A method and system for intelligent perturbation and analysis of hemodialysis data

Abstract

This invention discloses an intelligent perturbation and analysis method and system for hemodialysis data, relating to the field of dialysis data analysis. The method includes: during hemodialysis, superimposing a perturbation signal onto the blood input pump to cause a preset configuration change in the blood parameters input to the dialyzer, thereby forming an input perturbation signal; during the perturbation application period, real-time acquisition of blood parameters at the dialyzer's blood output end to form a corresponding output response signal; performing cross-correlation analysis on the input perturbation signal and the output response signal to calculate the time delay parameter between them; periodically repeating the detection and performing regression analysis on the time delay parameter; when the time delay parameter continuously increases over time, determining that the dialyzer has a clogging trend and outputting a warning message. This invention achieves early detection of dialyzer clogging through dynamic time delay characteristics without adding sensors, improving the safety of the hemodialysis system.

Description

Technical Field

[0001] This invention relates to the field of dialysis data analysis, and more specifically, to a method and system for intelligent perturbation and analysis of hemodialysis data. Background Technology

[0002] Hemodialysis is one of the most common blood purification methods for patients with end-stage renal disease. Its core process involves using a semipermeable membrane to remove metabolic waste and excess water from the blood, while replenishing fluid and electrolyte balance through dialysate. During operation, the dialysis machine relies on a blood pump and a dialysate inlet pump to maintain the flow of blood and dialysate. These pumps create a transmembrane pressure difference across the filter membrane, enabling solute diffusion and convection exchange. However, the dialysis process is affected by various complex factors, including changes in blood viscosity, filter fouling, fluctuations in transmembrane pressure difference, and nonlinear responses in pump control feedback, making precise monitoring of the system's dynamic state difficult.

[0003] In existing technologies, dialyzers primarily rely on static parameters such as pressure, flow rate, and conductivity for safety monitoring, lacking real-time analysis methods for the system's dynamic characteristics. Especially in the early stages of blockage, traditional methods often only detect it when the blockage becomes severe. Furthermore, the blood and dialysate sides are essentially coupled systems, their dynamic responses influencing each other. However, current dialyzer control strategies typically treat both sides as independent entities, failing to fully utilize this coupling characteristic for signal analysis and diagnosis. Therefore, how to achieve early detection and accuracy optimization of blockages through the dynamic relationship between the blood and dialysate pathways without adding additional sensors has become a crucial problem that needs to be solved. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent perturbation and analysis method and system for hemodialysis data, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent perturbation and analysis of hemodialysis data, comprising:

[0007] S1: During the blood input process into the dialysis machine, a disturbance signal is superimposed on the blood input pump to cause a preset configuration change in the blood parameters input into the dialysis machine, thereby forming an input disturbance signal;

[0008] S2: During the period of disturbance application, blood parameters at the blood output end of the dialyzer are collected in real time to form a corresponding output response signal;

[0009] S3: Perform cross-correlation analysis on the input disturbance signal and the output response signal, and calculate the time delay parameter Δτ corresponding to the disturbance signal between them;

[0010] S4: Execute steps S1~S3 once every preset time interval to obtain the time delay parameter Δτ that changes with time;

[0011] S5: Perform regression analysis on the time delay parameter Δτ to obtain the time delay change trend;

[0012] S6: If the regression analysis shows that the time delay parameter Δτ continues to increase over time, it is determined that the dialysis machine has a blockage trend and an early warning message is output.

[0013] Furthermore, the cross-correlation analysis includes: filtering and normalizing the input disturbance signal and the output response signal respectively; calculating the cross-correlation function of the input disturbance signal and the output response signal within a preset time window, and determining the time offset corresponding to the peak value of the cross-correlation function as the time delay parameter Δτ.

[0014] Furthermore, the method also includes:

[0015] Perform the same perturbation and cross-correlation analysis steps as S1~S6 on the dialysate input to the dialyzer and the dialysis waste liquid output from the dialyzer to obtain the time delay parameters on the dialysate side and perform corresponding regression analysis and early warning.

