Hemodialysis filtration membrane blockage early warning method based on machine learning

Through machine learning technology and real-time data acquisition system, the lack of accuracy and real-time early warning system for the hemodialysis filtration membrane blockage is solved, and the early warning of filtration membrane blockage and the safety of the dialysis process is improved.

CN118983054AInactive Publication Date: 2024-11-19THE FIRST AFFILIATED HOSPITAL OF HENAN UNIV OF TCM

Patent Information

Application Number
CN202411054544.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing hemodialysis filtration membrane blockage warning system has shortcomings in accuracy, sensitivity, data management, algorithm selection and real-time performance, resulting in the inability to timely identify early signals of filtration membrane blockage, affecting the treatment effect and patient safety.

Method used

Using machine learning technology, through data cleaning, feature extraction and model training, a real-time data acquisition system is built, using algorithms such as random forests, Bayesian optimization and deep learning to predict the risk of filtered membrane blockage in real time, and set risk thresholds to trigger early warnings .

Benefits of technology

It improves the accuracy and sensitivity of the risk of filtration membrane blockage, realizes early warning, reduces emergency situations due to blockage, and improves the safety and efficiency of the dialysis process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118983054A_ABST
    Figure CN118983054A_ABST
Patent Text Reader

Abstract

The invention relates to a hemodiafiltration membrane blockage early warning method based on machine learning. The method comprises the following steps: firstly, collecting data in a hemodialysis process, and carrying out data cleaning and standardization processing on the collected data; noise in the data is removed by adopting a self-adaptive noise recognition algorithm, and abnormal values in the data are recognized and processed by utilizing a self-encoder anomaly detection model based on deep learning; adopting an interpolation method based on a prediction model, and utilizing a time sequence prediction model or a deep learning network to predict and fill missing data values; a standardization technology for dynamically adjusting the data range is implemented, and standardization parameters are dynamically adjusted according to the characteristics of the real-time data flow; performing feature extraction and selection on the processed data; selecting a proper machine learning model, training the model through the training set, and evaluating the performance of the model by adopting a cross validation technology; and constructing a real-time data acquisition system to obtain data of various sensors in the dialysis process in real time, and updating input data of the machine learning model in real time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a hemodiafiltration membrane blockage early warning method, and in particular to a hemodiafiltration membrane blockage early warning method based on machine learning. Background Art

[0002] In recent years, the research and application of hemodiafiltration membrane blockage early warning methods have made significant progress. However, these methods still face many challenges and problems in practical applications, mainly including the following aspects: First, the existing hemodiafiltration membrane blockage early warning system still has limitations in accuracy and sensitivity. Due to various factors in the dialysis process, including individual differences in patients, changes in the physical and chemical properties of blood and dialysate, and different states of dialysis equipment, these factors will affect the performance of the filtration membrane. However, many early warning systems fail to fully consider these complexities, resulting in their unsatisfactory performance in accurately identifying and predicting the risk of filtration membrane blockage. Especially for the specific conditions of some patients, the existing models cannot identify the early signs of blockage in time, thus affecting the treatment effect and patient safety. Secondly, the problem of data collection and management is also one of the major challenges faced by the hemodialysis early warning system. The amount of data generated during the dialysis process is huge and complex, including multi-dimensional data of flow rate, pressure, and chemical component concentration. How to effectively collect, store, process and analyze this data to obtain high-quality input is a complex engineering problem. At present, many systems lack sufficient capabilities in data processing and management, and the problems of data loss, noise and lack of real-time performance are still prominent. In addition, data privacy and security issues also need to be addressed urgently. How to achieve efficient use of data while protecting patient privacy is a common problem that all health monitoring systems need to face.

[0003] In terms of algorithms, the model selection and optimization of existing early warning systems also face challenges. Although machine learning technology is increasingly used in dialysis early warning systems, the selection of appropriate algorithms remains a challenge due to the complexity of the physiological and chemical reactions involved in the dialysis process. Many models rely too much on specific data distributions and assumptions, making it difficult to adapt to various uncertainties and dynamic changes in practical applications. In addition, the hyperparameter tuning process of the model is usually complex and time-consuming, which can easily lead to overfitting or underfitting, thus affecting the predictive performance of the model. Another problem is the inadequacy of existing early warning systems in terms of response time and real-time performance. Filter membrane blockage often requires rapid response and processing, but many systems have a lag in the process of identifying and predicting risks, resulting in the inability to take timely measures, increasing the health risks of patients. Especially when the equipment processing capacity is limited or the system integration is not perfect, it is more difficult to achieve real-time monitoring and rapid response. In addition, existing systems also have deficiencies in integration and interoperability. Hemodialysis equipment and early warning systems are usually produced by different manufacturers, and the standards and protocols between them are inconsistent, resulting in many obstacles in data interfaces, system integration and collaborative operations. This not only increases the complexity and maintenance costs of the system, but also affects the efficiency and reliability of the early warning system. Summary of the invention

[0004] The purpose of the present invention is to provide a hemodiafiltration membrane blockage early warning method based on machine learning, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0006] S1. First, collect the flow sensor data, pressure sensor data, chemical sensor data and time series data during the hemodialysis process, clean and standardize the collected data, remove noise and outliers, and fill in missing data;

[0007] The data is subjected to standardization process:

[0008] S1.1. First, an adaptive noise recognition algorithm is used to learn and remove noise from historical data, and an autoencoder anomaly detection model based on deep learning is used to identify and process outliers in the data.

[0009] S1.2, then adopt the interpolation method based on the prediction model, using the time series prediction model or deep learning network to predict and fill in the missing data values;

[0010] S1.3, finally, implement the standardization technology of dynamically adjusting the data range, and dynamically adjust the standardization parameters according to the characteristics of the real-time data stream;

[0011] S2. Then, feature extraction and selection are performed on the processed data, and time domain features including mean, peak value, trend and periodic changes are extracted from each sensor data, and frequency domain features are extracted by Fourier transform and changes in chemical component concentrations in blood and dialysate are analyzed to extract key change features;

[0012] S3. Then use the machine learning model random forest to divide the processed data into training set and test set, train the model with the training set and use cross-validation technology to evaluate the model performance, use Bayesian optimization to adjust parameters to find hyperparameters, and further improve the model performance through feature engineering and ensemble learning methods;

[0013] S4. Finally, a real-time data acquisition system is built to obtain data from various sensors during the dialysis process in real time, update the machine learning model input data in real time, use the trained model to predict the clogging risk of the filter membrane in real time, set the risk threshold, and trigger the early warning system when the predicted clogging risk exceeds the threshold.

[0014] Furthermore, the adaptive noise recognition algorithm construction process includes: firstly, using the adaptive noise recognition algorithm to remove noise in the data based on historical data learning, and using the adaptive noise recognition function:

[0015]

[0016] Where x represents the input data, which is time or other continuous measurements related to the dialysis process, θ and λ are parameters that are dynamically adjusted based on real-time data to identify and filter out noise related to the dialysis pump frequency, and t is the integration variable;

[0017] At the same time, the autoencoder anomaly detection model based on deep learning is used to identify and process outliers in the data, and the anomaly measurement function is applied:

[0018]

[0019] Where z is the reconstruction error obtained from the autoencoder, and σ is the standard deviation dynamically adjusted according to the distribution of abnormal data;

[0020] In addition, these processed data are integrated into the dialysis machine control system to provide real-time feedback to adjust dialysis parameters or remind medical staff, using a dynamic threshold adjustment strategy:

[0021] h(t,α,β,γ)=αtexp(-βt)+γcos(βt)

[0022] Where t represents the usage time of the filter membrane, α, β, and γ are parameters that are dynamically adjusted based on the dialysate quality and the condition of the filter membrane.

