Blood purification equipment real-time control method based on edge calculation

Abstract

The invention relates to the crossing field of computer technology and edge calculation, and discloses a blood purification equipment real-time control method based on edge calculation. According to the method, multi-modal real-time physiological parameters, historical dialysis data and a current treatment target are fused through a local edge calculation unit, an individualized body fluid dynamic response prediction model is constructed, and an ultrafiltration rate instruction is optimized and generated within a millisecond level; the system comprises a multi-mode sensor array, a historical data storage module, a task planning interface, an edge computing unit and an ultrafiltration pump controller. According to the application, the hemodynamic state is accurately maintained, the risk of hypotension is remarkably reduced, and the treatment safety and individualization level are improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer technology and edge computing, and specifically relates to a real-time control method for blood purification equipment based on edge computing. Background Technology

[0002] With the widespread application of blood purification technology in the treatment of end-stage renal disease, higher demands are placed on the precise control and individualized adaptation of the dialysis process. One of the core goals of hemodialysis is to effectively remove excess water from the patient's body through ultrafiltration while maintaining hemodynamic stability.

[0003] The distribution of bodily fluids in the human body exhibits highly nonlinear and time-varying characteristics, influenced by multiple factors such as cardiac function, vascular permeability, autonomic nervous system regulation, and changes in body position during dialysis, resulting in a complex dynamic volume status. Traditional ultrafiltration control strategies often employ open-loop settings or simple closed-loop feedback based on fixed parameters (such as using transmembrane pressure or venous pressure as a single input). These strategies lack the ability to perceive and model the patient's real-time physiological state, making it difficult to adapt to inter-individual and intra-individual differences. This can easily lead to complications such as hypotension, muscle cramps, or volume overload, seriously threatening treatment safety and comfort.

[0004] Real-time control methods based on edge computing have become a key direction for improving the intelligence level of blood purification equipment. This direction aims to push computing power down to the device level, enabling localized fusion analysis and rapid decision-making of multi-source physiological signals (such as blood pressure, heart rate variability, bioimpedance, and ultrafiltration rate) during dialysis, while ensuring data privacy and communication efficiency. Its basic principle is to utilize the limited but low-latency computing resources at the edge to deploy lightweight intelligent algorithms, conduct online assessments of the patient's volume status, and dynamically adjust the ultrafiltration rate and replacement fluid flow rate to maintain circulatory stability.

[0005] Existing technologies face multiple challenges in achieving these goals: On the one hand, while high-precision physiological models can effectively characterize fluid transfer mechanisms, their high computational complexity makes it difficult to meet the stringent real-time requirements of edge devices; on the other hand, simplified control strategies (such as proportional-integral-derivative control) while possessing good robustness lack the ability to adapt to nonlinear dynamics and individual specificities, thus failing to achieve truly intelligent regulation. Furthermore, existing systems generally lack the ability to continuously learn from historical treatment experience and optimize long-term control strategies, leading to rigid control logic. Especially when facing sudden hemodynamic fluctuations or special pathological conditions (such as heart failure combined with dialysis), the system response is often delayed or excessive, making it difficult to balance safety and dehydration efficiency.

[0006] Therefore, there is an urgent need for an edge-side real-time control architecture that integrates the advantages of offline learning and online execution, and is both intelligent and robust, to solve the problem of dynamic and precise regulation of volume management during blood purification. Summary of the Invention

[0007] This invention provides a real-time control method for blood purification equipment based on edge computing, aiming to solve the technical problems of traditional ultrafiltration control, which is mostly open-loop or simple closed-loop, making it difficult to cope with individual patient differences and complex fluid dynamics during dialysis, easily leading to hypotension or volume overload. The method deploys an edge computing unit locally on the blood purification equipment to fuse multi-source physiological parameters, historical dialysis data, and current treatment progress information in real time, constructing a dynamically adaptive ultrafiltration rate regulation model. It then completes online optimization and execution of the ultrafiltration strategy within a millisecond timescale, thereby achieving precise maintenance of the patient's hemodynamic state.