[0016] Furthermore, perturbation signals are simultaneously applied to the blood infusion pump and the dialysate infusion pump. The two perturbation signals are set to have the same frequency and waveform. The coupling relationship between the two perturbations is changed by adjusting the phase difference and amplitude ratio. The cross-correlation function of the blood side is calculated after each perturbation cycle. When the peak amplitude of the cross-correlation function of the blood side reaches its maximum, the corresponding phase difference and amplitude ratio are recorded as the phase difference and amplitude ratio used in actual detection.

[0017] Furthermore, the phase difference and amplitude ratio of the disturbance signal are optimized based on a regression model, specifically including:

[0018] The peak amplitude of the cross-correlation function and the corresponding phase difference and amplitude ratio samples were collected during multiple perturbation cycles to construct a training dataset.

[0019] A regression model was trained using the training dataset to establish a mapping relationship between the perturbation parameters and the peak amplitude of the blood-side cross-correlation.

[0020] In actual testing, the phase difference and amplitude ratio that maximize the peak amplitude of the blood-side cross-correlation function are predicted by regression model.

[0021] Furthermore, the phase difference and amplitude ratio of the perturbation signal are self-learned and optimized based on a reinforcement learning model, specifically including:

[0022] The phase difference and amplitude ratio are used as action variables in reinforcement learning, and the peak amplitude of the cross-correlation function is set as the reward signal.

[0023] The agent performs actions and collects output signals in each perturbation cycle, and updates the policy network parameters based on the reward signal.

[0024] Through a multi-round perturbation cycle strategy iteration process, the algorithm gradually converges to the optimal phase difference and amplitude ratio that maximizes the peak amplitude of the blood cross-correlation function.

[0025] Furthermore, the waveform is a sine wave, a square wave, or multiple single-pulse signals, and the waveform or the frequency is included in the range of optimized parameters that maximize the peak amplitude of the blood-side cross-correlation function;

[0026] The blood parameter is either a flow rate parameter or a blood pressure parameter, and a threshold upper limit is set for the amplitude of the disturbance applied to the flow rate parameter or blood pressure parameter.

[0027] Furthermore, when performing regression analysis on the changing trend of the time delay parameter Δτ, the time delay parameter sequence obtained by multi-period sampling is used as the regression sample. The slope of the time delay parameter is calculated by the least squares method or linear regression algorithm. When the slope is positive and exceeds a preset threshold, it is determined that the time delay parameter Δτ continues to increase with time.

[0028] Furthermore, the method uses a blood simulation kit to replace a real patient for testing.

[0029] This invention also discloses an intelligent perturbation and analysis system for hemodialysis data to implement the above method, comprising:

[0030] The disturbance control module is used to generate an input disturbance signal by superimposing a disturbance signal on the blood input pump during the blood input process of the dialyzer, thereby causing a preset configuration change in the blood parameters input to the dialyzer.

[0031] The data acquisition module is used to collect blood parameters at the blood output end of the dialyzer in real time during the application of disturbance, and generate corresponding output response signals;

[0032] The time delay calculation module is used to perform cross-correlation analysis on the input disturbance signal and the output response signal, and calculate the time delay parameter Δτ corresponding to the disturbance signal between the two.

[0033] The periodic execution module is used to execute steps S1~S3 once at a preset time interval to obtain the time delay parameter Δτ that changes with time.

[0034] The regression analysis module is used to perform regression analysis on the time delay parameter Δτ to obtain the time delay change trend;

[0035] The blockage warning module is used to determine that the dialysis machine has a blockage trend when the regression analysis shows that the time lag parameter Δτ continues to increase over time, and to output warning information.

[0036] The advantage of this invention over existing technologies lies in its ability to intelligently analyze the dynamic characteristics of the dialysis system by introducing a micro-amplitude perturbation signal during the hemodialysis process. Its core innovation lies in applying a pre-configured perturbation signal to the blood input pump, collecting the response signal from the blood output end, and performing cross-correlation analysis to calculate the time delay parameter of the perturbation propagation. This time delay parameter reflects the delayed characteristics of blood transmission through the filter membrane and tubing, and can sensitively characterize changes in membrane resistance. When the time delay parameter continuously increases over time, an early warning of membrane clogging trends can be provided. This method requires no additional sensors, does not alter the original dialysis process, and is characterized by safety, non-invasiveness, and high sensitivity.

