Hemodialysis sufficiency prediction method based on transfer learning

Abstract

The invention belongs to the technical field of medical data analysis, and discloses a hemodialysis sufficiency prediction method based on transfer learning, and the method comprises the steps: S1, collecting and processing dialysis data: taking time increment, conductivity and temperature as input variables, and carrying out the standardization processing; and S2, model training: carrying out regression prediction model training on the data processed in the step S1 by using an improved multi-source domain TrAdaboost algorithm. According to the algorithm, variable data are input in multiple dimensions in the dialysis process, the nonlinear relation between the variable data and dialysis sufficiency is learned through a deep learning model, and accurate prediction of the hemodialysis sufficiency is achieved; and 3, dialysis sufficiency prediction: inputting new patient data into the final regression model trained in the step 2, and performing dialysis sufficiency prediction. The dialysis sufficiency is effectively predicted through the conductivity and the temperature.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data analysis, and in particular relates to a hemodialysis adequacy prediction method based on transfer learning. Background Art

[0002] In recent years, with the aggravation of the aging population, the incidence of chronic kidney disease has gradually increased, which will bring tremendous pressure to hemodialysis work. The demand pressure for dialysis equipment has also become increasingly greater. The dialysis adequacy index is a key parameter for evaluating the effectiveness of hemodialysis treatment. It directly affects the patient's health prognosis and the adjustment of dialysis treatment courses. At present, high-end imported dialysis equipment (such as Fresenius, B. Braun, etc.) can display the important indicator of dialysis adequacy (Kt / V) in real time. However, a large number of mid- and low-end equipment generally lack this function, resulting in many hospitals being unable to continuously monitor and adjust the patient's dialysis effect. In addition, due to the high technical threshold of existing dialysis equipment, the calculation method of dialysis adequacy is complex and protected by patents, which further limits the widespread application of this function.

[0003] There are two main methods for calculating dialysis adequacy. One involves collecting blood samples before and after dialysis and measuring urea nitrogen concentrations to calculate dialysis outcome indicators. However, this method only reflects the patient's physical condition at the time of blood draw and is relatively expensive. Another method for calculating dialysis adequacy indicators is to use real-time display equipment. These methods include ultraviolet absorbance measurement, conductivity measurement, and the effective ion dialyzability (EID) method. The ultraviolet absorbance measurement method enables real-time monitoring of urea clearance during dialysis. The device emits ultraviolet light of a specific wavelength. As it passes through the dialysis wastewater, urea molecules absorb some of the ultraviolet light, resulting in a decrease in the intensity of the ultraviolet light. This attenuation process is detected and recorded by a receiver within the device, reflecting the urea clearance rate during dialysis and enabling the calculation of the dialysis adequacy indicator. The effective ion dialyzability (EID) method assesses the adequacy of hemodialysis in real time by measuring the diffusion of sodium ions in the dialysate. Its principle is based on changes in dialysate conductivity during dialysis. When the dialysate's conductivity (primarily determined by sodium ions) briefly increases, sodium ions diffuse through the dialyzer membrane into the blood, causing the dialysate's conductivity at the dialyzer outlet to decrease. By analyzing the difference in conductivity between the dialyzer inlet and outlet, the conductivity clearance (Kecn) can be derived. This is comparable to the urea clearance rate and is used to calculate the dialysis dose. Conductivity testing involves installing conductivity sensors at the dialyzer inlet and outlet to measure changes in dialysate conductivity in real time, allowing for the automatic and noninvasive calculation of dialysate clearance. Dialysate conductivity is proportional to electrolyte concentration. Therefore, by monitoring the difference in dialysate conductivity before and after entering the dialyzer, the dialysis rate of electrolytes (such as NaCl) can be inferred. While ultraviolet absorbance measurement can monitor urea clearance in real time, it is expensive, requires a strict operating environment, and is significantly affected by interfering substances in the dialysate wastewater. The effective ion diffusion rate (EID) method assesses dialysis effectiveness by measuring sodium ion diffusion, but its results can be affected by conductivity fluctuations, and calculation accuracy decreases when electrolyte concentrations fluctuate dramatically. The conductivity measurement method is relatively complex, and the data requires further calculations to derive clearance and dialysis adequacy indicators.

