis cycle; and divide the historical data from previous dialysis cycles into a training set and a test set according to a preset ratio; S2. Based on blood test data and dialysis treatment data, the static dry weight baseline value was calculated using a uric acid kinetic model; S3. Construct an initial temporal machine learning model, and train the initial temporal machine learning model using the training set to obtain a trained temporal machine learning model; S4. Using the static dry weight baseline and the test set as input, the dynamic dry weight recommended value is predicted and output through a trained time-series machine learning model. At the same time, the static dry weight baseline is output as a reference baseline to adjust the treatment plan.

[0007] Furthermore, historical data includes historical blood test data, historical time-series clinical characteristic data, and dry weight labels when patients reached a comfortable state; The time-series clinical characteristics data include, but are not limited to: predialysis blood pressure, postdialysis blood pressure, ultrafiltration volume, interdialysis weight gain, BNP level, edema grade, heart rate variability, and previous dry weight sequence. A patient's comfort state is indicated by the following: no signs of volume overload, no symptoms of hypoperfusion, no muscle pain / spasm, SBP fluctuations of <15-20 mmHg throughout dialysis without the need for frequent intervention, and resting heart rate within the normal range without compensatory tachycardia.

[0008] Furthermore, the baseline value of static dry weight was calculated using a uric acid kinetic model, specifically including: Calculate the uric acid distribution volume after dialysis based on the pre-dialysis plasma uric acid concentration, post-dialysis plasma uric acid concentration, pre-dialysis sodium concentration, post-dialysis sodium concentration, dialysis duration, total water removal volume, dialyzer membrane area, and dialysate sodium concentration. Calculate the standardized uric acid distribution volume based on the uric acid distribution volume and body surface area after dialysis; The static dry weight baseline value is obtained by dividing the standardized uric acid distribution volume by a preset percentage.

[0009] Furthermore, S3 specifically includes: The initial time series machine learning model was constructed using an LSTM model; The training steps for a time-series machine learning model include: For each historical dialysis cycle, the corresponding static dry weight baseline value is calculated using the uric acid kinetic model based on its historical blood test data. The static dry weight baseline value of each historical dialysis cycle and historical time-series clinical feature data were used as input features for training samples, and the corresponding dry weight labels were used as training labels to train the time-series machine learning model, resulting in a well-trained time-series machine learning model.

[0010] Furthermore, the method also includes: After a patient finishes dialysis, the newly acquired data is automatically added to the training set to incrementally learn and update the trained time-series machine learning model, adapting to the changing trends of the patient's individual condition.

[0011] A dry weight dynamic prediction system based on uric acid kinetics and machine learning, used to implement any of the above methods, includes: The data acquisition module is used to acquire patients’ blood test data and dialysis treatment data in the current dialysis cycle, as well as time-series clinical characteristic data in multiple historical dialysis cycles. The kinetics calculation module is used to calculate the static dry weight baseline value based on blood test data and dialysis treatment data using a uric acid kinetic model. The model building and training module is used to build an initial temporal machine learning model and train the initial temporal machine learning model. The dynamic prediction and output module is used to input static dry weight baseline values ​​and time-series clinical feature data into a pre-trained time-series machine learning model to predict and output the recommended dynamic dry weight value for the current dialysis cycle.

[0012] Furthermore, the system also includes: The model update module is used to automatically add newly generated clinical data to the training set after each patient's dialysis session, and to incrementally learn and update the time series machine learning model.

[0013] Compared with the prior art, the present invention achieves the following technical effects: First, it upgrades from static assessment to dynamic prediction, enabling real-time reflection of changes in patient volume status and early warning of volume imbalance. Second, by incorporating multi-dimensional time-series clinical data, it utilizes time-series clinical characteristics to dynamically correct kinetic output, significantly improving dry weight prediction accuracy compared to pure kinetic models. Third, it eliminates the need for dedicated impedance equipment, requiring only routine blood biochemistry data, reducing medical costs and broadening its applicability. Fourth, the machine learning model can autonomously learn the volume change patterns of each patient based on their historical data, achieving personalized predictions without requiring patient cluster analysis. Fifth, the time-series machine learning model supports incremental online learning, continuously absorbing data from each patient's dialysis sessions to achieve individualized adaptive iteration. Sixth, the system can provide early warnings of the risk of volume overload or underload, assisting physicians in dynamically adjusting ultrafiltration protocols. Attached Figure Description

[0014] For ease of explanation, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0015] Figure 1 A flowchart of a method for dynamic prediction of dry weight based on uric acid kinetics and machine learning is provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the overall module architecture of a dry weight dynamic prediction system based on uric acid kinetics and machine learning, provided for an embodiment of the present invention. Detailed Implementation

[0016] The following are specific embodiments of the present invention, described in conjunction with the accompanying drawings, to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments. Specific details, such as particular configurations, are provided in the following description merely to aid in a comprehensive understanding of the embodiments of the present invention. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention.