[0023] Furthermore, the construction process of the interpolation method based on the prediction model is as follows:

[0024] S1. Use a long short-term memory network LSTM or a gated recurrent unit GRU to train the model specifically to identify data patterns that cause filter blockages, and use a dynamic window adjustment function:

[0025]

[0026] Where t represents the time point in the time series, δ and ω are parameters that are dynamically adjusted based on the historical data of the membrane pressure and flow rate;

[0027] S2. Then, we combine the convolutional neural network (CNN) and the recurrent neural network (RNN) to process the multi-dimensional data in the dialysis process, and use the advanced multi-dimensional feature analysis function:

[0028]

[0029] Where p represents the concentration of the chemical component, q and r represent measures of flow rate and pressure, respectively;

[0030] S3. In addition, a continuous learning mechanism is introduced to adjust the formula through feedback:

[0031]

[0032] Where s represents each new data point in the dialysis time series, and α and β are model performance parameters that are dynamically adjusted based on the new data.

[0033] Furthermore, the standardization technology construction method is as follows:

[0034] First, we select key parameters that directly affect the performance of the filtration membrane, such as pressure difference, flow rate, and chemical composition of the dialysate, and use the sliding window technology:

[0035]

[0036] Real-time data acquisition is performed, where t represents the time point in the time series, δ and ω are dynamically adjusted parameters, and the size of the data window is dynamically adjusted according to the precursor of clogging of the filter membrane;

[0037] Next, the adaptive normalization formula is applied according to the real-time monitored changes in the filter membrane status:

[0038]

[0039] Where t represents the time point in the time series, μ(t) and σ(t) are the dynamic mean and standard deviation calculated based on the real-time data window, and λ is the smoothing parameter introduced to adjust the sensitivity in the standardization process;

[0040] Finally, implement a feedback adjustment mechanism:

[0041] h(α,β,x)=α·log(1+e βx )

[0042] Where x represents the specific variable value of real-time data input, and α and β are parameters adjusted based on the performance feedback of the early warning model.

[0043] Furthermore, the data is subjected to feature extraction and selection implementation process as follows:

[0044] S1. First, extract temporal features from each sensor data, including mean, variance, peak, trend and periodic change. Dynamically calculate these features through sliding window technology to capture the temporal variation pattern in the data. And detect the long-term trend and periodic pattern in the data through segmented regression and period analysis algorithm.

[0045] S2. Secondly, the Fourier transform is applied to extract the frequency domain features of the processed time domain data, and the multi-resolution Fourier transform is introduced to adapt to the non-stationary characteristics of the data and capture the dynamic changes of different frequency components;

[0046] S3. Finally, analyze the concentration changes of chemical components in blood and dialysate, extract key change features to reveal the impact of chemical composition fluctuations on filtration membrane performance;

[0047] Furthermore, the method for extracting time domain features from each sensor data includes:

[0048] First, extract time domain features from each sensor data, including mean, variance, peak, trend, and periodic changes, and dynamically calculate these features through sliding window technology, using a dynamic window adjustment function:

[0049]

[0050] Where t represents the time point in the time series, α and t0 are adjustment parameters, and the window size is dynamically adjusted according to the data change rate;

[0051] In each sliding window, by calculating the time domain features, the mean:

[0052]

[0053] And the variance:

[0054]

[0055] Where μ(t) represents the mean value of the data in the sliding window at the current time t, W(t) represents the current window size, which is determined by the dynamic window adjustment function, and x(τ) represents the data value at the time point τ;

[0056] Detect long-term trends and cyclical patterns in data using piecewise regression and periodic analysis algorithms, and apply trend analysis functions:

[0057] T(t)=β0+β1t+β2t 2

[0058] Where t represents the time point in the time series, β0, β1 and β2 are regression coefficients dynamically calculated through piecewise regression to identify trend changes in the data;

[0059] Using periodic analysis function:

[0060] P(f)=γsin(2πft+φ)

[0061] Where t represents the time point in the time series, f represents the frequency, and γ and φ are parameters that are dynamically adjusted based on Fourier analysis to identify periodic patterns in the data.

[0062] Furthermore, the frequency domain feature extraction process is as follows:

[0063] S1. Apply fast Fourier transform FFT to convert the preprocessed time domain data into frequency domain to extract frequency domain features, using the Fourier transform formula:

[0064]

[0065] Where t represents the time point in the time series, x(t) represents the time domain signal, f represents the frequency, and T is the time period;

[0066] S2. Secondly, the multi-resolution Fourier transform is introduced to adapt to the non-stationary characteristics of the data and capture the dynamic changes of different frequency components. The formula is:

[0067]

[0068] Where t represents the time point in the time series, x(t) is the time domain signal, ψ is the mother wavelet function, and a and b are the scale and translation parameters respectively;

[0069] S3. Extract multi-scale frequency domain features from multi-resolution Fourier transform, including high-frequency components, low-frequency components, and medium-frequency components, using the frequency domain feature extraction formula:

[0070]

[0071] Where W k (f, t) is the transformation coefficient at the kth scale, f is the frequency, and t is the time.

[0072] Furthermore, the process of analyzing the concentration changes of chemical components in blood and dialysate is as follows:

[0073] First, a machine learning model is applied to analyze the dynamic data of chemical composition, and key change features are extracted through a series of feature engineering methods, including the time-varying coefficient model through:

[0074]

[0075] Where t0, t represent the start and end time of the integration, covering the time window of the analysis, c(τ) represents the concentration of the chemical component at time point τ, μ c (τ) is the moving average of the chemical component concentration;

[0076] Then, using the multi-parameter dynamic simulation model:

[0077]

[0078] t represents time, represents a collection of different chemical components, c j (t) is the concentration of the jth chemical component at time t, b0, b1, …, b 2m are the coefficients in the prediction model, where b0 is the baseline response, b1,…,b m is the influence intensity of each chemical component, b j+m is the time decay rate of the corresponding chemical component effect, and m is the total number of chemical components, which indicates the number of different chemical components used in the model.

[0079] Furthermore, the machine learning model construction method includes: first, using a customized machine learning model random forest, using the split cross-validation technology of time series data:

[0080]

[0081] Where k is the number of folds, J i is the index set of the i-th fold, y j is the actual value, is the predicted value, which evaluates the performance of the model on unseen data;

[0082] Next, the Bayesian optimization method is used to optimize the hyperparameters of the model, through the optimization formula:

[0083]

[0084] Where Θ is the set of hyperparameters, L(Θ) is the loss function, and λ is the regularization coefficient;

[0085] Finally, advanced feature engineering and ensemble learning methods are applied, using the feature engineering formula:

[0086]

[0087] Where x is the original feature set, x n is the nth feature, α n is the feature weight.