[0004] With the rapid development of deep learning technology, combining data-driven deep learning methods with medical devices, and training models using large amounts of patient data, could enable simple and effective prediction of dialysis adequacy. This would help address the inability of many low- and mid-range devices to display this important function in real time. By effectively preprocessing the collected data, the dialysis adequacy indicator can be accurately predicted based solely on the conductivity of the dialysate before and after dialysis and the dialysate temperature after dialysis, providing new insights into data-driven approaches to predicting dialysis adequacy (Kt / V). Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a method for predicting hemodialysis adequacy based on transfer learning. It uses the conductivity and temperature of the dialysate after dialysis to perform deep learning modeling to predict hemodialysis adequacy; and uses transfer learning to effectively predict the dialysis adequacy of different patients.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A method for predicting hemodialysis adequacy based on transfer learning, comprising the following steps:

[0008] Step S1: Collect and process dialysis data:

[0009] Data from different patients' dialysis sessions were collected as basic input variables, including pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and dialysis adequacy value. A time increment based on dialysis time was additionally constructed as a modeling input variable. The time increment, pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature were combined as input variables and standardized.

[0010] Step S2: Model training:

[0011] The data processed in step S1 is trained using an improved multi-source domain TrAdaboost algorithm for prediction model training. The algorithm first trains multiple weak regressors, gradually improves the performance of each weak regressor through multiple rounds of iterations, and finally constructs a strong regression model by integrating multiple weak regressors as the final prediction model.

[0012] Step 3: Dialysis adequacy prediction:

[0013] The new patient dialysis data is input into the final prediction model trained in step 2. The model automatically learns the nonlinear dynamic mapping relationship between the input variables and the dialysis adequacy index to predict dialysis adequacy.

[0014] Furthermore, step S1 is specifically implemented as follows:

[0015] Step S11: Data collection:

[0016] The dialysis equipment collects data from dialysis patients during the dialysis process. Specifically, sensors are used to monitor and collect pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature in real time, and the dialysis adequacy value displayed by the equipment is manually recorded. The pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and dialysis adequacy value collected during the dialysis process serve as the basic data source.

[0017] Step S12: Data processing:

[0018] The collected conductivity and temperature data were preprocessed by averaging the conductivity and temperature by minute to ensure that the conductivity and temperature data were consistent with the sampling frequency of the dialysis adequacy value. At the same time, the time increment was introduced and the time increment, pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature were used as input variables of the prediction model, and the input variables were standardized.

[0019] The prediction model uses pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature and time increment as input variables, and dialysis adequacy index as label value. It uses an improved multi-source domain TrAdaboost algorithm to extract shared features from heterogeneous patient data to train a model for predicting the adequacy of hemodialysis.

[0020] Furthermore, step S12 is specifically:

[0021] The dialysis adequacy value y at each moment recorded manually i A time increment is calculated accordingly. The time increment is based on the accumulated time from the start of dialysis. Its initial value t0 is set to 0. The calculation formula of the time increment is as follows:

[0022] Δt i =t i -t0(1);

[0023] Among them, t i is the time corresponding to the dialysis adequacy value manually collected at the i-th time point. At each sampling moment, the time increment, pre-dialysis conductivity, post-dialysis conductivity and post-dialysis dialysate temperature are used as the input variables of the model X. i , specifically expressed as:

[0024] X i =[C pre_i ,C post_i ,T i ,Δt i ] (2);

[0025] Among them, C pre_i At time point t i Measured conductivity before dialysis, C post_i At time point t i Measured post-dialysis conductivity, T i At time point t i Measured temperature, Δt i is the time increment calculated from the start of dialysis;

[0026] All input variables are standardized. The standardization formula is as follows:

[0027]

[0028] where μ X =[μ C_pre ,μ C_post ,μ T ,μ Δt ] are four variables μ C_pre 、μ C_post 、μ T 、μ Δt The mean vector, σ X =[σ C_pre ,σ C_post ,σ T ,σ Δt ] are four variables σ C_pre , σ C_post , σ T , σ Δt The standard deviation vector of C_pre and σ C_pre is the mean and standard deviation of conductivity before dialysis, μ C_post and σ C_post is the mean and standard deviation of conductivity after dialysis, μ T and σ T is the mean and standard deviation of the dialysate temperature after dialysis, μ Δt and σ Δt is the mean and standard deviation of the time increment;

[0029] After standardization, the data of each dialysis patient is expressed as Where n is the number of samples of all patients, X′ i The input variables after standardization for each patient, y i Manually annotated dialysis adequacy values ​​were collected for each patient.