[0037] This invention applies a perturbation simultaneously to the dialysate pathway and performs the same cross-correlation and regression analysis process to obtain the time lag trend on the dialysate side. This trend is then combined with the time lag trend on the blood side for joint determination, thereby achieving dynamic diagnosis of the transmembrane dual pathway. When the time lag changes on both sides increase in the same direction, the system can confirm the reliability of the decrease in membrane flux, effectively reducing the false alarm rate.

[0038] This invention further improves the signal-to-noise ratio and computational accuracy of time-delay measurements by setting the frequency and waveform of the perturbation signals on the blood and dialysate sides to be the same and adjusting the phase difference and amplitude ratio to change the coupling relationship between the two sides. This maximizes the peak amplitude of the cross-correlation function. Based on this, regression and reinforcement learning models are introduced to automatically search for the optimal phase difference, amplitude ratio, or waveform parameters, achieving adaptive optimization and self-learning adjustment of the perturbation signal. This ensures the system maintains high-precision detection performance under different dialysis conditions.

[0039] This invention not only enables early intelligent prediction of membrane clogging trends, but also significantly improves the stability and adaptability of detection through coupling analysis and machine learning optimization. It can be used for online monitoring of dialysis machines, system calibration, and blood simulant testing, and has high clinical and engineering application value. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0041] Figure 2 This is a schematic diagram of the main process of the blood test based on the present invention;

[0042] Figure 3 This is a schematic diagram of the present invention based on adjusting perturbations using regression models or reinforcement learning. Detailed Implementation

[0043] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0044] As shown in Figures 1 and 2, in the method of this invention, during hemodialysis, to monitor the potential risk of blockage inside the dialyzer in real time, a small perturbation signal can be actively introduced to stimulate the system response, thereby capturing subtle changes in the flow resistance inside the dialyzer. The core of this method lies in using the dynamic relationship between the controllable perturbation applied at the input end and the response signal collected at the output end to quantify the delay characteristics of blood flow in the dialyzer. When the fiber bundles inside the dialyzer gradually clot or become blocked, the propagation time of blood from the input end to the output end gradually increases. By continuously tracking this time lag change, problems can be detected in advance and warnings can be issued, avoiding serious malfunctions or treatment interruptions during dialysis.

[0045] In practice, a carefully designed perturbation signal is first superimposed on the blood input tubing of the dialyzer by controlling the rotational speed of the blood input pump. This perturbation signal can be a periodic, small fluctuation, such as a sine wave variation with an amplitude within 5% of the normal flow rate superimposed on the normal pump speed, or a short square wave pulse sequence, or even several discrete single pulse signals. These waveforms are chosen because they have clear spectral characteristics, facilitating subsequent signal processing and correlation analysis. The frequency of the perturbation signal is typically set between 0.1 Hz and 2 Hz. This range effectively stimulates the dynamic response within the dialyzer without significantly interfering with the patient's hemodynamics. To ensure safety, the perturbation amplitude must be strictly limited for blood parameters such as flow rate or blood pressure. For example, the flow rate perturbation should not exceed 8% of the baseline flow rate, and the blood pressure perturbation should not exceed 5 mmHg, thereby preventing any potential physiological risks. In another embodiment, the dialyzer input tubing can be divided into two sections: one close to the patient, and the other with the perturbation signal applied. A buffer zone is set between the two sections so that the section close to the patient is not affected by the perturbation. In a more stringent embodiment, to avoid risks to real patients, a blood dummy pack can be connected to replace the patient to detect potential blockages in the dialysis machine, thus preventing potential blockages when the dialysis machine is connected to a real patient.

[0046] Furthermore, after the perturbation signal is applied, parameter data from the dialyzer's blood output end needs to be continuously acquired in real time throughout the entire duration of the perturbation. These output parameters are consistent with the input parameters, i.e., monitoring flow rate or blood pressure, forming an output response signal corresponding to the input perturbation. The acquisition device can be a high-precision flow sensor or pressure sensor, with a sampling frequency at least 10 times the frequency of the perturbation signal to ensure sufficient signal detail is captured. Both the input perturbation signal and the output response signal are recorded, forming a pair of time-series data.