[0088] Beneficial effects of the present invention: By introducing machine learning algorithms, including support vector machines, random forests, and gradient boosting trees, the present invention can accurately capture complex multi-dimensional data during dialysis and identify potential blockage risks. These algorithms can process a large number of nonlinear data relationships and learn individual differences and physiological changes of different patients, improving the accuracy and sensitivity of predictions, thereby achieving early warning;

[0089] Secondly, the integrated real-time data acquisition and analysis system can monitor various physical and chemical parameters of blood and dialysate in real time during dialysis. With the help of fast data processing and analysis capabilities, the system can identify abnormal changes and quickly issue early warning signals, helping medical staff to quickly take intervention measures and reduce emergencies caused by filtration membrane blockage. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 This is a flow chart of the hemodiafiltration membrane blockage early warning method based on machine learning of the present invention.

[0091] Figure 2 This is a module diagram for building the adaptive noise recognition algorithm of the present invention.

[0092] Figure 3 A module diagram for constructing an interpolation method based on a prediction model of the present invention.

[0093] Figure 4 Flowchart for feature extraction and selection of data for the present invention. DETAILED DESCRIPTION

[0094] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0095] Combined with Figure 1 , collect key data, including flow, pressure, chemical sensor data, and related time series data, which are collected from various parts of the dialysis machine. To ensure the quality and applicability of the data, data cleaning is first performed to remove noise and outliers caused by sensor errors and external interference. An adaptive noise recognition algorithm is used to analyze historical data to identify and eliminate atypical noise patterns; at the same time, a deep learning autoencoder model is used to detect and isolate abnormal data points that do not conform to normal operating patterns. For missing data, time series-based prediction models or deep learning networks including recurrent neural networks are used to predict and fill these gaps to ensure data integrity.

[0096] Furthermore, the data range normalization technology is dynamically adjusted for the processed data, and the normalization parameters are dynamically adjusted by real-time analysis of the characteristics of the data stream to adapt it to the actual distribution of the data. Subsequently, the standardized data is subjected to in-depth feature extraction, including extraction of mean, peak, trend and periodic changes of data from the time domain, and extraction of key features of chemical component concentration changes from the frequency domain through Fourier transform.

[0097] In the model development phase, the filtered feature data was divided into training sets and test sets, and trained using random forest machine learning models. The performance of these models on different data sets was evaluated through cross-validation techniques to ensure that the models had good generalization capabilities. In addition, Bayesian optimization techniques were used to adjust hyperparameters and refine the most suitable model parameter settings. With the help of feature engineering and ensemble learning methods, including model fusion and stacking, the model's ability to predict potential blockage risks during dialysis was further enhanced.

[0098] Ultimately, this early warning system continuously monitors the dialysis process through a real-time data acquisition system, updates the input data of the machine learning model in real time, and uses the trained and optimized model to predict the risk of clogging of the filter membrane in real time. The system sets a predetermined risk threshold, and once the predicted risk exceeds this threshold, an early warning is immediately triggered to remind medical staff to take action.

[0099] Embodiment 1:

[0100] Combined with Figure 2 In this embodiment, a machine learning-based hemodiafiltration membrane blockage early warning system is developed. During the dialysis process, multiple sensor data including flow, pressure, and chemical composition are collected continuously and recorded as time series data. An adaptive noise recognition algorithm is used to process this data:

[0101]

[0102] in:

[0103] x is a point in the time series, indicating the length of time from the start of dialysis to the present;

[0104] θ and λ are adjustment parameters, where θ represents the sensitivity of noise identification and is set in the range of 0.1 to 0.5 to adjust the filtering strength, and λ controls the attenuation rate and ranges from 0.01 to 0.1 to adapt to different noise environments;

[0105] t is the integration variable used to calculate the cumulative noise contribution from the start of dialysis to the current point.

[0106] Assume that there is periodic noise in the data related to the frequency of the dialysis pump. The noise period occurs once every minute and is related to the operating frequency of the pump. Set θ = 0.3 and λ = 0.05. For a specific data point x = 30 minutes (i.e. 30 minutes after the start of dialysis), apply the above function to calculate the result of noise filtering. Assume The value of is 1.5, then:

[0107] f(30,0.3,0.05)=0.3exp(-0.05×1.5)=0.3exp(-0.075)=0.3×0.9275≈0.278

[0108] This value indicates that at 30 minutes, the coefficient of the original signal is adjusted through the adaptive noise identification algorithm to reduce the impact of noise.

[0109] The embodiment further utilizes a deep learning based autoencoder model to identify and handle outliers in the data, using the following anomaly metric function:

[0110]

[0111] Where z is the reconstruction error obtained from the autoencoder, and σ is the standard deviation dynamically adjusted according to the distribution of abnormal data, with a value range of 0.1 to 1.0.

[0112] Assume that during the dialysis process, the autoencoder identifies a data point with a reconstruction error z=0.5, and the standard deviation σ of the current abnormal data is 0.3. According to the above abnormality metric function, calculate the abnormal score of the data point:

[0113]

[0114] g(0.5,0.3)=log(1+exp(-2.777))

[0115] g(0.5,0.3)≈log(1+0.062)

[0116] g(0.5,0.3)≈log(1.062)

[0117] g(0.5,0.3)≈0.060

[0118] A low value indicates that the data point is slightly off but not significantly abnormal enough to trigger an alert.

[0119] The final example implements real-time feedback adjustment and dynamic threshold adjustment strategies, integrating the processed data into the dialysis machine control system to ensure that all relevant sensor readings and machine learning model outputs can be updated in real time and used to control the dialysis process. Function definition:

[0120] h(t,α,β,γ)=αtexp(-βt)+γcos(βt)

[0121] t is the usage time of the filter membrane; α, β, and γ are parameters that are dynamically adjusted according to the dialysate quality and the real-time status of the filter membrane. α is between 0.1 and 0.5 and is used to adjust the influence of time weight. β is between 0.01 and 0.1 and controls the attenuation rate and the sensitivity of periodic adjustment. γ is between 0.1 and 1.0 and is used to adjust the amplitude of the periodic function.

[0122] Assume that the filter membrane has been used for 3 hours (t=3), and based on the dialysate quality and the filter membrane condition, select α=0.3, β=0.05, and γ=0.7. These parameters will be substituted into the dynamic threshold adjustment formula to calculate:

[0123] h(3,0.3,0.05,0.7)=0.3×3×exp(-0.05×3)+0.7×cos(0.05×3)

[0124] h(3,0.3,0.05,0.7)=0.9×exp(-0.15)+0.7×cos(0.15)

[0125] h(3,0.3,0.05,0.7)≈0.9×0.861+0.7×0.988

[0126] h(3,0.3,0.05,0.7)≈0.775+0.692

[0127] h(3,0.3,0.05,0.7)≈1.467

[0128] This calculation result shows that as the filter membrane usage time increases, the dialysis machine system should adjust the dialysis parameters or remind medical staff to ensure that the filter membrane will not affect the patient's dialysis treatment effect due to clogging.

[0129] Embodiment 2:

[0130] Combined with Figure 3 In the machine learning-based hemodialysis filtration membrane blockage early warning system, the long short-term memory network LSTM and the gated recurrent unit GRU are used to identify and learn the signs of filtration membrane blockage, including specific pressure and flow rate change patterns, and use a dynamic window adjustment function:

[0131]

[0132] Where t represents the time point in the time series, δ and ω are dynamically adjusted parameters that adjust the size and shape of the data window; δ controls the overall scaling of the window, ranging from 0.5 to 2.0, and is adjusted according to the volatility of the data and the sensitivity of the model; ω adjusts the frequency response of the window, ranging from 0.1 to 10.0, matching the periodicity of the data and the detection of abnormal behavior.