[0030] Furthermore, step S2 is implemented as follows:

[0031] Step S21: Initialize the weights of the source domain dataset:

[0032] Data from multiple dialysis patients on the same device are collected, and the training data is used as the source domain dataset, and the test set is used as the target domain dataset; the source domain dataset is defined as D1, D2, ..., D S , where D1, D2, ..., D S Represent different source domain datasets respectively;

[0033] A patient’s dialysis data is used as the target domain D T ; For each source domain dataset D S , assign initial weights, the initial weights of samples in the source domain dataset are uniformly distributed:

[0034]

[0035] in represents the initial weight of the j-th source domain data when the number of iterations p = 0;

[0036] Step S22: Start iterative training:

[0037] Use samples from all source datasets for training and use them to update the parameters of the weak regressor; train the weak regressor and use h k A gated recurrent unit (GRU) is trained on each source dataset;

[0038] During training, the mean square error (MSE) is used as the loss function, and the maximum mean difference (MMD) loss is additionally introduced to calculate the difference between the source domain data and the target domain data. This loss is then added to the total loss function. By minimizing the combined loss, the model gradually learns the migration rules from the source dataset to the target dataset. The specific formula is expressed as follows:

[0039]

[0040] Where φ(·) is the kernel function that maps the data to the high-dimensional reproducing kernel Hilbert space RKHS, H represents the norm of the difference in the reproducing kernel Hilbert space, n is the number of samples in the target domain dataset, and m is the number of samples in the source domain dataset;

[0041] Step S23: Update the source domain data weight:

[0042] Adjust the weight parameter α of the source domain data p It is updated based on the mean square error and maximum mean difference (MMD) on the source domain data. The weight update formula is as follows:

[0043]

[0044] Among them S is the mean square error on the source domain dataset, reflecting the prediction ability of the current weak regressor on the source domain data; λ is the adjustment coefficient, which is used to balance the influence of error and MMD. MMD is the maximum mean difference, reflecting the difference in data distribution between the source domain and the target domain;

[0045] After each round of iteration, the sample weights are adjusted according to the prediction error of the weak regressor; for the samples in the source dataset The weight update formula for source domain data is:

[0046]

[0047] in is the weight of the source domain sample after p+1 rounds of iteration, is the weight of the source domain sample in p rounds of iteration;

[0048] Step S24: Weight normalization:

[0049] After each round of iteration, the weight of the source data is normalized. The specific formula is:

[0050]

[0051] Step S25: Output the final regression model:

[0052] After P rounds of iteration, the final model is obtained, and the k weak regressors are weighted averagely to obtain the final strong regression model H(x).

[0053] Furthermore, in step 3, when a new patient undergoes dialysis, the sensor collects conductivity and temperature data during the dialysis process in real time. Consistent with the processing method in step 1, the collected data will introduce time increment as an additional input variable; then the pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature and time increment are standardized; the processed data are input into the final regression model trained in step 2 to predict the dialysis adequacy index of the new patient.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] Although the existing pre- and post-dialysis blood sample detection method can directly and accurately reflect the changes in urea nitrogen concentration before and after dialysis, this method can only reflect the patient's physical condition at the time of blood sampling and cannot monitor dynamic changes during dialysis in real time. In addition, frequent blood sampling is invasive to patients and is therefore not suitable for frequent use in routine dialysis. Although the ultraviolet absorbance measurement method can monitor the clearance of urea in real time, it has high requirements for the operating environment, and the equipment cost is expensive and the maintenance cost is high. The effective ion dialyzability (EID) method evaluates the dialysis effect by the diffusion of sodium ions. However, this method depends on the sodium ion concentration in the dialysate, a parameter that does not always directly reflect the urea clearance rate, so its accuracy may be affected. Although the conductivity detection method can measure the conductivity changes of the dialysate in real time in a non-invasive manner and calculate the dialysis clearance rate by installing conductivity sensors at the inlet and outlet of the dialyzer, its clearance calculation process is complex and has limited accuracy.

[0056] The present invention introduces a prediction model based on a data-driven approach (such as multi-source transfer learning). Instead of directly calculating Kt / V based on the traditional substitution derivation formula, the present invention predicts dialysis adequacy (Kt / V) using several easily accessible parameters, such as pre- and post-dialysis conductivity and post-dialysis dialysate temperature. By introducing a multi-source transfer learning algorithm, the present invention allows the model to automatically learn the nonlinear dynamic mapping relationship between input variables and the dialysis adequacy indicator Kt / V, thereby extracting common features from data from different patients and training the model.

[0057] Furthermore, the Maximum Mean Difference (MMD) loss function is used during training to effectively reduce the differences in feature distributions between the source and target domains, further improving the robustness and generalization of predictions across different patient conditions. Essentially, this approach overcomes individual differences among dialysis patients by mining deep relationships between data.