[0008] According to one aspect of the present invention, a real-time control method for a blood purification device based on edge computing is provided, comprising: acquiring a patient's real-time physiological parameter sequence through a multimodal physiological sensor array on the blood purification device; acquiring the patient's historical dialysis record data through a historical dialysis data storage module built into the device; acquiring the target ultrafiltration volume, target treatment duration, and preset safety constraint boundaries of the current dialysis treatment through a task planning interface; and inputting the real-time physiological parameter sequence, historical dialysis record data, and current dialysis treatment parameters to an edge computing unit deployed locally on the device. In the edge computing unit, temporal features are extracted from the real-time physiological parameter sequence to generate a dynamic representation of the current physiological state; individualized baseline modeling is performed on the historical dialysis record data to generate a patient-specific fluid regulation capacity benchmark representation; the current physiological state dynamic representation and the patient-specific fluid regulation capacity benchmark representation are fused across time periods to generate an individualized fluid dynamic response prediction representation; based on the individualized fluid dynamic response prediction representation and the current dialysis treatment parameters, an ultrafiltration rate command for the next control cycle is generated through an online ultrafiltration rate optimization solver; the ultrafiltration rate command is sent to the ultrafiltration pump drive controller to adjust the operating parameters of the ultrafiltration pump in real time.

[0009] In one embodiment of the present invention, the multimodal physiological sensor array includes an arterial pressure sensor, a venous pressure sensor, a transmembrane pressure sensor, a blood flow meter, a heart rate monitoring module, a blood oxygen saturation monitoring module, and a bioimpedance analysis module. The arterial pressure sensor is used to acquire arterial pressure signals, the venous pressure sensor is used to acquire venous pressure signals, the transmembrane pressure sensor is used to acquire pressure difference signals across the dialyzer, the blood flow meter is used to acquire blood flow velocity signals in the extracorporeal circulation loop, the heart rate monitoring module is used to acquire the patient's heart rate variability signals, the blood oxygen saturation monitoring module is used to acquire arterial blood oxygen saturation signals, and the bioimpedance analysis module is used to acquire bioelectrical impedance spectrum signals from the patient's chest cavity or limbs.

[0010] In one embodiment of the present invention, the time sampling frequency of the real-time physiological parameter sequence is not less than 10 times per second, and the data processing latency of the edge computing unit does not exceed 50 milliseconds. The temporal feature extraction adopts a one-dimensional convolutional neural network structure, which contains three convolutional layers. The number of convolutional kernels in each convolutional layer is 32, 64, and 128 respectively, and the kernel size is 5. The activation function is a modified linear unit, followed by a max pooling layer with a pooling window size of 2.

[0011] In one embodiment of the present invention, the historical dialysis record data includes the total ultrafiltration volume, actual treatment duration, mean arterial pressure change curve, heart rate fluctuation curve, whether a hypotensive event occurred, the time and duration of the hypotensive event, and the weight change at the end of each treatment in the past 12 dialysis treatments. The individualized baseline modeling uses a Gaussian process regression model, with the ultrafiltration rate in historical treatments as the input variable and the slope of arterial pressure decrease as the output variable, to establish a patient's sensitivity response function to ultrafiltration load.

[0012] As one embodiment of the present invention, the cross-time feature fusion is achieved through an attention mechanism, wherein the query vector is generated by the dynamic representation of the current physiological state, the key vector and value vector are generated by the patient-specific fluid regulation ability benchmark, and the attention weight is used to weight the fusion of historical response patterns and the current real-time state, and the output is a fusion representation vector containing individualized dynamic prediction capabilities.

[0013] In one embodiment of the present invention, the safety constraint boundaries include a lower limit threshold for arterial pressure, an upper limit threshold for venous pressure, a threshold for transmembrane pressure change rate, a lower limit threshold for heart rate coefficient of variation, and a lower limit threshold for blood oxygen saturation. The online ultrafiltration rate optimization solver employs a model predictive control framework, with a prediction time domain of the next 30 seconds and a control time domain of the next 5 seconds. The objective function is to minimize the sum of squares of the deviations between the predicted arterial pressure and the target arterial pressure, while simultaneously satisfying the hard constraints of all safety constraint boundaries. The optimization variable is the ultrafiltration rate setpoint per second within the next 5 seconds.