[0133] Assume that during dialysis, the filter membrane begins to show signs of clogging after 2 hours of use. At this time point t = 2 hours, set δ = 1.5 and ω = 5, and calculate the value of the window adjustment function:

[0134]

[0135] Assuming sin(10) ≈ -0.544 (derived from a standard sine table or calculator):

[0136]

[0137] f(2,1.5,5)=1.5×0.002959

[0138] f(2,1.5,5)≈0.004439

[0139] This result is used to dynamically adjust the input window of the LSTM or GRU model so that the model can more accurately analyze the data pattern near the current data point.

[0140] The embodiment further enhances the processing capability of the model by combining convolutional neural network (CNN) and recurrent neural network (RNN) to process and analyze multi-dimensional data during dialysis. CNN automatically extracts spatial features from the sensor data of the dialysis machine, especially identifies patterns and features from the data collected by chemical sensors, including the concentration distribution pattern of chemical components; RNN processes the features extracted by CNN and the data directly obtained from time series sensors (flow and pressure sensors), and learns and predicts the dynamic changes of data over time.

[0141] Using advanced multidimensional feature analysis functions:

[0142]

[0143] Where p represents the concentration of the chemical component, and q and r are measures of flow rate and pressure, respectively.

[0144] Assume that during the dialysis process, the concentration of the chemical component p = 0.8, the flow rate q = 1.2 m / s, and the pressure r = 2.0 atm. According to the advanced multidimensional feature analysis function, the calculation is as follows:

[0145]

[0146] g(0.8,1.2,2.0)≈0.453

[0147] This result can be used to assess the risk level of the current dialysis status and further dynamically adjust the operating parameters of the dialysis machine or trigger an early warning when the risk threshold is exceeded.

[0148] Finally, the embodiment introduces a continuous learning mechanism. During the dialysis process, each new data point s represents the latest state of the dialysis time series, including multi-dimensional information such as flow rate, pressure, and chemical component concentration. New data points are used to continuously update and train machine learning models (including LSTM, GRU, and CNN) to maintain the model's sensitivity and responsiveness to the latest dialysis status.

[0149] Apply the feedback adjustment formula:

[0150]

[0151] Where s is the newly collected data at the dialysis time point; α and β are dynamically adjusted performance parameters that affect the response speed and sensitivity of the model. Parameter value range: α ranges from 0.01 to 0.5, controlling the influence and adjustment speed of linear growth. β ranges from 0.001 to 0.1, determining the rate of exponential decay and the sensitivity of integral adjustment.

[0152] Set s = 2 hours, select α = 0.1 and β = 0.05 to adjust the model. The calculation feedback adjustment formula is as follows:

[0153]

[0154] h(2,0.1,0.05)=0.2×e -0.2 +0.05×Numerical integration calculation

[0155] Assume that the numerical integration is calculated

[0156] h(2,0.1,0.05)=0.2×0.819+0.05×0.882

[0157] h(2,0.1,0.05)≈0.164+0.0441

[0158] h(2,0.1,0.05)≈0.2081

[0159] This computational result provides guidance on how to adjust the model parameters to better respond to new dialysis states, thereby maintaining the optimal performance of the early warning system.

[0160] Embodiment 3:

[0161] Example In a machine learning-based hemodiafiltration membrane blockage early warning system, key parameters pressure difference, flow rate, and chemical composition of the dialysate are monitored, using a sliding window technique:

[0162]

[0163] t represents the time point in the time series;

[0164] δ and ω are parameters that are dynamically adjusted according to real-time data and the state of the filter membrane, and are used to control the size and sensitivity of the data window;

[0165] δ adjusts the overall sensitivity, ranging from 0.1 to 1.0, and is adjusted according to the stability and variability of the dialysis data;

[0166] ω determines the width of the window, ranging from 10 to 1000 in seconds, and is adjusted according to the performance of the filter membrane and the warning requirements.

[0167] Assume that the time point t = 60 minutes, select δ = 0.5 and ω = 500 to adjust the window for real-time data collection. The window adjustment factor is calculated as follows:

[0168]

[0169] f(60,0.5,500)≈0.5·0.878

[0170] f(60,0.5,500)≈0.439

[0171] At the 60-minute time point, the adjusted window size allowed the system to acquire data at approximately 44% of the original data rate.

[0172] The embodiment further adopts an adaptive normalization formula to ensure that the actual operating status of the filter membrane can be accurately reflected during the data analysis process, and constructs a normalization formula:

[0173]

[0174] in:

[0175] x is any key parameter data point collected in real time, including pressure, flow rate, or chemical component concentration;

[0176] μ(t) is a dynamic mean value calculated based on the real-time data window, reflecting the average performance of the filter membrane over time;

[0177] σ(t) is the dynamic standard deviation, which represents the variability of the data within the current time window;

[0178] λ is a smoothing parameter, which is used to prevent the denominator from approaching zero and improve the stability and accuracy of the calculation.

[0179] Parameter value range:

[0180] The values ​​of μ(t) and σ(t) are calculated in real time, depending on the data in the current window;

[0181] λ takes a small positive value, ranging from 0.01 to 0.1, to ensure numerical stability and appropriately increase the robustness of the calculation.

[0182] Assume that at a certain time t = 120 minutes during the dialysis process, the monitored flow rate x is 200 mL / min, and μ(120) = 180 mL / min, σ(120) = 20 mL / min are calculated through the real-time data window, and λ = 0.05 is selected. According to the adaptive normalization formula, the normalized value x of the flow rate is calculated as follows:

[0183] norm

[0184]

[0185] x≈0.9999375

[0186] norm

[0187] This result shows that after the deviation of the current flow rate from the dynamic mean is standardized, the difference is very small, indicating that the flow rate is within the normal range.

[0188] The final embodiment implements a feedback adjustment mechanism using the function:

[0189] h(α,β,x)=α·log(1+e βx )

[0190] in:

[0191] x is the value of a specific variable extracted from real-time data, including pressure readings, flow rates, or concentrations of chemical components;

[0192] α and β are dynamically adjusted parameters that control the amplification factor and exponential growth rate of the function respectively;

[0193] α: The overall sensitivity and output amplitude of the adjustment formula, which ranges from 0.01 to 1.0, depending on the influence of the variable and the sensitivity requirements of the early warning system;

[0194] β: affects the exponential growth rate, with a value range of 0.1 to 5.0, adjusted to accommodate different data change speeds and warning thresholds.

[0195] Assume that during the dialysis process, the chemical component concentration x value monitored in real time is 0.5. The current system setting determines that α=0.5 and β=2 are used for feedback adjustment calculation based on model performance, and the feedback adjustment value h is calculated as follows:

[0196] h(0.5,2,0.5)=0.5·log(1+e 2×0.5 )

[0197] h(0.5,2,0.5)=0.5·log(1+e 1 )

[0198] h(0.5,2,0.5)=0.5·log(1+2.71828)

[0199] h(0.5,2,0.5)=0.5·log(3.71828)

[0200] h(0.5,2,0.5)≈0.5·1.311

[0201] h(0.5,2,0.5)≈0.6555

[0202] This calculation result provides information on how to dynamically adjust the parameters of the early warning system based on real-time data to ensure that the system output closely corresponds to the actual situation of the current dialysis status.