[0058] Compared to traditional methods, this method can effectively predict dialysis adequacy (Kt / V) using conductivity and temperature. Through large-scale data training, this data-driven approach effectively captures the complex relationship between conductivity, temperature, and dialysis adequacy. The model captures the changing trends of the urea clearance process, thereby improving prediction accuracy. This method also reduces the need for expensive and high-maintenance equipment, potentially offering broader applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0062] like Figure 1 As shown, this embodiment provides a method for predicting hemodialysis adequacy based on transfer learning, comprising the following steps:

[0063] Step S1: Collect and process dialysis data:

[0064] Data from different patients' dialysis processes were collected as basic input variables, including pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and dialysis adequacy value. Furthermore, a time increment generated based on dialysis time was constructed as a modeling input variable. The time increment, pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature were combined as input variables and standardized.

[0065] Step S11: Data collection:

[0066] The data of dialysis patients during the dialysis process are collected through dialysis equipment. Specifically, sensors are used to monitor and collect pre-dialysis conductivity, post-dialysis conductivity and post-dialysis dialysate temperature in real time, and the dialysis adequacy (Kt / V) value displayed by the equipment is manually recorded.

[0067] For each patient, the conductivity, temperature, and dialysis adequacy values ​​collected during dialysis were used as the basic data source.

[0068] Step S12: Data processing:

[0069] The collected conductivity and temperature data were preprocessed. Since the conductivity and temperature data collected during dialysis are collected by sensors at a high frequency of 5 seconds per time, while the dialysis adequacy value is usually recorded manually over a longer period of time (minutes), the collection intervals are long and vary (in minutes). To ensure the rationality and consistency of the collected data, the conductivity and temperature data were averaged by minute to ensure that the sampling frequency of the conductivity and temperature data was consistent with the dialysis adequacy value (Kt / V), so that these input variables can effectively correspond to the label values.

[0070] Because dialysis adequacy (Kt / V) is a cumulative variable and is manually recorded at varying intervals, directly using pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature as input variables often yields poor results. Therefore, the present invention introduces time increments, combining them with pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature as input variables for a prediction model, and standardizes these input variables. The model herein uses pre-dialysis conductivity, post-dialysis dialysate temperature, and time increments as input variables, with the dialysis adequacy index as the label value, and uses an improved multi-source domain TrAdaboost algorithm to train a model for predicting hemodialysis adequacy (Kt / V).

[0071] Specifically, step S12:

[0072] The dialysis adequacy (Kt / V) value y recorded manually at each moment i(in minutes) corresponds to a time increment. The time increment is the cumulative time based on the start of dialysis. Its initial value t0 is set to 0. The calculation formula of the time increment is as follows:

[0073] Δt i =t i -t0(1);

[0074] Among them, t i is the time corresponding to the dialysis adequacy (Kt / V) value manually collected at the i-th time point.

[0075] At each sampling moment, the time increment, pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature are used as the input variables of the model. i , specifically expressed as:

[0076] X i =[C pre_i ,C post_i ,T i ,Δt i ] (2);

[0077] Among them, C pre_i At time point t i Measured conductivity before dialysis, C post_i At time point t i Measured post-dialysis conductivity, T i At time point t i Measured temperature, Δt i The time increment is calculated from the start of dialysis. This input feature not only includes physical data (conductivity, temperature), but also reflects the change of sampling time through the time increment, solving the problem of inconsistent time intervals.

[0078] At each sampling moment, the manually labeled dialysis adequacy (Kt / V) value corresponds to the input feature X i By incorporating time increments into the input variables, the model is able to more accurately capture the temporal dependencies among variables during the dialysis process, especially when the sampling intervals are irregular.

[0079] In order to further optimize the input data, all input variables are standardized. The standardization formula is as follows:

[0080]

[0081] where μ X =[μ C_pre ,μ C_post ,μ T ,μ Δt ] are four variables μ C_pre 、μC_post 、μ T 、μ Δt The mean vector, σ X =[σ C_pre ,σ C_post ,σ T ,σ Δt ] are four variables σ C_pre , σ C_post , σ T , σ Δt The standard deviation vector of C_pre and σ C_pre is the mean and standard deviation of conductivity before dialysis, μ C_post and σ C_post is the mean and standard deviation of conductivity after dialysis, μ T and σ T is the mean and standard deviation of the dialysate temperature after dialysis, μ Δt and σ Δt are the mean and standard deviation of the time increments.

[0082] After standardization, the data of each dialysis patient is expressed as Where n is the number of samples of all patients, X i ′ is the input variable after standardization for each patient, y i Manually annotated dialysis adequacy values ​​were collected for each patient.