[0047] Next, cross-correlation analysis is performed on this pair of signals to extract time delay information. The purpose of cross-correlation analysis is to find the delay time of the output response relative to the input disturbance. First, the two signals are preprocessed, including using low-pass filters to remove high-frequency noise. For example, a Butterworth filter with a cutoff frequency of 5 Hz is used to filter the blood input and output signals to eliminate high-frequency components caused by pump control pulsation and electrical interference. Then, the filtered signals are normalized to standardize the signal amplitude to a range of zero mean and unit variance, thereby eliminating amplitude scale differences between different sampling channels.

[0048] After preprocessing, the cross-correlation function is calculated within a preset time window. This window length is typically set to 3 to 5 times the perturbation period to fully capture the main information of the signal interaction. Cross-correlation function Defined as:

[0049] ;

[0050] in, Input signals into the blood, For blood to output signals, This is the time offset. The initial time of integration can be chosen as the starting time of the disturbance application. The integration window length can be selected as a multiple of the estimated blood outflow time from the dialyzer; for example, it could be selected as 1.5 times the time required to observe the initial blood flow into and out of the dialyzer. By calculating different... Below This allows us to obtain the time offsets of the two signals. The degree of similarity between the two signals. The peak position of the cross-correlation function corresponds to the moment when the two signals are most similar, and its corresponding time offset. That is, the time delay parameter .

[0051] This time delay parameter reflects the dynamic delay characteristics of blood propagation from the inlet to the outlet, remaining relatively stable under normal conditions of unobstructed filter membrane. However, when the filter membrane resistance increases or there is local blockage, the inertia of blood flow and the transmembrane pressure response change, leading to… The trend of continuous increase can be used as an early indicator of a blocking trend.

[0052] To capture the dynamic changes in time lag, the process of applying perturbations, acquiring signals, and performing cross-correlation analysis is repeated at fixed intervals. This preset interval can be set according to clinical needs, for example, every 5 or 10 minutes, with each perturbation lasting 30 seconds to 1 minute. After multiple executions, a sequence of time lag parameters evolving over time is obtained, such as the Δτ values ​​at 10, 20, and 30 minutes after the start of dialysis. These values ​​form a time series for subsequent trend analysis.

[0053] In a further embodiment, regression analysis is performed on this time-delay parameter sequence to reveal its trend over time. The regression analysis uses time-delay parameters obtained from multi-period sampling as sample points, fits a straight line using the least squares method, and calculates the slope value. For more complex trend capture, multinomial regression or moving average filtering can be used to smooth the data. The specific process of slope calculation involves constructing a linear model with time as the independent variable and the time-delay parameter as the dependent variable. ,in The slope For time, It is the intercept. The solution is found to minimize the sum of squared residuals. Value, when this slope If the time lag parameter is positive and exceeds a preset threshold, such as 0.5 seconds per hour, it is determined that the time lag parameter is continuously increasing over time, indicating that the internal resistance of the dialysis machine is increasing and there is a tendency for blockage. At this point, an early warning message is immediately output, such as a flashing red alarm on the dialysis machine's display, an audible alert, or a message sent to the monitoring system at the nurse station. The warning message includes the current time lag value, slope estimate, and estimated time of blockage, helping medical staff to intervene promptly, such as replacing the filter or adjusting the anticoagulant dosage.

[0054] Furthermore, to make the detection more comprehensive, the same perturbation and analysis can be performed simultaneously on the dialysate side. At the dialysate input end, a perturbation signal is superimposed by controlling the dialysate input pump. In some embodiments, this perturbation can be applied independently of the blood side, while in others, it can be coordinated. At the dialysate output end, waste fluid parameters are collected to form an input perturbation and output response signal pair. Then, the same cross-correlation analysis is performed to obtain the time delay parameters on the dialysate side. This process is repeated at preset time intervals to obtain a time delay sequence, which is then subjected to regression analysis. If the slope is positive and exceeds a threshold, a blockage trend is determined on the dialysate side, and an early warning is issued. This dual-sided monitoring can distinguish between fibrin coagulation on the blood side and scaling on the dialysate side, providing more accurate fault location.