[0203] Embodiment 4:

[0204] Embodiment In a hemodiafiltration membrane blockage early warning system based on machine learning, time domain features are first extracted from various sensor data, including mean, variance, peak, trend and periodic change, and a dynamic window adjustment function is used:

[0205]

[0206] in:

[0207] t is the time point in the time series;

[0208] α and t0 are adjustment parameters used to dynamically adjust the window size according to the rate of data change. α is adjusted between 0.01 and 0.1 to match the sensitivity of data changes; t0 is set to the time point, midpoint or specific time when the change is expected to be captured.

[0209] Then calculate the mean and variance within the sliding window:

[0210] Mean (μ(t)) calculation formula:

[0211]

[0212] Here W(t) is the width of the dynamic window, which is calculated according to the dynamic window adjustment function above.

[0213] Variance (σ 2 (t)) Calculation formula:

[0214]

[0215] Here μ(t) is the mean calculated within the current window W(t), and x(τ) is the data value at time point τ.

[0216] Assume that flow rate data is monitored during dialysis and that you wish to use the above techniques to analyze data changes. Set α = 0.05 and t0 = 30 minutes as the midpoint of the monitoring period. Considering the calculation at t = 30 minutes, assume that the actual data value of x(τ) varies within the window as follows:

[0217] Assume that the actual data of x(τ) within the window range is as follows:

[0218] Data values ​​for time 25 to 35 minutes: 200, 205, 210, 215, 220, 225, 230, 235, 240, 245

[0219] Calculate μ(30):

[0220]

[0221] Calculate σ 2 (30):

[0222]

[0223] These computational steps show how to analyze dialysis data in real time, using dynamically adjusted windowing techniques to capture key changes and calculate time-domain features.

[0224] The embodiment further utilizes segmented regression and period analysis algorithms to gain insight into the time-dependent characteristics of the data, thereby dynamically adjusting the dialysis process to prevent potential problems, using trend analysis functions:

[0225] T(t)=β0+β1t+β2t 2

[0226] Where t is the time point in the time series, β0, β1, and β2 are regression coefficients calculated dynamically by piecewise regression of the data, calculated by least squares or other regression techniques, which depend on the distribution of time points t in the data set and the change pattern of the data.

[0227] Periodic analysis functions:

[0228] P(f)=γsin(2πft+φ)

[0229] Where t represents the time point in the time series, f is the specific frequency of the analysis, and γ and φ are amplitude and phase parameters that are dynamically adjusted through Fourier analysis. γ ranges from 0.1 to 1.0 and is adjusted according to the amplitude changes of the data. φ ranges from 0 to 2π and is dynamically adjusted according to the phase shift of the data.

[0230] Assume that in the dialysis data analysis, it is observed that the pressure and flow rate data have certain trends and periodic changes. Assume that the data from the past few months is calculated as follows:

[0231] β0=10, β1=0.05, β2=0.001; γ=0.5, f=1 (daily cycle), φ=π / 4.

[0232] For a specific time t = 180 days (approximately half a year):

[0233] T(180)=10+0.05×180+0.001×180 2

[0234] T(180)=10+9+32.4=51.4

[0235] This means that the values ​​predicted by the trend analysis show a significant increase over time, reflecting an increase in system load or a long-term burden on the filter membrane.

[0236] For cycle analysis:

[0237] P(180)=0.5sin(2π×1×180+π / 4)

[0238] P(180)=0.5sin(360π+π / 4)

[0239] P(180)=0.5sin(π / 4)≈0.5×0.707=0.3535

[0240] This means that at the 180-day cycle point, the impact of cyclical variations is relatively small, but still exists, related to seasonal factors or scheduled maintenance activities.

[0241] Embodiment 5:

[0242] Combined with Figure 4 ,Implementation In the hemodiafiltration membrane blockage warning system based on machine learning, frequency domain feature extraction is achieved by Fourier transform, and time series data x(t) is obtained from the sensor, with a recording interval of once per minute, and data points from t=0 to t=60 minutes.

[0243] Assume that there is preprocessed data x(t), which contains signal noise, but after noise processing and missing data interpolation, a set of stable time domain data is obtained.

[0244] Fast Fourier Transform (FFT) will be applied to convert these time domain data to frequency domain. The Fourier transform formula is:

[0245]

[0246] Where t is the time point, x(t) is the time domain signal, f is the frequency, and T is the time period.

[0247] Set T = 60 minutes, the length of the time series is 60 minutes. Use discrete Fourier transform DFT or its efficient algorithm fast Fourier transform FFT to complete the frequency domain conversion. The frequency range is from 0Hz to 1Hz to cover possible periodic changes.

[0248] Assume that the following frequency domain characteristics are calculated:

[0249] Amplitude |F(0.1)|=5.3 at frequency f=0.1 Hz;

[0250] Amplitude |F(0.2)|=2.1 at frequency f=0.2 Hz;

[0251] The amplitude |F(0.3)|=1.0 at the frequency f=0.3 Hz.

[0252] These frequency domain features represent the strength of the signal at different frequencies, where the amplitude of the high-frequency component represents a rapidly changing signal, while the low-frequency component represents a slower changing trend. The high amplitude at f = 0.1 Hz indicates some periodic phenomenon or potential abnormality.

[0253] In order to adapt to the non-stationary characteristics of the data, the multi-resolution Fourier transform technology is introduced. The multi-resolution Fourier transform (using STFT) divides the data into small time windows and applies Fourier transform to each window to analyze the change of frequency over time.

[0254] Use STFT to analyze the data, set the time window to 5 minutes, and overlap 2 minutes. Get the frequency domain feature spectrum in each time window to help observe how the frequency components change over time:

[0255] In the window from 0 to 5 minutes, the amplitude of the frequency f = 0.1 Hz is higher;

[0256] In the window between 55 and 60 minutes, the amplitude of the frequency f = 0.2 Hz increases significantly.

[0257] These changes indicate that the performance of the filter membrane is beginning to deteriorate. The early warning system can detect these changes and trigger corresponding warnings or adjustment measures to prevent the filter membrane from clogging.

[0258] The embodiment further introduces a multi-resolution Fourier transform to adapt to the non-stationary characteristics of the data and capture the dynamic changes of different frequency components:

[0259]

[0260] in:

[0261] t is the time point in the time series;

[0262] x(t) is the time domain signal;

[0263] ψ is the mother wavelet function;

[0264] a and b are scale and translation parameters, respectively, which adjust the scaling and position of the wavelet.

[0265] Parameter value range:

[0266] a: scale parameter, usually ranging from 0.1 to 10, used to adjust the expansion and contraction of the wavelet to capture signals of different frequencies;

[0267] b: Translation parameter, which is dynamically adjusted according to the length of the time series data and analysis requirements.

[0268] Assume that during the dialysis process, the frequency components and dynamic changes of the flow rate data x(t) need to be analyzed. The flow rate data is recorded as follows (in mL / min):

[0269] x(t)={200,205,210,215,220,225,230,235,240,245,250}

[0270] The time points t range from 0 to 10 minutes.

[0271] Select Morlet wavelet as the mother wavelet function ψ, and its formula is:

[0272]

[0273] Where f0 is the center frequency, select f0=1Hz.

[0274] The scale parameter a is set to 0.5, 1, 2, representing different frequency components respectively; the translation parameter b ranges from 0 to 10 minutes.

[0275] Compute the wavelet transform:

[0276] For each scale a and time point b, compute the wavelet transform W(a,b). For a=1 and b=5 minutes:

[0277]

[0278] Perform numerical integration at each time point t. Assume that the integral result is 3.5.