[0083] In summary, by introducing time increments as input features and normalizing the collected dialysis data for each patient, a deep learning model can be effectively applied to accurately predict dialysis adequacy (Kt / V). This method not only addresses the issue of inconsistent data collection intervals, but also enhances the model's ability to learn time dependencies by including cumulative time, thereby improving the model's training and prediction performance. It also addresses the issue of predicting dialysis adequacy (Kt / V) using cumulative values ​​of conductivity and temperature that lack temporal sequence information.

[0084] Step S2: Model training:

[0085] The data processed in step S1 is trained using an improved multi-source domain TrAdaboost algorithm (i.e., Multi Source TrAdaboost algorithm) to train a prediction model. The algorithm first trains multiple weak regressors, and after multiple rounds of iterations, gradually improves the performance of each weak regressor, and finally constructs a strong regression model by integrating multiple weak regressors as the final prediction model.

[0086] Multi Source TrAdaboost is a transfer learning algorithm in machine learning. In Chinese, it can usually be translated as "Multi-source domain TrAdaboost algorithm. TrAdaboost is the abbreviation of TransferAdaBoost, which means an AdaBoost algorithm based on transfer learning. It can be directly transliterated as "TrAdaboost" in Chinese.

[0087] Although the input variables do not directly include static individual attributes such as weight and residual renal function, the introduction of a multi-source transfer learning algorithm enables the extraction of common features from data from different patients. Furthermore, the maximum mean difference (MMD) loss function is used during training to effectively reduce the differences in feature distributions between the source and target domains, further improving the robustness and generalization of predictions across different patient conditions. Essentially, this overcomes the individual differences between dialysis patients by exploring deep relationships between the data.

[0088] Specifically, step S2 is implemented as follows:

[0089] Step S21: Initialize the weights of the source domain dataset:

[0090] The present invention collects data from multiple dialysis patients on the same device, uses the training data as the source domain dataset, and the test set as the target domain dataset. The source domain dataset is defined as D1, D2, ..., D S , where D1, D2, ..., D S They represent different source domain datasets respectively.

[0091] A patient’s dialysis data is used as the target domain D T ; For each source domain dataset D S , assign initial weights, the initial weights of samples in the source domain dataset are uniformly distributed:

[0092]

[0093] in It represents the initial weight of the j-th source domain data when the number of iterations p = 0.

[0094] Step S22: Start iterative training:

[0095] Use samples from all source datasets for training and use them to update the parameters of the weak regressor; train the weak regressor and use h k A gated recurrent unit GRU, k is 3, and each source data set is trained; GRU is a recurrent neural network, where k is 3, and each source data set is trained.

[0096] During training, the mean square error (MSE) is used as the loss function, and the maximum mean difference (MMD) loss is additionally introduced to calculate the difference between the source domain data and the target domain data. This loss is then added to the total loss function. By minimizing the combined loss, the model gradually learns the migration rules from the source dataset to the target dataset. The specific formula is expressed as follows:

[0097]

[0098] where φ(·) is the kernel function that maps the data to the high-dimensional reproducing kernel Hilbert space RKHS, H represents the norm of the difference in the reproducing kernel Hilbert space, n is the number of samples in the target domain dataset, and m is the number of samples in the source domain dataset.

[0099] Step S23: Update the source domain data weight:

[0100] Adjust the weight parameter α of the source domain data p It is updated based on the mean square error and maximum mean difference (MMD) on the source domain data. The weight update formula is as follows:

[0101]

[0102] Among them S is the mean square error on the source domain dataset, reflecting the prediction ability of the current weak regressor on the source domain data; λ is the adjustment coefficient, which is used to balance the influence of error and MMD. MMD is the maximum mean difference, reflecting the difference in data distribution between the source domain and the target domain.

[0103] After each round of iteration, the sample weights are adjusted according to the prediction error of the weak regressor; for the samples in the source dataset If the prediction error of the source data is large, the weight of the source data will be reduced to avoid the negative impact of the source data on the model; the update formula is:

[0104]

[0105] in is the weight of the source domain sample after p+1 rounds of iteration, is the weight of the source domain sample in p rounds of iteration; if Smaller, then As it approaches 1, the weight update amplitude is smaller, so the weight remains relatively large. A negative exponent means that the larger the prediction error of the source domain data, the smaller its weight is, thereby reducing its impact on subsequent model training.

[0106] Step S24: Weight normalization:

[0107] After each round of iteration, in order to prevent the sample weight from increasing or decreasing infinitely, the weight of the source data must be normalized. The specific formula is:

[0108]

[0109] Formula (8) is the weight normalization process for the source domain. Weight normalization ensures that the weight sum of the samples remains 1 in each iteration, avoiding excessive weights affecting the training process.

[0110] Step S25: Output the final regression model:

[0111] After P rounds of iteration, the final model is obtained, and the k weak regressors are weighted averagely to obtain the final strong regression model H(x).