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0018] Example 1 This invention discloses a method for dynamic prediction of dry weight based on uric acid kinetics and machine learning, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the patient's blood test data and dialysis treatment data in the current dialysis cycle, as well as historical data from previous dialysis cycles; and divide the historical data from previous dialysis cycles into a training set and a test set according to a preset ratio.

[0019] Specifically, blood test data and dialysis treatment data include, but are not limited to, predialysis serum uric acid concentration. Postdialysis serum uric acid concentration Predialysis serum sodium concentration Postdialysis serum sodium concentration Total water removal volume of dialysis Dialysis duration , the permeate membrane area A.

[0020] Historical data for dialysis cycles includes historical blood test data, historical time-series clinical characteristic data, and dry weight labels when patients reach a comfortable state.

[0021] The time-series clinical characteristics data include, but are not limited to: predialysis blood pressure, postdialysis blood pressure, ultrafiltration volume, interdialysis weight gain, BNP level, edema grade, heart rate variability, and previous dry weight sequence.

[0022] A patient's comfort state is indicated by the following: no signs of volume overload, i.e., no pulmonary rales, no jugular venous distension, and no symptoms of acute left heart failure; no symptoms of hypoperfusion, i.e., no dizziness, no blurred vision, no cold sweats, no altered consciousness, and no muscle pain / spasm, i.e., no severe cramps or limb numbness; stable blood pressure, with SBP fluctuations of <15-20 mmHg throughout dialysis, requiring no frequent intervention; and stable heart rate, with resting heart rate within the normal range and without compensatory tachycardia.

[0023] In this embodiment, historical data from historical dialysis cycles are randomly divided into a training set and a test set in an 8:2 ratio. The training set is used for model training, and the test set is used to evaluate the model's predictive performance.

[0024] S2. Based on blood test data and dialysis treatment data, the static dry weight baseline value is calculated using a uric acid kinetic model.

[0025] Specifically, the process of analyzing and calculating the uric acid distribution volume based on the kinetic model is as follows: The amount of uric acid removed by dialysis is the difference between the product of the predialysis uric acid distribution volume and the predialysis plasma uric acid concentration, and the product of the postdialysis uric acid distribution volume and the plasma uric acid concentration that rebounds after dialysis.

[0026] Based on existing research, uric acid does not pass through cell membranes during dialysis; therefore, the uric acid removed during dialysis is only that in the extracellular fluid. Further, it can be deduced that uric acid does not flow from inside the cell to the extracellular fluid during dialysis, so theoretically, the volume of distribution of uric acid is equal to the volume of extracellular fluid. The volume of distribution of uric acid before dialysis should be the sum of the volume of distribution of uric acid after dialysis, the total amount of water removed during dialysis, and the amount of water flowing into the cell due to changes in serum sodium. Therefore, the formula for calculating the volume of distribution of uric acid after dialysis can be derived as follows: (1) in, This indicates the uric acid distribution volume after dialysis. This indicates the amount of uric acid removed during dialysis. This indicates the total amount of water removed by dialysis. This indicates the amount of water flowing into cells due to changes in serum sodium levels during dialysis. This indicates the plasma uric acid concentration before dialysis. This indicates the plasma uric acid concentration that rebounds after dialysis.

[0027] The amount of water flowing into cells due to changes in serum sodium during dialysis is calculated as follows: (2) in, It is the urea distribution volume after dialysis. It is the sodium concentration after dialysis. It refers to the sodium concentration before dialysis; the urea distribution volume after dialysis is calculated based on the most recent biochemical test data.

[0028] Uric acid removal through dialysis The determination method is as follows: Based on the kinetic model of uric acid, the product of the plasma uric acid concentration and the uric acid clearance rate of the dialyzer is calculated every minute. The amount of uric acid cleared every minute from the start to the end of dialysis is summed to obtain the amount of uric acid cleared during dialysis. .

[0029] The uric acid clearance rate of the dialyzer is calculated by substituting the plasma flow rate, dialysate flow rate, and uric acid-related area solute transport coefficient into the Michaelis formula.