[0279] The wavelet coefficients of different scales and time points are obtained according to the calculation results. These coefficients reflect the changes of velocity data at different frequencies and times. By analyzing the time-frequency diagram of the wavelet coefficients, the abnormal frequency components and dynamic change patterns in the velocity data can be identified.

[0280] For a=0.5, at time t=5 minutes:

[0281]

[0282] Assume that the result of numerical integration is 2.0.

[0283] For a=2, at time t=5 minutes:

[0284]

[0285] Assume that the numerical integration result is 4.5.

[0286] These results indicate that the frequency content and dynamics of flow velocity data vary at different scales.

[0287] The last embodiment extracts multi-scale frequency domain features from the multi-resolution Fourier transform, including high-frequency components, low-frequency components, and medium-frequency components. The frequency domain feature extraction formula is:

[0288]

[0289] in:

[0290] W k (f, t) is the transform coefficient at the kth scale, representing the frequency domain information at a specific scale and time point;

[0291] f is the frequency, ranging from low frequency to high frequency, between 0 and 1 Hz;

[0292] t is the time point, which is adjusted according to the specific time series data.

[0293] set up:

[0294] Transformation coefficients of high frequency W1(f,t): W1(0.3,t)=2.5, W1(0.6,t)=1.8;

[0295] Transformation coefficients of intermediate frequency W2(f,t): W2(0.3,t)=3.1, W2(0.6,t)=2.3;

[0296] The transformation coefficients of low frequency W3(f,t): W3(0.3,t)=4.2, W3(0.6,t)=3.7.

[0297] The frequency domain feature extraction formula is used to calculate the eigenvalue at each frequency and time point.

[0298] For f = 0.3 and time point t:

[0299] E(0.3,t)=|W1(0.3,t)| 2 +|W2(0.3,t)| 2 +|W3(0.3,t)| 2

[0300] E(0.3,t)=2.5 2 +3.1 2 +4.2 2

[0301] E(0.3,t)=6.25+9.61+17.64

[0302] E(0.3,t)=33.5

[0303] For f = 0.6 and time point t:

[0304] E(0.6,t)=|W1(0.6,t)| 2 +|W2(0.6,t)| 2 +|W3(0.6,t)| 2

[0305] E(0.6,t)=1.8 2 +2.3 2 +3.7 2

[0306] E(0.6,t)=3.24+5.29+13.69

[0307] E(0.6,t)=22.22

[0308] The high eigenvalue E(0.3, t) = 33.5 at the frequency f = 0.3 indicates the presence of periodic changes or potential abnormal conditions, while the eigenvalue at the frequency f = 0.6 is relatively low, indicating that the influence of this frequency component is small.

[0309] Embodiment 6:

[0310] Embodiments use machine learning models and feature engineering methods to understand the impact of chemical composition fluctuations on filtration membrane performance, improve the accuracy and timeliness of early warning, and use a time-varying coefficient model:

[0311]

[0312] in:

[0313] t0 and t represent the start and end time of the integration, respectively, covering the time window of the analysis;

[0314] c(τ) represents the concentration of chemical components at time point τ;

[0315] μ c (τ) is the moving average of the chemical component concentration.

[0316] Assume that the concentration change of potassium ions in the dialysate is monitored during dialysis, and the data from t0 = 0 minutes to t = 60 minutes are recorded. The chemical component concentration c(τ) is recorded as follows (unit: mmol / L):

[0317] c(t)={4.0,4.1,4.2,4.3,4.4,4.5,4.6,4.7,4.8,4.9,5.0}

[0318] Time points t range from 0 to 60 minutes.

[0319] Moving average μ c (τ) can be calculated by simple moving average. Assuming the average value is calculated every 10 minutes:

[0320]

[0321] Calculate the eigenvalue E(c,t) using the time-varying coefficient model:

[0322]

[0323] Assume that the moving average is updated every 10 minutes, and the moving average is as follows:

[0324] μ c (0-10)=4.05

[0325] μ c (10-20)=4.15

[0326] μ c (20-30)=4.25

[0327] μ c (30-40)=4.35

[0328] μ c (40-50)=4.45

[0329] μ c (50-60)=4.55

[0330] Compute the sum of squares of differences over the time period:

[0331]

[0332] Assume that the sum of squared differences in each 10-minute period is calculated as:

[0333] (4.0-4.05) 2 +(4.1-4.05) 2 +…+(5.0-4.55) 2

[0334] =0.0025+0.0025+0.0625+0.0025+0.0625+0.0025+0.0025+0.0625+0.0025

[0335] +0.0025=0.208

[0336] Therefore, the integral calculation result is:

[0337]

[0338] Integration interval length:

[0339]

[0340] Final eigenvalue calculation:

[0341]

[0342] This result E(c,60)=0.0208 indicates that the fluctuation of potassium ion concentration is relatively small within the 60-minute time window.

[0343] The embodiment further uses a multi-parameter dynamic simulation model to analyze the concentration changes of different chemical components and their effects on the performance of the filtration membrane, using the formula:

[0344]

[0345] in:

[0346] t represents time;

[0347] Represents a collection of different chemical components;

[0348] c j (t) is the concentration of the jth chemical component at time t;

[0349] b0,b1,…,b 2m are the coefficients in the prediction model;

[0350] b0 is the baseline response;

[0351] b1,…,b m is the influence intensity of each chemical component;

[0352] b j+m is the time decay rate corresponding to the effect of chemical composition;

[0353] m is the total number of chemical components.

[0354] Parameter value range:

[0355] b0 ranges from 0 to 1, indicating the strength of the baseline response;

[0356] b j The value range is between 0.1 and 2, indicating the influence intensity of chemical composition;

[0357] b j+m The value range is between 0.01 and 0.1, indicating the time decay rate.

[0358] Assume that the concentration changes of two key chemical components, urea and potassium ions, are monitored during dialysis. The urea and potassium ion concentration data from 0 to 60 minutes are recorded and represented as c1(t) and c2(t), respectively.

[0359] Urea concentration c1(t) = {6.0, 6.1, 6.2, 6.3, 6.4, 6.5, 6.6, 6.7, 6.8, 6.9, 7.0} (unit: mmol / L);

[0360] Potassium ion concentration c2(t) = {4.0, 4.1, 4.2, 4.3, 4.4, 4.5, 4.6, 4.7, 4.8, 4.9, 5.0} (unit: mmol / L);

[0361] Time point t = {0, 10, 20, 30, 40, 50, 60} minutes.

[0362] Assume that the following model coefficients are selected: b0=0.5, b1=1.0, b2=0.8, b3=0.02, b4=0.03.

[0363] The predicted value at t = 60 minutes is calculated using a multi-parameter dynamic simulation model:

[0364]

[0365] Substitute specific data:

[0366]

[0367] Calculate each term:

[0368] 1.0e1.2 ≈1.0×3.32=3.32

[0369] 0.8e 1.8 ≈0.8×6.05=4.84

[0370] Final calculation results:

[0371]

[0372] The predicted value calculated The influence of the current chemical component concentration on the filtration membrane performance can be determined.

[0373] Embodiment 6:

[0374] In the hemodiafiltration membrane blockage early warning system based on machine learning, a random forest is used to integrate multiple decision trees to enhance the stability and prediction ability of the model. The formula is defined as follows:

[0375]

[0376] Where k is the fold number; J i is the index set of the i-th fold; y j is the actual value; is the predicted value.