[0112] This data-driven modeling method, based on deep learning, does not rely on traditional direct calculations of Kt / V by substitution. Key variables from the dialysis process (such as pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature) are collected and input into the model for training. The model automatically learns the nonlinear dynamic mapping relationship between these input variables and the dialysis adequacy indicator, Kt / V. This model captures the changing trends of the dynamic urea clearance process, enabling accurate predictions.

[0113] Step 3: Dialysis adequacy prediction:

[0114] The new patient data is input into the final prediction model trained in step 2 to predict dialysis adequacy.

[0115] When a new patient undergoes dialysis, the sensor collects conductivity and temperature data during the dialysis process in real time. Similar to the processing method in step 1, the collected data will introduce time increment as an additional input variable. The pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and time increment are then standardized. The processed data are input into the final regression model trained in step 2 to predict the dialysis adequacy index of the new patient.

[0116] In this example, dialysis data from six patients was collected using a dialysis device. Table 1 shows the collected data variables. To ensure the effectiveness of model training, this example preprocessed each patient's data, specifically introducing a key variable, time increment, to reflect dynamic changes in the time series. The preprocessed data is shown in Table 2.

[0117] In order to comprehensively evaluate the generalization ability of the model, this embodiment adopts a 6-fold cross-validation method. Specifically, in each experiment, the data of one patient is selected as the test set (Test), and the data of the remaining 5 patients are used as the training set for model training. In this way, the performance and generalization ability of the model on different patients can be fully evaluated. It is worth noting that all variables in Tables 1 and 2 are in minutes, among which the manually collected dialysis adequacy index (Kt / V) is based on data sampled at different time intervals, while the data collection frequency of conductivity and temperature is once every 5 seconds. Therefore, during data processing, the average values ​​of conductivity and temperature per minute are calculated to match the corresponding dialysis adequacy index.

[0118] To prevent information leakage during model training, this example does not use labeled test set patient data when using the Maximum Mean Difference (MMD) loss. Furthermore, because the dialysis equipment does not immediately display the dialysis adequacy (Kt / V) indicator at the beginning of dialysis, this example uses the patient's conductivity and temperature data from the first 15 minutes as the basis for MMD loss calculations and performs target domain migration. This approach prevents information leakage during the migration between the source and target domains.

[0119] Table 1 Manually collected data

[0120]

[0121] Table 2 Preprocessed data

[0122]

[0123] Among them C pre_i represents the conductivity of the dialysate before dialysis at the i-th time point, C post_i T represents the conductivity of the dialysate after dialysis at the i-th time point. i Δt represents the temperature of the dialysate after dialysis at the i-th time point. i represents the cumulative time increment between the i-th time point and the start of dialysis. i represents the manually recorded dialysis adequacy value (Kt / V) at the i-th time point.

[0124] During model training, this example uses time increment, conductivity, and temperature as input variables and dialysis adequacy (Kt / V) as a label, applying an improved MultiSource TrAdaboost model. This model successfully addresses the imbalance between multi-source data by dynamically adjusting the weights of source and target domain data, effectively improving prediction performance in the target domain.

[0125] To further mitigate the discrepancy between the source and target domain data distributions, the model introduces the Maximum Mean Difference (MMD) loss. MMD loss helps the model better adapt to diverse data distributions, thereby improving prediction accuracy. In each iteration, the MultiSource TrAdaboost model optimizes the contribution of source domain data by adjusting sample weights, gradually reducing prediction error. As iterations proceed, multiple weak models are gradually integrated to form a strong prediction model. This weighted integration strategy enables the model to better adapt to target domain data, significantly improving the accuracy of dialysis adequacy prediction.

[0126] Performance evaluation and comparison:

[0127] The prediction of dialysis adequacy is a regression task. Therefore, in order to evaluate the performance of the present invention in the prediction of dialysis adequacy (Kt / V), this embodiment selected a variety of regression evaluation indicators, including root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ). At the same time, the consistency evaluation Bland-Altman method is used as an additional evaluation index to calculate the prediction value and the true value y i These indicators comprehensively evaluate the prediction accuracy, stability, and consistency of the model from multiple perspectives. Specifically, the root mean square error (RMSE) is used to measure the average deviation between the predicted value and the actual value. The specific formula is:

[0128]

[0129] The mean absolute error is the average of the absolute errors between the predicted value and the actual value. The specific formula is:

[0130]

[0131] Coefficient of determination (R 2 ) represents the model's ability to explain actual data changes. The specific formula is:

[0132]