[0014] In one embodiment of the present invention, after receiving the ultrafiltration rate command, the ultrafiltration pump drive controller converts it into a pulse width modulation signal to drive a DC brushless motor to precisely control the speed of the ultrafiltration pump, thereby realizing real-time adjustment of the ultrafiltration liquid volume flow rate with an adjustment accuracy of ±5 ml per minute.

[0015] According to another aspect of the present invention, a real-time control system for a blood purification device based on edge computing is provided, comprising: a multimodal physiological sensor array for real-time acquisition of a patient's real-time physiological parameter sequence; a historical dialysis data storage module for storing and providing the patient's historical dialysis record data; a task planning interface for receiving the target ultrafiltration volume, target treatment duration, and preset safety constraint boundaries of the current dialysis treatment; an edge computing unit deployed locally on the blood purification device for performing data fusion, feature extraction, individualized modeling, and ultrafiltration rate optimization calculation; and an ultrafiltration pump drive controller for receiving ultrafiltration rate commands and controlling the operating parameters of the ultrafiltration pump.

[0016] As one embodiment of the present invention, the edge computing unit includes a dedicated neural network acceleration chip with a computing power of no less than 1 trillion integer operations per second, a memory bandwidth of no less than 20 gigabytes per second, supports floating-point and fixed-point mixed precision calculations, and has a hardware-level security encryption module to ensure the privacy and integrity of patient data during local processing.

[0017] As one embodiment of the present invention, the system further includes an emergency handling module for abnormal states. When any physiological parameter is detected to exceed the safety constraint boundary and the duration exceeds 2 seconds, an emergency reduction command for the ultrafiltration rate is immediately triggered to reduce the ultrafiltration rate to 0, and an audible and visual alarm device is activated. At the same time, the abnormal event is recorded and written to an immutable log storage area.

[0018] As one embodiment of the present invention, the bioimpedance analysis module uses a four-electrode method to apply an alternating current excitation current in the frequency range of 5000 Hz to 500 kHz, measure the voltage response, and calculate the ratio of extracellular fluid to total body fluid to assess the patient's current volume status. This ratio value serves as a key input feature in the generation of a dynamic representation of the current physiological state.

[0019] As one embodiment of the present invention, the individualized fluid dynamic response prediction characterization has 256 dimensions, each of which corresponds to a trained feature channel, characterizing the expected response intensity and direction of the patient to ultrafiltration intervention under specific physiological conditions.

[0020] As one embodiment of the present invention, the system dynamic model in the model predictive control framework is described by a nonlinear state-space equation. The state variables include arterial pressure, venous pressure, and estimated effective circulating blood volume. The control input is the ultrafiltration rate, and the output is a measurable physiological parameter. The model parameters are updated in real time using an online recursive least squares method, and the forgetting factor is set to 0.98.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves millisecond-level perception of the patient's physiological state and real-time closed-loop optimization of the ultrafiltration strategy by integrating edge computing capabilities locally into the blood purification device. This is in contrast to traditional open-loop control or simple closed-loop systems that rely solely on single blood pressure feedback.

[0022] 2. This invention integrates multi-dimensional real-time physiological signals, individualized historical treatment data, and strict physical safety constraints to construct a patient-specific dynamic prediction model. This dynamic prediction model can predict the impact of fluid removal on hemodynamics in advance and proactively adjust the ultrafiltration rate before hypotension events occur, thereby significantly reducing the incidence of hypotension during treatment.

