High-precision pressure automatic regulation system for peritoneal dialysis based on machine learning

Abstract

This invention relates to the field of medical information technology, specifically to a high-precision automatic pressure control system for peritoneal dialysis based on machine learning. The system's state perception unit determines the estimated glucose concentration based on the initial glucose concentration and retention time; determines the instantaneous osmosis rate based on the estimated glucose concentration and body temperature; determines the volume-pressure response coefficient based on the instantaneous osmosis rate and fluid pressure, and determines the volume-pressure state index in conjunction with a preset reference peritoneal volume; determines the action risk index in real time based on inertial measurement data and the volume-pressure state index; during the drainage phase, it determines the drainage efficiency index based on the real-time drainage rate and the preset reference drainage rate; and a prediction unit fuses the determined indicators through a pre-trained correlation matrix to determine the dynamic weights of each sensor data point. The sensor data is then weighted using these dynamic weights and input into a time-series training model to obtain a pressure prediction value, thereby improving the accuracy of pressure prediction.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a high-precision automatic pressure control system for peritoneal dialysis based on machine learning. Background Technology

[0002] Peritoneal dialysis involves infusing dialysate into the patient's peritoneal cavity, using the peritoneum as a semipermeable membrane to remove metabolic waste and excess water. During treatment, dynamic monitoring and stable control of intraperitoneal pressure are crucial. Too low a pressure may affect dialysis efficiency, while too high a pressure can lead to abdominal distension and pain, and significantly increase the risk of complications such as hernias and dialysate leakage, severely impacting treatment safety and patient comfort.

[0003] Currently, pressure regulation in existing peritoneal dialysis equipment largely relies on traditional control methods, such as on / off control based on a fixed pressure threshold or classic proportional-integral-derivative (PID) control. These methods primarily monitor intra-abdominal pressure using a single pressure sensor, and when the pressure exceeds a preset safety threshold, they take lagging measures such as pausing perfusion or reducing the flow rate. However, intra-abdominal pressure is a complex variable influenced by various dynamic factors, including but not limited to the osmotic pressure gradient of the dialysate, the patient's individualized peritoneal transport characteristics, real-time changes in intra-abdominal volume, changes in patient position, and physical activity. Traditional control methods struggle to comprehensively and accurately capture and respond to these complex time-varying characteristics, leading to delayed regulatory responses, frequent misjudgments, and an inability to achieve proactive and precise control. Summary of the Invention

[0004] To address the technical problem of delayed response and frequent misjudgments caused by monitoring intra-abdominal pressure with a single pressure sensor, which fails to consider objective factors such as the osmotic pressure gradient of the dialysate, the individualized peritoneal transport characteristics of the patient, real-time changes in abdominal volume, changes in patient position, and physical activity, this invention provides a high-precision automatic pressure control system for peritoneal dialysis based on machine learning. The specific technical solution adopted is as follows: This invention proposes a high-precision automatic pressure control system for peritoneal dialysis based on machine learning. The system includes: The acquisition unit is used to acquire multi-source sensor data and initial glucose concentration of dialysate in real time during peritoneal treatment. The sensor data includes peritoneal fluid pressure, retention time, drainage rate, individual body temperature, and inertial measurement data. The state sensing unit is used to determine the estimated glucose concentration in real time based on the initial glucose concentration and retention time; determine the instantaneous osmosis rate in real time based on the estimated glucose concentration and body temperature; determine the volume-pressure response coefficient based on the instantaneous osmosis rate and fluid pressure, and determine the volume-pressure state index in combination with the preset reference abdominal cavity volume; determine the action risk index in real time based on inertial measurement data and volume-pressure state index; and determine the drainage efficiency index in the drainage stage based on the real-time drainage rate and the preset reference drainage rate. The prediction unit is used to determine the dynamic weights of each sensor data at each moment of the current working stage of the system by fusing instantaneous permeation rate, volume-pressure state index, action risk index and drainage efficiency index through a pre-trained correlation matrix. After weighting the sensor data with dynamic weights, the data is input into the pre-trained time series training model to obtain the pressure prediction value. The control unit is used to adjust the operating parameters of the peritoneal dialysis machine based on the predicted pressure value.

[0005] Furthermore, the process for determining the estimated glucose concentration includes: Based on the initial glucose concentration of the dialysate and the current retention time of the dialysate, the estimated glucose concentration at the current moment is calculated using a preset glucose concentration decay model. The glucose concentration decay model decreases exponentially with the increase of retention time.