[0377] Assume that 1000 data points are collected during dialysis, recording flow rate, pressure, and urea and potassium ion concentration data. Apply 5-fold cross validation to evaluate model performance. Divide the 1000 data points into 5 subsets, each containing 200 data points. Calculate the prediction error for each subset and take the average as the overall error of the model.

[0378] Assume the actual and predicted values ​​are as follows:

[0379] Subset 1: actual value y = [1.0, 1.2, 1.4, 1.6, 1.8], predicted value

[0380] Subset 2: actual value y = [2.0, 2.2, 2.4, 2.6, 2.8], predicted value

[0381] Subset 3: actual value y = [3.0, 3.2, 3.4, 3.6, 3.8], predicted value

[0382] Subset 4: actual value y = [4.0, 4.2, 4.4, 4.6, 4.8], predicted value

[0383] Subset 5: actual value y = [5.0, 5.2, 5.4, 5.6, 5.8], predicted value

[0384] Calculate the root mean square error RMSE for each subset:

[0385]

[0386] RMSE2=0.1

[0387] RMSE3=0.1

[0388] RMSE4=0.1

[0389] RMSE5=0.1

[0390] Average RMSE:

[0391]

[0392] Through cross-validation technology, the average RMSE of the model is 0.1, indicating that the model has a small prediction error on unseen data. The stability and high prediction ability of the random forest model enable it to effectively detect possible abnormalities during dialysis and issue early warnings in practical applications.

[0393] Next, the embodiment uses the Bayesian optimization method to optimize the hyperparameters of the model to improve the performance and prediction accuracy of the model. By optimizing the formula:

[0394]

[0395] Θ is a set of hyperparameters; L(Θ) is the loss function; λ is the regularization coefficient.

[0396] Parameter value range:

[0397] The range of the hyperparameter θ depends on the specific model setting, including the number of trees and the maximum depth of the random forest.

[0398] The regularization coefficient λ ranges from 0.01 to 0.1 and is used to control the strength of regularization to prevent overfitting.

[0399] Assuming that the random forest model is used for optimization, the hyperparameters include the number of trees (n_estimators) and the maximum depth (max_depth). The best hyperparameter combination is found through Bayesian optimization to minimize the model's loss function L(Θ).

[0400] The number of trees n_estimators ranges from 10 to 100;

[0401] The maximum depth max_depth ranges from 3 to 20;

[0402] The regularization coefficient λ is set to 0.05.

[0403] Assume that the loss function L(θ) represents the mean square error MSE of the model.

[0404] Assume that some of the iterations during Bayesian optimization result in the following:

[0405] Iteration 1: n_estimators = 30, max_depth = 10, L(Θ) = 0.12

[0406] Iteration 2: n_estimators = 50, max_depth = 15, L(Θ) = 0.10

[0407] Iteration 3: n_estimators = 40, max_depth = 12, L(Θ) = 0.11

[0408] Finally, Bayesian optimization determines the optimal hyperparameter combination:

[0409] Optimal Θ * ={n_estimators=50,max_depth=15}

[0410] Minimum loss L(Θ * )=0.10

[0411] Optimization formula application:

[0412]

[0413] The optimal hyperparameter combination of the random forest model is found through Bayesian optimization, so that the loss function of the model reaches the minimum value of 0.10.

[0414] The last example extracts and optimizes features and combines ensemble learning technology to more accurately predict the risk of clogging of the filtration membrane and improve the safety and effectiveness of the dialysis process. The feature engineering formula is used:

[0415]

[0416] Where x is the original feature set; x_n is the nth feature; α n is the feature weight.

[0417] Parameter value range:

[0418] Feature weight α n The value range is between 0 and 1, and is adjusted according to the importance and influence of the feature.

[0419] The collected original feature set x includes: flow rate x1 = 200 HmL / min, pressure x2 = 120 HmmHg, urea concentration x3 = 6.0 mmol / L, ion concentration x4 = 4.0 mmol / L.

[0420] Assume that the feature weights α are set as follows: α1=0.8, α2=0.7, α3=0.9, α4=0.6.

[0421] Use the feature engineering formula to calculate the new eigenvalue F(x):

[0422]

[0423] Calculate the contribution of each feature:

[0424] F(x1)=0.8log(1+e 200 )≈0.8×200=160(using log(1+e 200 )≈200)

[0425] F(x2)=0.7log(1+e 120 )≈0.7×120=84(using log(1+e 120 )≈120)

[0426] F(x3)=0.9log(1+e 6.0 )≈0.9×6.0=5.4

[0427] F(x4)=0.6log(1+e 4.0 )≈0.6×4.0=2.4

[0428] Total eigenvalue:

[0429] F(x)=160+84+5.4+2.4=251.8

[0430] During dialysis, the random forest and gradient boosted tree models produced the following predictions, respectively:

[0431] Random Forest Prediction:

[0432] Gradient Boosted Tree Prediction:

[0433] These two predicted values ​​are input into the meta-learner as new features. Assume that the meta-learner is a linear regression model:

[0434]

[0435] Assume weights w1=0.6 and w2=0.4:

[0436]

[0437] Through advanced feature engineering and ensemble learning methods, the system predicts the value It is close to the true value 1, indicating that the model has high prediction accuracy.

[0438] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A hemodiafiltration membrane blockage early warning method based on machine learning, characterized in that The following steps are involved: S1. First, collect the flow sensor data, pressure sensor data, chemical sensor data and time series data during the hemodialysis process, clean and standardize the collected data, remove noise and outliers, and fill in missing data; The data is subjected to standardization process: S1.

1. First, an adaptive noise recognition algorithm is used to learn and remove noise from historical data, and an autoencoder anomaly detection model based on deep learning is used to identify and process outliers in the data. S1.2, then adopt the interpolation method based on the prediction model, using the time series prediction model or deep learning network to predict and fill in the missing data values; S1.3, finally, implement the standardization technology of dynamically adjusting the data range, and dynamically adjust the standardization parameters according to the characteristics of the real-time data stream; S2. Then, feature extraction and selection are performed on the processed data, and time domain features including mean, peak value, trend and periodic changes are extracted from each sensor data, and frequency domain features are extracted by Fourier transform and changes in chemical component concentrations in blood and dialysate are analyzed to extract key change features; S3. Then use the machine learning model random forest to divide the processed data into training set and test set, train the model with the training set and use cross-validation technology to evaluate the model performance, use Bayesian optimization to adjust parameters to find hyperparameters, and further improve the model performance through feature engineering and ensemble learning methods; S4. Finally, a real-time data acquisition system is built to obtain data from various sensors during the dialysis process in real time, update the machine learning model input data in real time, use the trained model to predict the clogging risk of the filter membrane in real time, set the risk threshold, and trigger the early warning system when the predicted clogging risk exceeds the threshold.

2. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 1 is characterized in that The process of constructing the adaptive noise recognition algorithm includes: firstly, using the adaptive noise recognition algorithm to remove the noise in the data based on historical data learning, and then using the adaptive noise recognition function: Where x represents the input data, which is time or other continuous measurements related to the dialysis process, θ and λ are parameters that are dynamically adjusted based on real-time data to identify and filter out noise related to the dialysis pump frequency, and t is the integration variable; At the same time, the autoencoder anomaly detection model based on deep learning is used to identify and process outliers in the data, and the anomaly measurement function is applied: Where z is the reconstruction error obtained from the autoencoder, and σ is the standard deviation dynamically adjusted according to the distribution of abnormal data; In addition, these processed data are integrated into the dialysis machine control system to provide real-time feedback to adjust dialysis parameters or remind medical staff, using a dynamic threshold adjustment strategy: h(t,α,β,γ)=αtexp(-βt)+γcos(βt) Where t represents the usage time of the filter membrane, α, β, and γ are parameters that are dynamically adjusted based on the dialysate quality and the condition of the filter membrane.

3. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 1 is characterized in that The construction process of the interpolation method based on the prediction model is as follows: S1. Use a long short-term memory network LSTM or a gated recurrent unit GRU to train the model specifically to identify data patterns that cause filter blockages, and use a dynamic window adjustment function: Where t represents the time point in the time series, δ and ω are parameters that are dynamically adjusted based on the historical data of the membrane pressure and flow rate; S2. Then, we combine the convolutional neural network (CNN) and the recurrent neural network (RNN) to process the multi-dimensional data in the dialysis process, and use the advanced multi-dimensional feature analysis function: Where p represents the concentration of the chemical component, q and r represent measures of flow rate and pressure, respectively; S3. In addition, a continuous learning mechanism is introduced to adjust the formula through feedback: Where s represents each new data point in the dialysis time series, and α and β are model performance parameters that are dynamically adjusted based on the new data.

4. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 1 is characterized in that The standardized technology construction method is as follows: First, we select key parameters that directly affect the performance of the filtration membrane, such as pressure difference, flow rate, and chemical composition of the dialysate, and use the sliding window technology: Real-time data acquisition is performed, where t represents the time point in the time series, δ and ω are dynamically adjusted parameters, and the size of the data window is dynamically adjusted according to the precursor of clogging of the filter membrane; Next, the adaptive normalization formula is applied according to the real-time monitored changes in the filter membrane status: Where t represents the time point in the time series, μ(t) and σ(t) are the dynamic mean and standard deviation calculated based on the real-time data window, and λ is the smoothing parameter introduced to adjust the sensitivity in the standardization process; Finally, implement a feedback adjustment mechanism: h(α,β,x)=α·log(1+e βx ) Where x represents the specific variable value of real-time data input, and α and β are parameters adjusted based on feedback from the early warning model performance.

5. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 1 is characterized in that The data feature extraction and selection implementation process is as follows: S1. First, extract temporal features from each sensor data, including mean, variance, peak, trend and periodic change. Dynamically calculate these features through sliding window technology to capture the temporal variation pattern in the data. And detect the long-term trend and periodic pattern in the data through segmented regression and period analysis algorithm. S2. Secondly, the Fourier transform is applied to extract the frequency domain features of the processed time domain data, and the multi-resolution Fourier transform is introduced to adapt to the non-stationary characteristics of the data and capture the dynamic changes of different frequency components; S3. Finally, analyze the concentration changes of chemical components in blood and dialysate, and extract key change features to reveal the impact of chemical composition fluctuations on filtration membrane performance.

6. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 5 is characterized in that The method for extracting time domain features from each sensor data includes: First, extract time domain features from each sensor data, including mean, variance, peak, trend, and periodic changes, and dynamically calculate these features through sliding window technology, using a dynamic window adjustment function: Where t represents the time point in the time series, α and t0 are adjustment parameters, and the window size is dynamically adjusted according to the data change rate; In each sliding window, by calculating the time domain features, the mean: And the variance: Where μ(t) represents the mean value of the data in the sliding window at the current time t, W(t) represents the current window size, which is determined by the dynamic window adjustment function, and x(τ) represents the data value at the time point τ; Detect long-term trends and cyclical patterns in data using piecewise regression and periodic analysis algorithms, and apply trend analysis functions: T(t)=β0+β1t+β2t 2 Where t represents the time point in the time series, β0, β1 and β2 are regression coefficients dynamically calculated through piecewise regression to identify trend changes in the data; Using periodic analysis function: P(f)=γsin(2πft+φ) Where t represents the time point in the time series, f represents the frequency, and γ and φ are parameters that are dynamically adjusted based on Fourier analysis to identify periodic patterns in the data.

7. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 5 is characterized in that The frequency domain feature extraction process is as follows: S1. Apply fast Fourier transform FFT to convert the preprocessed time domain data into frequency domain to extract frequency domain features, using the Fourier transform formula: Where t represents the time point in the time series, x(t) represents the time domain signal, f represents the frequency, and T is the time period; S2. Secondly, the multi-resolution Fourier transform is introduced to adapt to the non-stationary characteristics of the data and capture the dynamic changes of different frequency components. The formula is: Where t represents the time point in the time series, x(t) is the time domain signal, ψ is the mother wavelet function, and a and b are the scale and translation parameters respectively; S3. Extract multi-scale frequency domain features from multi-resolution Fourier transform, including high-frequency components, low-frequency components, and medium-frequency components, using the frequency domain feature extraction formula: Where Wk(f,t) is the transformation coefficient at the kth scale, f is the frequency, and t is the time.

8. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 5 is characterized in that The process of analyzing the concentration changes of chemical components in blood and dialysate is as follows: First, a machine learning model is applied to analyze the dynamic data of chemical composition, and key change features are extracted through a series of feature engineering methods, including the time-varying coefficient model through: Where t0, t represent the start and end time of the integration, covering the time window of the analysis, c(τ) represents the concentration of the chemical component at time point τ, μ c (τ) is the moving average of the chemical component concentration; Then, using the multi-parameter dynamic simulation model: t represents time, represents a collection of different chemical components, c j (t) is the concentration of the jth chemical component at time t, b0, b1, …, b 2m are the coefficients in the prediction model, where b0 is the baseline response, b1,…,b m is the influence intensity of each chemical component, b j+m is the time decay rate of the corresponding chemical component effect, and m is the total number of chemical components, which indicates the number of different chemical components used in the model.

9. The hemodiafiltration membrane blockage early warning method based on machine learning according to claim 1, characterized in that The method for constructing a machine learning model includes: first, using a customized machine learning model random forest and using a split cross-validation technique for time series data: Where k is the number of folds, J i is the index set of the i-th fold, y j is the actual value, is the predicted value, which evaluates the performance of the model on unseen data; Next, the Bayesian optimization method is used to optimize the hyperparameters of the model, through the optimization formula: Where Θ is the set of hyperparameters, L(Θ) is the loss function, and λ is the regularization coefficient; Finally, advanced feature engineering and ensemble learning methods are applied, using the feature engineering formula: Where x is the original feature set, x n is the nth feature, α n is the feature weight.

Citation Information

Patent Citations

  • Satellite internet traffic identification system based on machine learning

    CN117033928A

  • Hemodialysis monitoring system and method based on hemodialysis machine

    CN117153336A

  • Personalized coronary artery interventional therapy planning system based on multi-modal data fusion

    CN118280522A

  • System and method for prediction of intradialytic adverse event and computer readable medium thereof

    US20230094657A1

  • Blood treatment systems and methods

    US20230338632A1

Cited By

  • Intelligent voltage regulating system of emergency direct-current power supply

    CN119179364A

  • Method, system and equipment for dynamically monitoring quality of premixed concrete

    CN119851837A

  • Hemodialysis loop blockage early warning device based on big data

    CN120617667A