[0133] In clinical medicine, especially in the research of medical devices, the Bland-Altman consistency evaluation is widely used because of its characteristics of simple understanding, intuitive analysis, and reliable results. In order to ensure the accuracy of the results, it is also necessary to check whether the randomness, homoscedasticity, and normality of the comparison data differences are met before using this calculation indicator. The Bland-Altman inconsistency interval ratio is mainly used. By calculating the mean difference (Mean Bias) and the limits of agreement (Limits of Agreement), the LowerLimit is the lower limit of consistency, calculated as the mean deviation minus 1.96 times the standard deviation, and the Upper Limit is the upper limit of consistency, calculated as the mean deviation plus 1.96 times the standard deviation. The inconsistency interval ratio refers to the proportion of samples that fall outside the consistency limit, reflecting the degree of inconsistency between the predicted value and the actual value. A lower inconsistency interval ratio indicates that there is good consistency between the model prediction results and the actual value, and is an important indicator for evaluating the reliability of the model.

[0134] This example analyzed data from six patients and evaluated the model's predictive performance using six-fold cross-validation, with each fold representing a single patient's data set as the test set. In each experiment, data from one patient was selected as the test set, and data from the remaining five patients served as the training set, for a total of six cross-validations. Table 3 shows the specific evaluation results.

[0135] The fluctuation of evaluation indicators between folds is within a certain range, indicating that the performance of the model under different test sets is relatively stable. In particular, the coefficient of determination (R 2 ) was 0.8023, the lowest of all compromises, but still relatively high, indicating that even under the most suboptimal conditions, the model maintains strong predictive capabilities and reliable results. This further demonstrates that effective data preprocessing enables effective data-driven modeling under diverse conditions. Overall, the model's predictive performance is stable and reliable, providing strong support for predicting dialysis adequacy.

[0136] Table 3 Prediction results of transfer learning model

[0137]

[0138] Before the Bland-Altman analysis, prerequisites must be verified. For each fold of dialysis adequacy predictions and actual values, analysis of difference randomness, homoscedasticity, and normality was performed. The assumptions of homoscedasticity and normality were fully met, but the assumption of difference randomness was largely not met, possibly due to deficiencies in the model or incomplete capture of data characteristics. The Bland-Altman consistency evaluation results are shown in Table 4. For the six-fold experimental results, the calculated proportions of inconsistency intervals fluctuated around 5%. The low inconsistency ratios indicate that the differences between the predicted and actual values ​​mostly fell within the limits of agreement, demonstrating the effectiveness of this method.

[0139] Table 4 Bland-Altman consistency analysis results

[0140]

[0141] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should fall within the scope of protection of the present invention.

Claims

1. A hemodialysis adequacy prediction method based on transfer learning, characterized in that: The following steps are involved: Step S1: Collecting and processing dialysis data: Data from different patients' dialysis sessions were collected as basic input variables, including pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and dialysis adequacy value. A time increment based on dialysis time was additionally constructed as a modeling input variable. The time increment, pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature were combined as input variables and standardized. Step S2: Model training: The data processed in step S1 is trained using an improved multi-source domain TrAdaboost algorithm for prediction model training. The algorithm first trains multiple weak regressors, gradually improves the performance of each weak regressor through multiple rounds of iterations, and finally constructs a strong regression model by integrating multiple weak regressors as the final prediction model. Step 3: Dialysis adequacy prediction: The new patient dialysis data is input into the final prediction model trained in step 2. The model automatically learns the nonlinear dynamic mapping relationship between the input variables and the dialysis adequacy index to predict dialysis adequacy.

2. A hemodialysis adequacy prediction method based on transfer learning according to claim 1, characterized in that: Step S1 is specifically implemented as follows: Step S11: Data collection: The dialysis equipment collects data from dialysis patients during the dialysis process. Specifically, sensors are used to monitor and collect pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature in real time, and the dialysis adequacy value displayed by the equipment is manually recorded. The pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and dialysis adequacy value collected during the dialysis process serve as the basic data source. Step S12: Data processing: The collected conductivity and temperature data were preprocessed by averaging them by minute to ensure that they were consistent with the sampling frequency of the dialysis adequacy value. Furthermore, a time increment was introduced and used as input variables for the prediction model along with pre-dialysis conductivity, post-dialysis conductivity, and post-dialysis dialysate temperature. These input variables were then standardized. The prediction model uses pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature and time increment as input variables, and dialysis adequacy index as label value. It uses an improved multi-source domain TrAdaboost algorithm to extract shared features from heterogeneous patient data to train a model for predicting the adequacy of hemodialysis.