[0006] Furthermore, the instantaneous permeation rate determination process includes: Obtain individual peritoneal transport characteristics, intraperitoneal reference body temperature, and blood reference blood glucose concentration; Calculate the difference between the initial glucose concentration and the reference blood glucose concentration at each time point during the treatment process, and use it as the concentration difference. The absolute difference between the individual's body temperature and the abdominal reference body temperature at each moment during treatment is calculated as the temperature difference; the product of the preset empirical coefficient and the temperature difference is calculated as the temperature compensation coefficient at each moment. The ratio of the concentration difference to the peritoneal transport characteristic parameter is calculated as the first ratio; the product of the first ratio and the temperature compensation coefficient is calculated as the instantaneous permeation rate at each time point.

[0007] Furthermore, the process for determining the volumetric-pressure response coefficient includes: Obtain the initial perfusion volume of the dialysate; Starting from the beginning of treatment, the time interval between each moment and the previous moment is recorded; the product of the instantaneous osmosis rate at each moment and the corresponding time interval is calculated as the volume increment; all volume increments before each moment are accumulated to obtain the volume change caused by osmosis from the beginning of treatment to each moment. The volume change at each time point is added to the initial perfusion volume to obtain the estimated abdominal cavity volume. Based on the estimated abdominal cavity volume and the corresponding fluid pressure at each time point, the volume-pressure response coefficient is determined.

[0008] Furthermore, determining the volume-pressure response coefficient based on the estimated abdominal cavity volume and fluid pressure at each time point includes: The estimated abdominal cavity volume and corresponding fluid pressure at each time point were statistically analyzed to obtain a volume-pressure dataset. Based on the volume-pressure dataset, a curve fitting algorithm is used to establish the volume-pressure relationship curve for each individual. Calculate the slope of the volume-pressure relationship curve at each time point, and use the statistical value of the slope as the volume-pressure response coefficient.

[0009] Furthermore, the process for determining the volumetric-pressure state index includes: Obtain the preset reference abdominal cavity volume; Calculate the square of the preset abdominal cavity reference volume and use it as the volume square value; Calculate the ratio of the estimated abdominal cavity volume to the square of the volume at each time point, and use it as the second ratio; The product of the second ratio and the volumetric-pressure response coefficient is calculated and used as the volumetric-pressure state index.

[0010] Furthermore, the inertial measurement data includes acceleration data; the process for determining the motion risk index includes: Based on inertial measurement data, the attitude change at each moment during the treatment process is calculated using an attitude calculation algorithm; High-pass filtering is applied to the acceleration data to separate the high-frequency components; The norm of the posture change is calculated and normalized to obtain the intensity of the posture change; the variance of the high-frequency component is normalized to obtain the high-frequency impact intensity. Based on preset weights, the intensity of body position change and the intensity of high-frequency impact are weighted and fused to obtain the original disturbance intensity; Calculate the product of the original disturbance intensity and the volume-pressure state index as the action risk index.

[0011] Furthermore, the process for determining the diversion efficiency index includes: Obtain the preset minimum constant; The comprehensive drainage rate is obtained by adding the real-time drainage rate at each moment during the treatment process to a preset minimum constant. The ratio of the preset reference diversion rate to the comprehensive diversion rate is calculated and used as a diversion efficiency indicator.

[0012] Furthermore, the dynamic weights of each sensor data are determined by fusing instantaneous penetration rate, volume-pressure state index, action risk index, and drainage efficiency index through a pre-trained correlation matrix, including: Based on the current working stage of the system, a binary gating vector is generated. The binary gating vector is used to select effective indicators from instantaneous permeation rate, volume-pressure state indicators, action risk indicators and drainage efficiency indicators to form an effective feature vector. The binary gating vector is used to disable sensor data channels that have unclear physical meaning or are invalid in a specific working stage. The effective feature vectors are input into the pre-trained correlation matrix to obtain the dynamic weights of each sensor data channel. The pre-trained correlation matrix is ​​learned through end-to-end training.

[0013] Furthermore, the step of weighting the sensor data using dynamic weights and then inputting it into the pre-trained time-series training model to obtain the pressure prediction value includes: The dynamic weights are then weighted and fused with the corresponding original sensor data to form weighted feature data; Weighted feature data is input into a pre-trained time series training model to obtain stress prediction values ​​for future time periods.