[0055] To further optimize the bilateral perturbation, this application applies perturbation signals simultaneously to both the blood inlet pump and the dialysate inlet pump, preferably setting both perturbations to the same frequency and waveform, such as a 1 Hz sine wave. Under this condition, the two perturbations create a coupling-like vibration effect on both sides of the dialyzer's semipermeable membrane. Pressure or flow perturbations on both the blood and dialysate sides are transmitted to the other side through the semipermeable membrane. Although the dialysis process itself has a significant damping effect, and the signal amplitude inevitably attenuates during transmission, the simultaneous perturbations on both sides can, to some extent, superimpose, making the effective perturbation signal strength reaching the blood output end after passing through the dialyzer higher than that of a single-sided perturbation. This compensates for some signal attenuation and improves the observable perturbation signal strength.

[0056] In the coupling process of the two perturbations, the phase difference has the most direct and significant impact on the final output signal. When the phase difference between the perturbations on the blood side and the dialysate side is adjusted appropriately, a superposition enhancement effect can be formed on both sides of the semipermeable membrane, that is, peaks superimpose with peaks and troughs superimpose with troughs, thereby enhancing the total perturbation amplitude. If the phase difference is not adjusted appropriately, when the two interact internally, one side is at a peak while the other side is close to a trough, the perturbations partially cancel each other out, and the total perturbation amplitude will be weakened. By adjusting the phase difference, the output perturbation signal can reach the highest possible amplitude, thereby making the peak of the cross-correlation function more prominent and the contrast between the main peak and the noise background higher.

[0057] Amplitude ratio is also a factor to consider. In actual dialysis, the fluid resistance, volume, and pump response characteristics on the blood and dialysate sides are not entirely consistent. Even if the amplitudes of the two disturbances are the same at input, the actual disturbance amplitudes transmitted to the semipermeable membrane inside the dialyzer will differ. For example, if one disturbance amplitude is much larger than the other near the semipermeable membrane, the coupling effect between them will be significantly weakened because the effect of a smaller amplitude on a larger amplitude is almost negligible. Only when the amplitudes of the two disturbances near the semipermeable membrane are comparable can the coupling be significant. Therefore, appropriately adjusting the amplitude ratio of the disturbances on both sides of the input to make the actual disturbance amplitudes transmitted to the semipermeable membrane as close as possible can enhance the coupling effect.

[0058] In practice, there is no strict restriction on which side is used to adjust the phase difference and amplitude ratio. In principle, the phase difference and amplitude ratio can be changed by adjusting either or both sides of the blood infusion pump and dialysate infusion pump. For safety and clinical feasibility considerations, the preferred approach is to primarily adjust the phase and amplitude through the dialysate side, while keeping the perturbation amplitude on the blood side within a small range or not adjusting it at all, using it only as a basic perturbation source.

[0059] The peak amplitude of the cross-correlation function can be used to measure the accuracy of time delay detection because when the proportion of the component related to the input disturbance in the output signal is higher and the coherence is stronger, the cross-correlation function curve will show a sharp main peak, a high peak value, and relatively low background noise. In this case, the time offset corresponding to the peak value is more stable and less affected by noise disturbances. Conversely, if the effective disturbance component in the output signal is weak and the noise proportion is high, the main peak of the cross-correlation function curve will not be obvious, the peak top will be flat, or even multiple approximate peak values ​​will appear, leading to jitter and misjudgment in the estimation of the time delay parameter. This invention adjusts the phase difference and amplitude ratio to make the effective disturbance signal at the blood output end generate the highest possible and single prominent main peak under cross-correlation analysis, thereby improving the discriminability and repeatability of the time delay parameter. Simply put, the higher the peak value and the sharper the peak shape, the clearer the correspondence between the useful signal and the disturbance, the more accurate the time delay estimation, and thus the higher the reliability of judging the filter membrane blockage trend.

[0060] As shown in Figure 3, in a further embodiment, in order to achieve automatic optimization of phase difference and amplitude ratio, prediction can be made based on regression model or reinforcement learning.