[0023] 3. This invention avoids the risk of capacity overload due to insufficient ultrafiltration by precisely controlling the total ultrafiltration volume and rate. The entire control process is completed at the device end, without relying on cloud communication, ensuring the system's real-time performance, reliability, and data security. It is suitable for various clinical scenarios, and has outstanding clinical value, especially for elderly or critically ill patients with unstable cardiovascular function. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the individualized body fluid dynamic response prediction and characterization generation and online optimization solution of ultrafiltration rate in this invention; Figure 3 This is a logical flowchart of the multimodal physiological parameter acquisition and dynamic representation of the current physiological state in this invention; Figure 4 This is a logical flowchart of the individualized baseline modeling and cross-time period feature fusion of historical dialysis data in this invention; Figure 5 This is a flowchart illustrating the logical flow of online optimization of ultrafiltration rate and handling of safety constraints under the model predictive control framework of this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of the edge computing unit, ultrafiltration pump drive controller and abnormal state emergency handling module in this invention. Detailed Implementation

[0025] This invention provides a real-time control method for blood purification equipment based on edge computing, aiming to solve the technical problems of traditional ultrafiltration control, which is mostly open-loop or simple closed-loop, making it difficult to cope with individual patient differences and complex fluid dynamics during dialysis, easily leading to hypotension or volume overload. The method deploys an edge computing unit locally on the blood purification equipment, fusing multi-source physiological parameters, historical dialysis data, and current treatment progress information in real time to construct a dynamically adaptive ultrafiltration rate regulation model. It then completes online optimization and execution of the ultrafiltration strategy within a millisecond timescale, thereby achieving precise maintenance of the patient's hemodynamic state. The following will be combined with the appendix... Figures 1 to 5 The specific implementation methods of each component of the system are described in detail.

[0026] The method includes the following steps: S1, real-time physiological parameter sequences of patients are collected in real time through the multimodal physiological sensor array on the blood purification device; S2, obtains the historical dialysis record data of the dynamically predicted patient through the device's built-in historical dialysis data storage module; S3 obtains the target ultrafiltration volume, target treatment duration, and preset safety constraint boundaries for the current dialysis treatment through the task planning interface; S4, input the real-time physiological parameter sequence, historical dialysis record data and current dialysis treatment parameters to the edge computing unit deployed locally on the device; S5, In the edge computing unit, the time-series features of the real-time physiological parameter sequence are extracted to generate a dynamic representation of the current physiological state; S6, Perform individualized baseline modeling on the historical dialysis record data to generate a patient-specific baseline characterization of fluid regulation capacity; S7, The current physiological state dynamic representation and the patient-specific body fluid regulation capacity benchmark representation are fused across time periods to generate an individualized body fluid dynamic response prediction representation. S8, Based on the individualized dynamic response prediction characterization of body fluids and the current dialysis treatment parameters, the ultrafiltration rate command for the next control cycle is generated through the online ultrafiltration rate optimization solver; S9, the ultrafiltration rate command is sent to the ultrafiltration pump drive controller to adjust the operating parameters of the ultrafiltration pump in real time.

[0027] In step S1, the multimodal physiological sensor array includes an arterial pressure sensor, a venous pressure sensor, a transmembrane pressure sensor, a blood flow meter, a heart rate monitoring module, a blood oxygen saturation monitoring module, and a bioimpedance analysis module. The arterial pressure sensor is used to acquire arterial pressure signals, the venous pressure sensor is used to acquire venous pressure signals, the transmembrane pressure sensor is used to acquire pressure difference signals across the dialyzer, the blood flow meter is used to acquire blood flow velocity signals in the extracorporeal circulation loop, the heart rate monitoring module is used to acquire the patient's heart rate variability signals, the blood oxygen saturation monitoring module is used to acquire arterial blood oxygen saturation signals, and the bioimpedance analysis module is used to acquire bioelectrical impedance spectrum signals from the patient's chest cavity or limbs.