3. The method for predicting hemodialysis adequacy based on transfer learning according to claim 2, characterized in that: Specifically, step S12: The dialysis adequacy value y recorded manually at each moment i A time increment is calculated accordingly. The time increment is based on the accumulated time from the start of dialysis. Its initial value t0 is set to 0. The calculation formula of the time increment is as follows: Δt i =t i -t0(1); Among them, t i is the time corresponding to the dialysis adequacy value manually collected at the i-th time point. At each sampling moment, the time increment, pre-dialysis conductivity, post-dialysis conductivity and post-dialysis dialysate temperature are used as the input variables of the model X. i , specifically expressed as: X i =[C pre_i ,C post_i ,T i ,Δt i ] (2); Among them, C pre_i At time point t i Measured conductivity before dialysis, C post_i At time point t i Measured post-dialysis conductivity, T i At time point t i Measured post-dialysis dialysate temperature, Δt i is the time increment calculated from the start of dialysis; All input variables are standardized. The standardization formula is as follows: where μ X =[μ C_pre ,μ C_post ,μ T ,μ Δt ] are four variables μ C_pre 、μ C_post 、μ T 、μ Δt The mean vector, σ X =[σ C_pre ,σ C_post ,σ T ,σ Δt ] are four variables σ C_pre , σ C_post , σ T , σ Δt The standard deviation vector of C_pre and σ C_pre is the mean and standard deviation of conductivity before dialysis, μ C_post and σ C_post is the mean and standard deviation of conductivity after dialysis, μ T and σ T is the mean and standard deviation of the dialysate temperature after dialysis, μ Δt and σ Δt is the mean and standard deviation of the time increment; After standardization, the data of each dialysis patient is expressed as Where n is the number of samples of all patients, X i ′ is the input variable after standardization for each patient, y i Manually annotated dialysis adequacy values ​​were collected for each patient.

4. The method for predicting hemodialysis adequacy based on transfer learning according to claim 2, wherein: Step S2 is achieved by: Step S21: Initialize the weight of the source domain dataset: Data from multiple dialysis patients on the same device are collected, and the training data is used as the source domain dataset, and the test set is used as the target domain dataset; the source domain dataset is defined as D1, D2, ..., D S , where D1, D2, ..., D S Represent different source domain datasets respectively; A patient’s dialysis data is used as the target domain D T ; For each source domain dataset D S , assign initial weights, the initial weights of samples in the source domain dataset are uniformly distributed: in represents the initial weight of the j-th source domain data when the number of iterations p = 0; Step S22: Start iterative training: Use samples from all source datasets for training and to update the parameters of the weak regressor; Train the weak regressor using h k A gated recurrent unit (GRU) is trained on each source dataset; During training, the mean square error (MSE) is used as the loss function, and the maximum mean difference (MMD) loss is additionally introduced to calculate the difference between the source domain data and the target domain data. This loss is then added to the total loss function. By minimizing the combined loss, the model gradually learns the migration rules from the source dataset to the target dataset. The specific formula is expressed as follows: Where φ(·) is the kernel function that maps the data to the high-dimensional reproducing kernel Hilbert space RKHS, H represents the norm of the difference in the reproducing kernel Hilbert space, n is the number of samples in the target domain dataset, and m is the number of samples in the source domain dataset; Step S23: Update the source domain data weight: Adjust the weight parameter α of the source domain data p It is updated based on the mean square error and maximum mean difference (MMD) on the source domain data. The weight update formula is as follows: in is the mean square error on the source domain dataset, reflecting the prediction ability of the current weak regressor on the source domain data; λ is the adjustment coefficient, which is used to balance the influence of error and MMD. MMD is the maximum mean difference, reflecting the difference in data distribution between the source domain and the target domain; After each round of iteration, the sample weights are adjusted according to the prediction error of the weak regressor; for the samples in the source dataset The weight update formula for source domain data is: in is the weight of the source domain sample after p+1 rounds of iteration, is the weight of the source domain sample in p rounds of iteration; Step S24: Weight normalization: After each round of iteration, the weight of the source data is normalized. The specific formula is: Step S25: Output the final regression model: After P rounds of iteration, the final model is obtained, and the k weak regressors are weighted averagely to obtain the final strong regression model H(x).

5. The method for predicting hemodialysis adequacy based on transfer learning according to claim 1, characterized in that: In step 3, when a new patient undergoes dialysis, the sensor collects real-time conductivity and temperature data during the dialysis process. Similar to the processing method in step 1, the collected data will introduce time increment as an additional input variable. The pre-dialysis conductivity, post-dialysis conductivity, post-dialysis dialysate temperature, and time increment are then standardized. The processed data are input into the final prediction model trained in step 2 to predict the dialysis adequacy index of the new patient.

Images