[0030] Rebound in plasma uric acid concentration after dialysis The result is calculated using Tattersall's formula, as follows: (3) in, tp This is a standard coefficient, based on research on uric acid. tp It lasted 64 minutes.

[0031] After obtaining the post-dialysis uric acid distribution volume, the standardized uric acid distribution volume is calculated based on the post-dialysis uric acid distribution volume and body surface area. The standardized uric acid distribution volume is calculated as follows: (4) in, To standardize the uric acid distribution volume, BSA This represents the surface area of ​​the body after dialysis.

[0032] The body surface area for uric acid production after dialysis is calculated as follows: (5) Where W represents the patient's weight and H represents the patient's height.

[0033] Statistical analysis of standardized uric acid distribution volume (UDI) in dehydrated, normal, and water-retaining patients yielded a reasonable range for standardized UDI in patients with ideal dry weight, enabling the assessment of dry weight using standardized UDI. UDI is 27% of ideal body weight; therefore, dry weight can be calculated by dividing UDI by 27%. The static dry weight baseline value is calculated as follows: (6) in, This is the baseline value for static dry weight.

[0034] S3. Construct an initial temporal machine learning model, and train the initial temporal machine learning model using the training set to obtain a trained temporal machine learning model.

[0035] Specifically, the initial temporal machine learning model is constructed using an LSTM (Long Short-Term Memory) network, a GRU (Gated Recurrent Unit), or a temporal Transformer model. In this embodiment, an LSTM model is preferred for constructing the initial temporal machine learning model.

[0036] After the initial time series machine learning model is built, the model is trained.

[0037] First, for each historical dialysis cycle in the training set, the corresponding static dry weight baseline value is calculated using the uric acid kinetic model based on its blood test data and dialysis treatment data, as described in step S2.

[0038] Secondly, using the static dry weight baseline values ​​and historical time-series clinical feature data of each historical dialysis cycle as input features for training samples, a time-series input sequence of length N (N ranges from 4 to 8) is constructed, with the corresponding dry weight label used as the training label. A sliding window mechanism is used to generate training samples: for the t-th dialysis cycle, the input features include the static dry weight baseline values ​​from the (t-N+1)-th to the t-th dialysis cycle and the corresponding time-series clinical feature data for each dialysis cycle, and the output label is the dry weight label for the (t+1)-th dialysis cycle.

[0039] Mean squared error (MSE) was used as the loss function, and the model was trained using the backpropagation algorithm and the Adam optimizer. The initial learning rate was set to 0.001, the batch size to 32, and the number of training epochs to 100. Training was stopped early when the validation set loss no longer decreased for 10 consecutive epochs to prevent overfitting. After training, the model parameters were saved to obtain the trained time-series machine learning model.

[0040] After each dialysis session, the newly acquired data is automatically added to the training set to incrementally learn and update the model. This involves using new samples to fine-tune and update the LSTM model, enabling it to continuously adapt to changes in the individual patient's condition.

[0041] S4. Using the static dry weight baseline and the test set as input, a trained time-series machine learning model is used to predict the recommended dynamic dry weight. It also outputs the static dry weight baseline value. As a reference benchmark. Recommended values ​​based on dynamic dry weight. Compared with the baseline value of static dry weight The treatment plan was adjusted, and the recommended dynamic dry weight value was adjusted as treatment progressed. Gradually fine-tune to reach the static dry weight baseline value .

[0042] This invention employs a two-layer fusion architecture. The first layer utilizes a uric acid kinetic model to calculate a baseline static dry weight based on blood test data and dialysis treatment data from the current dialysis cycle. The second layer will use the static dry weight baseline value. The model inputs current and historical time-series clinical feature data into a pre-trained time-series machine learning model, which outputs a dynamic recommended dry weight value. Simultaneously output the static dry weight baseline value. It serves as a clinical reference benchmark, allowing doctors to compare and analyze data to adjust treatment plans.

[0043] Example 2 This invention discloses a dry weight dynamic prediction system based on uric acid kinetics and machine learning, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire patients’ blood test data and dialysis treatment data in the current dialysis cycle, as well as time-series clinical characteristic data in multiple historical dialysis cycles. The kinetics calculation module is used to calculate the static dry weight baseline value based on blood test data and dialysis treatment data using a uric acid kinetic model. The model building and training module is used to build an initial temporal machine learning model and train the initial temporal machine learning model. The dynamic prediction and output module is used to input static dry weight baseline values ​​and time-series clinical feature data into a pre-trained time-series machine learning model to predict and output the recommended dynamic dry weight value for the current dialysis cycle.