[0028] All sensor data outputs are synchronized to a unified time reference, with a sampling frequency of no less than 10 times per second to ensure strict alignment of physiological parameters on the time axis. Arterial and venous pressure signals are measured in millimeters of mercury (mmHg) with an accuracy of ±0.5 mmHg; transmembrane pressure signals are measured in kilopascals (kPa) with an accuracy of ±0.1 Pa; blood flow signals are measured in milliliters per minute (mL / min) with an accuracy of ±2 mL / min; heart rate variability signals are recorded in milliseconds to indicate the interval between adjacent R waves; blood oxygen saturation signals are expressed as a percentage with a resolution of 0.1%; the bioimpedance analysis module uses a four-electrode method, applying an alternating current excitation current within a frequency range of 5000 Hz to 500 kHz, measuring the voltage response, and calculating the ratio of extracellular fluid to total body fluid. This ratio serves as a key input feature in the generation of a dynamic representation of the current physiological state.

[0029] In step S2, the historical dialysis record data is stored in the device's local non-volatile storage medium, including the total ultrafiltration volume, actual treatment duration, mean arterial pressure change curve, heart rate fluctuation curve, whether a hypotension event occurred, the time and duration of the hypotension event, and the weight change at the end of each treatment for the past 12 dialysis treatments. All historical data is timestamped and outlier removed to ensure data quality. The historical records are organized in a structured database format, with each treatment session corresponding to an independent data entry, including time, operator identifier, device model, patient unique identifier, and the aforementioned physiological and treatment parameters. When a patient uses this device for dialysis for the first time, if there is no historical record, the system initializes to an empty baseline model, and after the treatment, the complete data is stored in the historical database for individualized modeling of subsequent treatments.

[0030] In step S3, the current dialysis treatment parameters are input by clinical medical staff through the human-computer interaction interface, and received and verified by the task planning interface. The target ultrafiltration volume is in liters, ranging from 0.5 to 6 liters; the target treatment duration is in minutes, ranging from 120 to 300 minutes; safety constraints include lower arterial pressure threshold, upper venous pressure threshold, transmembrane pressure rate of change threshold, lower heart rate coefficient of variation threshold, and lower oxygen saturation threshold. The lower arterial pressure threshold is set to 90 mmHg, the upper venous pressure threshold to 250 mmHg, the transmembrane pressure rate of change threshold to 10 kPa per minute, the lower heart rate coefficient of variation threshold to 15%, and the lower oxygen saturation threshold to 92%. All parameters undergo validity checks after input; values ​​exceeding clinically reasonable ranges will trigger a rejection mechanism and prompt for re-entry.

[0031] In step S4, the edge computing unit is a dedicated computing module integrated within the blood purification device. Its hardware platform includes a central processing unit, a neural network acceleration chip, high-speed memory, and a secure encryption coprocessor. After startup, the edge computing unit loads a pre-trained deep learning model and optimization solver program, and establishes communication channels with various sensors, storage modules, and controllers. All input data is normalized before entering the computing unit, mapped to the 0-1 range to eliminate dimensional differences. Data transmission employs a timestamp-based queue buffering mechanism to ensure that high-priority physiological signals are processed first, with a delay not exceeding 50 milliseconds.

[0032] In step S5, temporal features are extracted from the real-time physiological parameter sequence to generate a dynamic representation of the current physiological state. This process is implemented using a one-dimensional convolutional neural network structure, which contains three convolutional layers. The number of convolutional kernels in each layer is 32, 64, and 128 respectively, and the kernel size is 5. The activation function is a modified linear unit, followed by a max pooling layer with a pooling window size of 2. The input is a seven-channel physiological signal temporal vector, with each channel containing 500 sampling points, corresponding to the data from the most recent 50 seconds. The first convolutional layer extracts local patterns such as the slope of the instantaneous drop in blood pressure and sudden changes in heart rate; the second convolutional layer combines local patterns to form intermediate features such as circulatory stability indices; and the third convolutional layer generates a high-level abstract representation. The final output is a 128-dimensional dynamic representation vector of the current physiological state, with each dimension corresponding to a trained feature channel, representing different aspects of the patient's current hemodynamic stability.