[0014] The present invention has the following beneficial effects: This invention uses a state-sensing unit to analyze key state indicators such as glucose concentration decay, osmotic rate, and volume-pressure response in real time. Combined with motion disturbance and drainage efficiency analysis, it dynamically allocates the weights of multi-source sensor data through a pre-trained correlation matrix, and then outputs accurate pressure prediction values ​​through a time-series model. This solves the problem that traditional fixed-parameter models cannot cope with dynamic interference, thus improving the accuracy of pressure prediction. By introducing a dynamic weight fusion mechanism, the system can adaptively allocate weights to different sensor data according to the current treatment stage and real-time physiological state. This allows the time-series prediction model to focus on the most valuable information at each moment, effectively suppressing noise interference and significantly improving the accuracy of pressure prediction and the stability of the system in the face of complex situations. Through stage gating and dynamic weighting, the system can automatically ignore irrelevant sensor data at specific stages, reducing computational redundancy and allowing computing resources to be more concentrated on processing critical tasks, thereby improving the overall efficiency of the system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a high-precision automatic pressure control system for peritoneal dialysis based on machine learning, provided in one embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the instantaneous permeation rate determination process provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a machine learning-based high-precision automatic pressure control system for peritoneal dialysis proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a high-precision automatic pressure control system for peritoneal dialysis based on machine learning, provided by the present invention.

[0020] Please see Figure 1 The diagram illustrates a high-precision automatic pressure control system for peritoneal dialysis based on machine learning, according to an embodiment of the present invention. The system includes: The acquisition unit 101 is used to acquire multi-source sensor data and initial glucose concentration of dialysate in real time during peritoneal treatment. The sensor data includes peritoneal fluid pressure, retention time, drainage rate, individual body temperature and inertial measurement data.

[0021] Intraperitoneal fluid pressure data: Acquired using a high-precision pressure sensor integrated into the front end of the peritoneal dialysis machine's drainage tubing, which remains positioned at the individual's mid-axillary level throughout the treatment process. The precise positioning of the high-precision pressure sensor ensures that its measurement baseline aligns with clinically standard intraperitoneal pressure measurement locations. The intraperitoneal fluid pressure data directly reflects the actual physical pressure within the abdominal cavity during treatment.

[0022] It is important to understand that, due to the constant intestinal peristalsis and respiratory movements within the abdominal cavity, it is necessary to extract an effective value representing the stable intra-abdominal fluid pressure from the raw, noisy pressure signal. Therefore, the acquired raw pressure signal needs to undergo real-time signal processing to extract the effective value representing the stable intra-abdominal fluid pressure. The specific process includes bandpass filtering: filtering the raw signal to retain the slow pressure fluctuations caused by respiration, while removing high-frequency noise (such as sensor electronic noise) and extremely low-frequency drift; feature extraction: calculating the filtered pressure signal within a continuous sliding time window (e.g., 30 seconds), using its trough value (i.e., end-expiratory pressure) as the effective representative value of the stable intra-abdominal fluid pressure within that time window, and using the effective representative value of the stable intra-abdominal fluid pressure as the intra-abdominal fluid pressure used in subsequent calculations.

[0023] Drainage rate data: This data is obtained in real time by a flow sensor (such as an ultrasonic flow meter) installed on the dialysis fluid outflow line. The drainage rate data is used to quantify the volume of fluid discharged from the abdominal cavity per unit time during the drainage phase.

[0024] Inertial measurement data: acquired through an inertial measurement unit worn on the surface of an individual's abdomen, such as an IMU. An IMU typically contains a three-axis accelerometer and a three-axis gyroscope to collect raw data characterizing the individual's body posture and acceleration in three-dimensional space. It is the basis for identifying changes in body position and external mechanical disturbances.

[0025] Individual body temperature data: This can be obtained through temperature sensors. In order to measure the temperature of the dialysate in order to indirectly reflect the core body temperature, the temperature sensor can be integrated into the dialysate tubing to obtain the individual temperature. The individual body temperature data is used to perform temperature compensation correction for physiological processes.

[0026] Retention time: It is not directly measured by physical sensors, but is automatically calculated and recorded by the system's internal clock module according to the treatment prescription. The time is started from the moment a perfusion ends and is updated in real time, reflecting the duration of the dialysis fluid in the peritoneal cavity.

[0027] Initial glucose concentration of dialysate: The initial glucose concentration is a preset known parameter, which is usually set by medical staff according to the prescription and entered into the system before the start of treatment, or automatically obtained by the system by recognizing the QR code on the dialysate bag. It should be noted that the initial glucose concentration of dialysate is the initial benchmark that determines the driving force of osmotic pressure.

[0028] It should be noted that after the raw sensor data is collected, preprocessing is required, such as signal filtering, timestamp alignment, and data encapsulation and transmission. The specific preprocessing process is a common technical means, which will not be described in detail in this embodiment.