[0061] In one embodiment of the present invention, a regression model is introduced to characterize the mapping relationship between the combination of perturbation parameters and the peak amplitude of the cross-correlation signal for optimizing the phase difference and amplitude ratio of the perturbation signal. Based on this, the optimal combination of perturbation parameters is obtained by reverse search. Specifically, in multiple perturbation cycles, the system automatically changes the combination of phase difference and amplitude ratio. For example, the phase difference is taken sequentially within a fixed step size in the range of 0 to 360 degrees within one cycle, and the amplitude ratio is taken sequentially within a preset range with a fixed step size. Under the action of each set of phase difference and amplitude ratio, a perturbation is applied and the peak amplitude of the blood-side cross-correlation function is collected to form a sample dataset. Each sample includes at least the phase difference, amplitude ratio, and the corresponding measured peak amplitude of the cross-correlation. It can also additionally record the current system state parameters at the corresponding time, such as baseline blood flow rate, baseline blood pressure, baseline dialysate flow rate, or baseline dialysate pressure, to characterize the operating conditions of the dialysis process.

[0062] Using the above sample dataset, a regression model is trained to approximate the functional relationship between "phase difference, amplitude ratio, current system state and cross-correlation peak amplitude".

[0063] During the training phase, phase difference, amplitude ratio, and one or more current system state parameters are selected as input features. The corresponding cross-correlation peak amplitude is used as the regression target. The model is trained using support vector regression, random forest regression, or other supervised learning algorithms. Cross-validation or hold-out validation is used to evaluate the model performance and optimize the model hyperparameters. The goal is to minimize the error between the predicted peak amplitude and the actual measured peak amplitude. After training, a regression model is obtained that can predict the cross-correlation peak amplitude under given perturbation parameter combinations and current system state conditions.

[0064] In the actual detection phase, a pre-trained regression model is used to evaluate a set of candidate perturbation parameters, thereby indirectly obtaining the optimal parameter combination. Specifically, given the current system state parameters, the "phase difference, amplitude ratio, and current system state" are input point-by-point into the regression model on a preset discrete grid of phase difference and amplitude ratio. This yields the predicted peak amplitude for each set of candidate perturbation parameters. Then, the phase difference and amplitude ratio with the largest predicted peak amplitude are selected from all candidate combinations as the parameters used for the actual applied perturbation. In this way, the regression model always maintains a consistent direction of "input as perturbation parameters and system state, output as predicted peak amplitude," and the optimal phase difference and amplitude ratio are obtained by searching for the maximum peak value in the model output.

[0065] In another embodiment of the invention, to achieve adaptive optimization of the perturbation parameters, a reinforcement learning model is introduced, modeling the "selection process of phase difference and amplitude ratio" as a decision-making process based on reward feedback. Specifically, the combination of phase difference and amplitude ratio is regarded as the action of the agent, the operating conditions of the dialysis process are regarded as the environmental state, and the peak amplitude of the blood-side cross-correlation function is regarded as the reward signal.

[0066] In this implementation, the states, actions, and rewards in reinforcement learning are first clearly defined. The state characterizes the operating conditions of the current dialysis process and may include one or more system state variables such as baseline blood flow rate, baseline blood pressure, baseline dialysate flow rate, baseline dialysate pressure, and recent trends in time-delay parameters. These variables are not directly optimized but serve as reference information for the agent's decision-making. An action is a set of perturbation parameters, including at least phase difference and amplitude ratio. The action space can be set to discrete or continuous form as needed. The reward signal is defined as the peak amplitude of the blood-side cross-correlation function obtained through cross-correlation analysis under the action. A larger peak value results in a higher reward, guiding the agent to find the combination of perturbation parameters that maximizes the cross-correlation peak.