[0033] In step S6, individualized baseline modeling is performed on the historical dialysis records to generate a patient-specific baseline representation of fluid regulation capacity. This individualized baseline modeling employs a Gaussian process regression model, using the ultrafiltration rate during historical treatments as the input variable and the arterial pressure drop slope as the output variable to establish a patient-specific sensitivity response function to ultrafiltration load. The covariance function of the Gaussian process regression uses radial basis functions, and its length scale and signal variance are learned from historical data through maximum likelihood estimation. For each historical treatment, the system extracts a piecewise constant sequence of ultrafiltration rate and its corresponding first derivative sequence of arterial pressure to construct a training sample set. The model output is a probability distribution, where the mean function reflects the expected arterial pressure change trend at a given ultrafiltration rate, and the variance function reflects the uncertainty of the individual response. This model is encoded as a 256-dimensional baseline representation vector of fluid regulation capacity, where the first 128 dimensions are the sampled values ​​of the mean function on a standard ultrafiltration rate grid, and the last 128 dimensions are the square roots of the corresponding variance points.

[0034] In step S7, the current physiological state dynamic representation and the patient-specific fluid regulation capacity benchmark representation are fused across time periods to generate a personalized fluid dynamic response prediction representation. This patient-specific fluid regulation capacity benchmark representation undergoes cross-time period feature fusion via an attention mechanism. The query vector is obtained from the current physiological state dynamic representation through a fully connected layer, while the key and value vectors are obtained from the fluid regulation capacity benchmark representation through another fully connected layer. The attention weight calculation formula is as follows: ,in The value matrix is ​​used. This attention mechanism enables the system to dynamically weight historical response patterns based on the current real-time state. For example, when the current heart rate variability coefficient is low, the system assigns higher weights to responses related to historical hypotension events. The final output is a 256-dimensional personalized fluid dynamic response prediction representation vector, where each dimension corresponds to a trained feature channel, representing the expected intensity and direction of the patient's response to ultrafiltration intervention under specific physiological conditions.

[0035] In step S8, based on the individualized dynamic response prediction characterization of body fluids and the current dialysis treatment parameters, an online ultrafiltration rate optimization solver is used to generate the ultrafiltration rate command for the next control cycle. This online ultrafiltration rate optimization solver employs a model predictive control framework, with a prediction time domain of the next 30 seconds and a control time domain of the next 5 seconds. The objective function is to minimize the sum of squares of the deviations between the predicted arterial pressure and the target arterial pressure, while simultaneously satisfying all hard constraints of the safety boundary. The system dynamic model is described using nonlinear state-space equations. State variables include arterial pressure, venous pressure, and an estimated effective circulating blood volume. The control input is the ultrafiltration rate, and the output is measurable physiological parameters. Model parameters are updated in real-time using an online recursive least squares method, with a forgetting factor set to 0.98. The objective function expression is: in, For the first Predict arterial pressure in seconds Target arterial pressure, For the first Ultrafiltration rate per second To smooth out the penalty coefficient, the optimization variable is the ultrafiltration rate setpoint per second for the next 5 seconds. Constraints include arterial pressure not falling below the lower threshold, venous pressure not exceeding the upper threshold, transmembrane pressure change rate not exceeding the threshold, heart rate coefficient of variation not falling below the lower limit, and blood oxygen saturation not falling below the lower limit. The solver uses a sequential quadratic programming algorithm to complete the optimization calculation within 50 milliseconds.

[0036] In step S9, the ultrafiltration rate command is sent to the ultrafiltration pump drive controller. Upon receiving the ultrafiltration rate command in milliliters per minute (mL / min), the ultrafiltration pump drive controller converts it into a pulse-width modulation (PWM) signal to drive a brushless DC motor for precise control of the ultrafiltration pump's rotational speed. The duty cycle of the PWM signal is linearly related to the ultrafiltration rate and is calibrated and stored in the controller firmware. The ultrafiltration pump employs a volumetric design, discharging a fixed volume of liquid per revolution, with a rotational speed control accuracy of ±5 mL / min. The controller reads the command and updates its output every 10 milliseconds to ensure dynamic response performance.