[0029] The state sensing unit 102 is used to determine the estimated glucose concentration in real time based on the initial glucose concentration and the retention time; determine the instantaneous osmosis rate in real time based on the estimated glucose concentration and body temperature; determine the volume-pressure response coefficient based on the instantaneous osmosis rate and fluid pressure, and determine the volume-pressure state index in combination with the preset reference abdominal cavity volume; determine the action risk index in real time based on inertial measurement data and the volume-pressure state index; and determine the drainage efficiency index in the drainage stage based on the real-time drainage rate and the preset reference drainage rate.

[0030] It is important to understand that glucose in the dialysate is gradually absorbed by the peritoneum over time, leading to a decrease in glucose concentration and a corresponding reduction in osmotic pressure. Furthermore, directly measuring intraperitoneal glucose concentration in real time is very difficult and impractical. Therefore, a clinically validated attenuation model can be used to estimate the gradually changing glucose concentration, providing a crucial time-dynamic parameter for subsequent analysis of osmotic driving forces.

[0031] In this embodiment, based on the initial glucose concentration of the dialysate and the current retention time of the dialysate, the estimated glucose concentration at the current moment is calculated by using a preset glucose concentration decay model. The glucose concentration decay model decreases exponentially with the increase of retention time.

[0032] It should be noted that the pre-set glucose concentration decay model does not depend on the operation of this system in real-time treatment. Instead, it is a basic model that is pre-established and built into the system using clinical data. For example, researchers can collect a large amount of clinical data to record the actual glucose concentration in the dialysate at different time points during the patient's abdominal wall period. Then, using these large number of "time-concentration" data points, they can fit an exponential decay curve that best represents the overall trend through mathematical regression methods (such as nonlinear least squares method), which is the pre-set glucose concentration decay model.

[0033] The process of determining the instantaneous permeation rate is as follows: Figure 2 As shown, it includes: S101-1, obtain the peritoneal transport characteristics parameters of an individual, the intraperitoneal reference body temperature and the blood reference blood glucose concentration.

[0034] Peritoneal transport characteristics parameters characterize the "permeability" or "resistance" of an individual peritoneum as a semipermeable membrane. Peritoneal transport characteristics parameters can be obtained through standard peritoneal equilibration experiments. The specific peritoneal equilibration experiment process is a well-known technique in the art, and will not be described in detail in this embodiment.

[0035] It's important to understand that individual peritoneal transport characteristics can change in clinical practice, for example, due to factors such as peritonitis, long-term dialysis, and glycemic control. Therefore, the determination and updating of peritoneal transport characteristic parameters can follow three mechanisms: First, during the initial treatment, the system uses individualized parameter values ​​measured by a standard peritoneal equilibration test as the initial peritoneal transport characteristic parameters. Second, in subsequent treatments, the system uses real-time monitoring data to micro-calibrate the peritoneal transport characteristic parameters. For example, in the mid-to-late stages of the peritoneal retention phase, the system will use data such as the decay curve based on the estimated glucose concentration and the actual measured ultrafiltration volume. According to the algorithm, an equivalent peritoneal transport characteristic parameter for the current treatment cycle is derived through a pre-set inverse optimization algorithm. This equivalent parameter will be used to smoothly update the initial peritoneal transport characteristic parameter (e.g., using an exponentially weighted moving average method), so that the system can gradually adapt to the long-term slow changes in individual peritoneal function. The third point is that when the prediction unit calculates the dynamic weight, it will take the update magnitude of the peritoneal transport characteristic parameter as an implicit feature: if the recent calibration magnitude of the peritoneal transport characteristic parameter is large, the system will tend to reduce its dependence on the historical time series model and increase its attention to real-time status indicators such as instantaneous permeability rate.

[0036] It should be noted that a lower peritoneal transport characteristic parameter indicates high transport (good membrane permeability and fast substance exchange), while a higher peritoneal transport characteristic parameter indicates low transport (poor membrane permeability and slow substance exchange). The peritoneal transport characteristic parameter reflects the peritoneum's substance exchange capacity and is not directly related to the peritoneal compliance (peritoneal elasticity) reflected by changes in peritoneal volume-pressure. The former affects the permeation rate, while the latter affects the volume-pressure response coefficient.

[0037] Abdominal reference temperature (IRT) represents a baseline value for an individual's core abdominal temperature under stable conditions, following a personalized-first principle: the system prioritizes obtaining an individual's personalized baseline value from their electronic health record or historical treatment data; when individual data is lacking, a default value based on prior knowledge of healthy individuals (e.g., an IRT of 37.0°C) is used as the initial estimate. This ensures that the starting point for physiological parameter estimation is closer to the patient's actual condition. For example, in the absence of individual data, the system can directly use a standard medical common-sense value, 37.0°C, as the initial IRT for everyone. This is a simplified starting point.