[0067] During the training phase, a reinforcement learning model consisting of a policy network and a value network is constructed. The policy network takes the current system state as input and outputs the probability distribution or expected action value of selecting different perturbation parameter combinations under a given state. The value network takes the current system state and the actions already performed as input and outputs the value assessment of the action pair for that state. During training, at the beginning of each perturbation cycle, the agent first collects the current system state and inputs it into the policy network. Based on the output of the policy network, it selects a set of phase difference and amplitude ratio as the perturbation parameters for this cycle and applies them to the blood infusion pump and dialysate infusion pump through the perturbation control module. Subsequently, the output signal is collected, cross-correlation analysis is performed, and the peak amplitude of the cross-correlation function on the blood side is calculated. This peak value is used as the reward signal for this cycle. The time-series difference error between this reward signal and the output of the value network is used to update the parameters of the value network. At the same time, a policy gradient-based method is used to update the parameters of the policy network to increase the probability of generating high-reward actions under the same state. Through continuous interaction and policy updates over multiple perturbation cycles, the policy network gradually converges, making the actions given under different system states tend to output phase difference and amplitude ratio combinations that can obtain a larger peak amplitude of cross-correlation.

[0068] The preceding stages are the exploration and training phases. In the actual testing phase, the reinforcement learning model no longer focuses on exploration but instead utilizes the converged policy to directly provide recommended perturbation parameter combinations for the given current system state. Specifically, at the start of the testing cycle, the system collects the current baseline blood flow, baseline blood pressure, and other selected system state variables, inputting them into the policy network. The policy network outputs a set of phase differences and amplitude ratios as the target parameters for this perturbation, which are applied to the pump control end through the perturbation control module. This allows for adaptive selection of perturbation parameters more suitable for the current operating conditions without manual intervention. Through this design, the reinforcement learning model maintains a consistent direction throughout the process: "input is the current system state, output is the perturbation parameter combination." The system uses the cross-correlation peak amplitude as a reward signal to drive policy convergence, ultimately achieving adaptive optimization for individual differences in different dialyzers and different operating states.

[0069] In a further embodiment, waveform selection can also be included in the optimization. Besides sine waves, square waves, with their sharp rising edges that can elicit more high-frequency responses, can be tried; or multiple single-pulse signals, such as three 0.1-second pulses per cycle, can be used to simulate transient impacts. The frequency can also be variable, for example, from 0.5 Hz to 1.5 Hz. During optimization, waveform type and frequency are used as additional parameters, along with phase difference and amplitude ratio, to maximize the peak amplitude of the blood-side cross-correlation.

[0070] For the selection of blood parameters, flow rate is measured using an ultrasonic flow meter, and blood pressure is acquired using a pressure sensor. Regardless of the method, an upper limit threshold is set for the disturbance amplitude; for example, the upper limit for flow rate disturbance is 10% of the base flow rate, and the upper limit for blood pressure is 8 mmHg. The system monitors the actual disturbance amplitude in real time and automatically reduces it if it approaches the upper limit.

[0071] Through the implementation of the above detailed steps, the entire system can continuously and non-invasively monitor blockage trends during dialysis. The calculation of time lag parameters accurately captures changes in internal resistance, regression analysis provides trend prediction, and two-sided monitoring and parameter optimization further enhance detection sensitivity and specificity. The adaptive capabilities of reinforcement learning enable the system to cope with aging dialysis machines or changes in patient coagulation status, ultimately achieving early warning and significantly improving dialysis safety and efficiency.

[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent perturbation and analysis of hemodialysis data, characterized in that, include: S1: During the blood input process into the dialysis machine, a disturbance signal is superimposed on the blood input pump to cause a preset configuration change in the blood parameters input into the dialysis machine, thereby forming an input disturbance signal; S2: During the period of disturbance application, blood parameters at the blood output end of the dialyzer are collected in real time to form a corresponding output response signal; S3: Perform cross-correlation analysis on the input disturbance signal and the output response signal, and calculate the time delay parameter Δτ corresponding to the disturbance signal between them; the cross-correlation analysis includes calculating the cross-correlation function between the input disturbance signal and the output response signal within a preset time window, and determining the time offset corresponding to the peak value of the cross-correlation function as the time delay parameter Δτ; S4: Execute steps S1~S3 once every preset time interval to obtain the time delay parameter Δτ that changes with time; S5: Perform regression analysis on the time delay parameter Δτ to obtain the time delay change trend; S6: If the regression analysis shows that the time lag parameter Δτ continues to increase over time, it is determined that the dialysis machine has a blockage trend and an early warning message is output; The method further includes: Perform the same perturbation and cross-correlation analysis steps as S1~S6 on the dialysate input to the dialyzer and the dialysis waste liquid output from the dialyzer to obtain the time delay parameters on the dialysate side and perform corresponding regression analysis and early warning. Simultaneously, perturbation signals are applied to the blood infusion pump and the dialysate infusion pump. The two perturbation signals are set to have the same frequency and waveform. The coupling relationship between the two perturbations is changed by adjusting the phase difference and amplitude ratio. The cross-correlation function of the blood side is calculated after each perturbation cycle. When the peak amplitude of the cross-correlation function of the blood side reaches its maximum, the corresponding phase difference and amplitude ratio are recorded as the phase difference and amplitude ratio used in actual detection.