[0037] The system also includes an emergency response module for abnormal states. This module continuously monitors all physiological parameters to ensure they do not exceed safety constraints. If any parameter exceeds the limit for more than 2 seconds, an emergency reduction command for the ultrafiltration rate is immediately triggered, lowering the ultrafiltration rate to 0 and activating an audible and visual alarm. Simultaneously, the timestamp of the abnormal event, the name of the parameter exceeding the limit, the value exceeding the limit, the duration, and a snapshot of all physiological parameters at that time are written to an immutable log storage area. The log uses a hash chain structure, with each record containing the hash value of the previous record, ensuring data integrity.

[0038] The edge computing unit includes a dedicated neural network acceleration chip with a computing power of no less than 1 trillion integer operations per second and a memory bandwidth of no less than 20 gigabytes per second, supporting mixed-precision floating-point and fixed-point calculations. This dedicated neural network acceleration chip incorporates a hardware-level security encryption module that encrypts all patient data in memory in real time. The key is generated jointly by the device's unique identifier and the patient's biometric characteristics, ensuring the privacy and integrity of the data during local processing.

[0039] The entire control loop executes 20 times per second, with an end-to-end latency of no more than 50 milliseconds from data acquisition to ultrafiltration pump adjustment. This method is entirely performed on the device side, without relying on cloud communication, ensuring the system's real-time performance, reliability, and data security. It is suitable for various clinical scenarios, and has outstanding clinical value, especially for elderly or critically ill patients with unstable cardiovascular function.

Claims

1. A real-time control method for a blood purification device based on edge computing, characterized in that, include: The patient's real-time physiological parameter sequence is acquired in real time through a multimodal physiological sensor array on the blood purification device. The real-time physiological parameter sequence includes arterial pressure signal, venous pressure signal, transmembrane pressure signal, blood flow signal, heart rate variability signal, blood oxygen saturation signal, and bioelectrical impedance spectrum signal. The device acquires the real-time historical dialysis record data of the patient through its built-in historical dialysis data storage module. The target ultrafiltration volume, target treatment duration, and preset safety constraint boundaries for the current dialysis treatment can be obtained through the task planning interface. The real-time physiological parameter sequence, historical dialysis record data, and current dialysis treatment parameters are input to the edge computing unit deployed locally on the device. In the edge computing unit, temporal features are extracted from the real-time physiological parameter sequence to generate a dynamic representation of the current physiological state; Individualized baseline modeling was performed on the historical dialysis record data to generate a patient-specific baseline characterization of fluid regulation capacity; The current physiological state dynamic representation is fused with the patient-specific fluid regulation capacity benchmark representation across time periods to generate a personalized fluid dynamic response prediction representation. Based on the individualized dynamic response prediction characterization of body fluids and the current dialysis treatment parameters, the ultrafiltration rate command for the next control cycle is generated through the online ultrafiltration rate optimization solver. The ultrafiltration rate command is sent to the ultrafiltration pump drive controller to adjust the operating parameters of the ultrafiltration pump in real time.

2. The real-time control method for blood purification equipment based on edge computing according to claim 1, characterized in that, Extracting temporal features from the real-time physiological parameter sequence to generate a dynamic representation of the current physiological state includes: The real-time physiological parameter sequence is organized into a seven-channel time-series vector according to the time dimension, with each channel containing data from no less than 500 sampling points within the most recent 50 seconds; The seven-channel temporal vector is input into a one-dimensional convolutional neural network. The one-dimensional convolutional neural network contains three sequentially connected convolutional layers. The number of convolutional kernels in each convolutional layer is 32, 64 and 128 respectively. The kernel size is 5. The activation function is a modified linear unit. Each convolutional layer is followed by a max pooling layer with a pooling window size of 2. Output a 128-dimensional dynamic representation vector of the current physiological state as the dynamic representation of the current physiological state.