[0044] In addition, the system also includes a model update module, which automatically adds newly generated clinical data to the training set after each patient's dialysis session to incrementally learn and update the time-series machine learning model.

[0045] Those skilled in the art to which this invention pertains may make various modifications or additions to the specific embodiments described or use similar methods to replace them, without departing from the inventive concept of this invention or exceeding the scope defined by the appended claims.

Claims

1. A method for dynamic prediction of dry weight based on uric acid kinetics and machine learning, characterized in that, Includes the following steps: S1. Obtain the patient's blood test data, dialysis treatment data, and historical data from previous dialysis cycles in the current dialysis cycle; and divide the historical data from previous dialysis cycles into a training set and a test set according to a preset ratio; S2. Based on the blood test data and dialysis treatment data, the static dry weight baseline value is calculated using the uric acid kinetic model; S3. Construct an initial temporal machine learning model, and train the initial temporal machine learning model using the training set to obtain a trained temporal machine learning model; S4. Using the static dry weight baseline value and the test set as input, predict and output the dynamic dry weight recommended value through the trained time-series machine learning model, and at the same time output the static dry weight baseline value as a reference to adjust the treatment plan.

2. The method for dynamic prediction of dry weight based on uric acid kinetics and machine learning according to claim 1, characterized in that, The historical data includes historical blood test data, historical time-series clinical characteristic data, and dry weight labels when the patient reached a comfortable state. The time-series clinical characteristics data include, but are not limited to: predialysis blood pressure, postdialysis blood pressure, ultrafiltration volume, interdialysis weight gain, BNP level, edema grade, heart rate variability, and previous dry weight sequence. The patient's comfortable state is defined as follows: the patient has no signs of volume overload, no symptoms of hypoperfusion, no muscle pain / spasm, SBP fluctuations of <15-20 mmHg throughout dialysis, no need for frequent intervention, and resting heart rate within the normal range without compensatory tachycardia.

3. The method for dynamic prediction of dry weight based on uric acid kinetics and machine learning according to claim 1, characterized in that, The static dry weight baseline value calculated using the uric acid kinetic model specifically includes: Calculate the uric acid distribution volume after dialysis based on the pre-dialysis plasma uric acid concentration, post-dialysis plasma uric acid concentration, pre-dialysis sodium concentration, post-dialysis sodium concentration, dialysis duration, total water removal volume, dialyzer membrane area, and dialysate sodium concentration. Calculate the standardized uric acid distribution volume based on the uric acid distribution volume and body surface area after dialysis; The static dry weight baseline value is obtained by dividing the standardized uric acid distribution volume by a preset percentage.

4. The method for dynamic prediction of dry weight based on uric acid kinetics and machine learning according to claim 1, characterized in that, S3 specifically includes: The initial time series machine learning model was constructed using an LSTM model; The training steps for the time-series machine learning model include: For each historical dialysis cycle, the corresponding static dry weight baseline value is calculated using the uric acid kinetic model based on its historical blood test data. The time-series machine learning model is trained by using the static dry weight baseline value of each historical dialysis cycle and historical time-series clinical feature data as training sample input features, and the corresponding dry weight label as training label.

5. The method for dynamic prediction of dry weight based on uric acid kinetics and machine learning according to claim 1, characterized in that, The method further includes: After a patient completes dialysis, the newly acquired data is automatically added to the training set to incrementally learn and update the trained time-series machine learning model, adapting to the changing trends of the patient's individual state.

6. A dry weight dynamic prediction system based on uric acid kinetics and machine learning, characterized in that, To implement any one of the methods described in claims 1-5, the method comprises: The data acquisition module is used to acquire patients’ blood test data and dialysis treatment data in the current dialysis cycle, as well as time-series clinical characteristic data in multiple historical dialysis cycles. The kinetic calculation module is used to calculate the static dry weight baseline value based on the blood test data and dialysis treatment data using the uric acid kinetic model; The model building and training module is used to build an initial temporal machine learning model and train the initial temporal machine learning model. The dynamic prediction and output module is used to input the static dry weight baseline value and the time-series clinical feature data into a pre-trained time-series machine learning model, predict and output the recommended dynamic dry weight value for the current dialysis cycle.

7. The dry weight dynamic prediction system based on uric acid kinetics and machine learning according to claim 6, characterized in that, The system also includes: The model update module is used to automatically add newly generated clinical data to the training set after each patient's dialysis session, and to incrementally learn and update the time-series machine learning model.

Images