2. The intelligent perturbation and analysis method for hemodialysis data according to claim 1, characterized in that, The cross-correlation analysis further includes filtering and normalizing the input disturbance signal and the output response signal respectively before calculating the cross-correlation function.

3. The intelligent perturbation and analysis method for hemodialysis data according to claim 1, characterized in that, The phase difference and amplitude ratio of the disturbance signal are optimized based on a regression model, specifically including: The peak amplitude of the cross-correlation function and the corresponding phase difference and amplitude ratio samples were collected during multiple perturbation cycles to construct a training dataset. A regression model was trained using the training dataset to establish a mapping relationship between the perturbation parameters and the peak amplitude of the blood-side cross-correlation. In actual testing, the phase difference and amplitude ratio that maximize the peak amplitude of the blood-side cross-correlation function are predicted by regression model.

4. The intelligent perturbation and analysis method for hemodialysis data according to claim 3, characterized in that, The phase difference and amplitude ratio of the perturbation signal are optimized through self-learning based on a reinforcement learning model, specifically including: The phase difference and amplitude ratio are used as action variables in reinforcement learning, and the peak amplitude of the cross-correlation function is set as the reward signal. The agent performs actions and collects output signals in each perturbation cycle, and updates the policy network parameters based on the reward signal. Through a multi-round perturbation cycle strategy iteration process, the algorithm gradually converges to the optimal phase difference and amplitude ratio that maximizes the peak amplitude of the blood cross-correlation function.

5. The intelligent perturbation and analysis method for hemodialysis data according to any one of claims 3 to 4, characterized in that... ； The waveform is a sine wave, a square wave, or multiple single-pulse signals, and the waveform or the frequency is included in the optimization parameter range that maximizes the peak amplitude of the blood-side cross-correlation function. The blood parameter is either a flow rate parameter or a blood pressure parameter, and a threshold upper limit is set for the amplitude of the disturbance applied to the flow rate parameter or blood pressure parameter.

6. The intelligent perturbation and analysis method for hemodialysis data according to claim 1, characterized in that, When performing regression analysis on the changing trend of the time delay parameter Δτ, the time delay parameter sequence obtained by multi-period sampling is used as the regression sample. The slope of the time delay parameter is calculated by the least squares method or linear regression algorithm. When the slope is positive and exceeds a preset threshold, it is determined that the time delay parameter Δτ continues to increase with time.

7. The intelligent perturbation and analysis method for hemodialysis data according to claim 1, characterized in that, The method uses a blood simulation kit to replace a real patient for testing.

8. An intelligent perturbation and analysis system for implementing the method of claim 1 in hemodialysis data, characterized in that, include: The disturbance control module is used to generate an input disturbance signal by superimposing a disturbance signal on the blood input pump during the blood input process of the dialyzer, thereby causing a preset configuration change in the blood parameters input to the dialyzer. The data acquisition module is used to collect blood parameters at the blood output end of the dialyzer in real time during the application of disturbance, and generate corresponding output response signals; The time delay calculation module is used to perform cross-correlation analysis on the input disturbance signal and the output response signal, and calculate the time delay parameter Δτ corresponding to the disturbance signal between the two. The periodic execution module is used to execute steps S1~S3 once at a preset time interval to obtain the time delay parameter Δτ that changes with time. The regression analysis module is used to perform regression analysis on the time delay parameter Δτ to obtain the time delay change trend; The blockage warning module is used to determine that the dialysis machine has a blockage trend when the regression analysis shows that the time lag parameter Δτ continues to increase over time, and to output warning information.

Images