3. The real-time control method for blood purification equipment based on edge computing according to claim 2, characterized in that, Individualized baseline modeling was performed on the historical dialysis records to generate patient-specific baseline representations of fluid regulation capabilities, including: A Gaussian process regression model was constructed using the ultrafiltration rate from historical dialysis records as the input variable and the slope of arterial pressure drop as the output variable. Radial basis functions are used as covariance functions, and length scale and signal variance parameters are learned from historical data through maximum likelihood estimation; The sampled values ​​of the mean function of the Gaussian process regression model on the standard ultrafiltration rate grid and the square roots of the corresponding variance points are encoded into the first 128 dimensions and the last 128 dimensions, respectively, forming a 256-dimensional humoral regulation capacity benchmark representation vector as the patient-specific humoral regulation capacity benchmark representation.

4. The real-time control method for blood purification equipment based on edge computing according to claim 3, characterized in that, The current physiological state dynamic representation is fused with the patient-specific humoral regulation capacity benchmark representation across time periods to generate a personalized humoral dynamic response prediction representation, including: The dynamic representation of the current physiological state is transformed into a query vector through the first fully connected layer; The benchmark characterization of body fluid regulation capability is transformed into a bond vector and a value vector through a second fully connected layer; Based on the query vector, key vector, and value vector, attention weights are calculated using an attention mechanism to generate a fusion output. The fusion output is used as a 256-dimensional individualized fluid dynamic response prediction representation vector as the individualized fluid dynamic response prediction representation.

5. The real-time control method for blood purification equipment based on edge computing according to claim 4, characterized in that, The formula for calculating the attention weight of the attention mechanism is as follows: ,in It is a value matrix.

6. The real-time control method for blood purification equipment based on edge computing according to claim 5, characterized in that, Based on the individualized dynamic response prediction characterization of body fluids and the current dialysis treatment parameters, an online ultrafiltration rate optimization solver is used to generate the ultrafiltration rate command for the next control cycle, including: A model predictive control framework is adopted, with the prediction time domain set to the next 30 seconds and the control time domain set to the next 5 seconds. A nonlinear state-space equation is constructed to describe the dynamic model of the system. The state variables include arterial pressure, venous pressure and estimated effective circulating blood volume. The control input is the ultrafiltration rate and the output is a measurable physiological parameter. The objective function is to minimize the sum of squares of the deviations between the predicted arterial pressure and the target arterial pressure, and hard constraints are imposed on all safety boundary conditions. The ultrafiltration rate setting value per second for the next 5 seconds is obtained by using a sequential quadratic programming algorithm and is used as the ultrafiltration rate command.

7. The real-time control method for blood purification equipment based on edge computing according to claim 6, characterized in that, The objective function expression is: in, For the first Predict arterial pressure in seconds Target arterial pressure, For the first Ultrafiltration rate per second To smooth out the penalty coefficient.

8. The real-time control method for blood purification equipment based on edge computing according to claim 7, characterized in that, The model parameters of the system dynamic model are updated in real time using the online recursive least squares method, and the forgetting factor is set to 0.

98.

9. The real-time control method for blood purification equipment based on edge computing according to claim 8, characterized in that, The ultrafiltration rate command is sent to the ultrafiltration pump drive controller to adjust the operating parameters of the ultrafiltration pump in real time, including: The ultrafiltration rate command is converted into a pulse width modulation signal; A DC brushless motor is driven to control the speed of the ultrafiltration pump, so that the adjustment accuracy of the ultrafiltration liquid volume flow rate can reach ±5 ml per minute. The ultrafiltration rate command is read every 10 milliseconds and the pulse width modulation signal output is updated.

10. The real-time control method for blood purification equipment based on edge computing according to claim 9, characterized in that, The bioelectrical impedance spectrum signal is obtained by applying an alternating current in the frequency range of 5000 Hz to 500 kHz and measuring the voltage response using a four-electrode method. This signal is used to calculate the ratio of extracellular fluid to total body fluid, and this ratio serves as a key input feature in the generation of dynamic representations of the current physiological state.

Images