[0038] Blood-side reference glucose concentration (BSG) represents a baseline value of blood glucose concentration for an individual before dialysis or in a stable state. It follows an individualization-first principle: the system prioritizes estimation based on the individual's anthropometric data (such as height, weight, and gender). Furthermore, in subsequent treatments, the system continuously calibrates the BSG using the individual's actual maximum tolerable volume data recorded in their historical treatment history, thereby making it more reflective of the individual's physiological characteristics and pathological state (such as changes in abdominal volume due to diseases like polycystic kidney disease). When individual data is lacking, a fasting blood glucose value based on prior knowledge of healthy individuals is used. For example, the median of a standard medical reference range, such as 5.5 mmol / L, can be used as the initial BSG.

[0039] S101-2, calculate the difference between the initial glucose concentration and the reference blood glucose concentration at each time point during the treatment process, and use it as the concentration difference value.

[0040] The concentration difference represents the osmotic pressure gradient between blood and dialysate, which is the fundamental driving force for the transfer of water from blood to dialysate (i.e., ultrafiltration).

[0041] It should be noted that in actual medical scenarios, the concentration difference generates osmotic pressure, which is the fundamental driving force that drives water to flow from the blood (low concentration side) to the dialysate (high concentration side). Without the concentration difference, ultrafiltration cannot be performed. Therefore, it can be understood that if the initial glucose concentration is much greater than the reference blood glucose concentration, then the concentration difference is a positive number.

[0042] S101-3 Calculate the absolute difference between the individual's body temperature and the abdominal reference body temperature at each moment during the treatment process, as the temperature difference value; calculate the product of the preset empirical coefficient and the temperature difference value, as the temperature compensation coefficient at each moment.

[0043] It is important to understand that since the diffusion rate of substances is positively correlated with temperature, when an individual's body temperature (e.g., due to fever) is higher than the reference body temperature, the transmembrane transport rate of water and solute will increase. Therefore, the effect of body temperature fluctuations on osmotic efficiency can be dynamically corrected by using a temperature compensation coefficient, making it more in line with physiological laws.

[0044] Since the temperature difference reflects the degree to which body temperature deviates from the normal baseline, a larger temperature difference indicates a more significant abnormality in body temperature, and thus a greater impact on osmotic efficiency. In physiology, the effect of temperature on the rate of substance diffusion is approximately linear within a certain physiological range. Therefore, calculating the temperature compensation coefficient by multiplying a preset empirical coefficient by the temperature difference is a reasonable and effective linear approximation method. The temperature compensation coefficient is used to dynamically adjust the calculated osmotic rate, making it more consistent with actual physiological conditions.

[0045] It should be noted that the preset experience coefficient can be determined in advance through historical data analysis. For example, researchers collect a large amount of historical clinical data, which needs to include the actual ultrafiltration volume measured during peritoneal dialysis at different body temperatures. Then, they use a linear regression model to analyze this data, find the quantitative relationship between changes in body temperature and changes in dialysis efficiency, and extract the regression coefficient in front of the variable "body temperature" from the fitted linear regression model. This regression coefficient quantifies the average influence of body temperature on the permeation rate, i.e., the preset experience coefficient. For example, based on historical data analysis, the preset experience coefficient is usually a coefficient that fluctuates around 1. Assuming it is set to 0.05, this means that when an individual's real-time body temperature is 1°C higher than their reference body temperature, their instantaneous permeation rate is expected to increase by 5%.

[0046] S101-4, calculate the ratio of the concentration difference to the peritoneal transport characteristic parameter, as the first ratio; calculate the product of the first ratio and the temperature compensation coefficient, as the instantaneous permeation rate at each time point.

[0047] It is important to understand that the instantaneous osmotic rate quantifies the theoretical rate at which the volume of water transferred from blood to dialysate per unit time under the current osmotic pressure gradient, individualized peritoneal characteristics, and real-time body temperature. If an individual's peritoneal transport characteristic parameters are small, resulting in a larger first ratio, this means that under the same osmotic pressure, dialysis efficiency is high and fluid inflow is fast. If an individual's temperature compensation coefficient is greater than 1 at a certain moment, the effect of body temperature at that moment on accelerating the molecular diffusion rate will increase the first ratio.

[0048] Since the concentration difference represents the current osmotic pressure gradient, it is the "driving force" that propels water from the blood to the dialysate. However, the same driving force produces different effects in different patients because each person's peritoneum, as a semipermeable membrane, has different resistances. The peritoneal transport characteristic parameter quantifies this physiological resistance. Therefore, dividing the concentration difference by the peritoneal transport characteristic parameter yields the first ratio, which reflects the theoretical rate of water transfer under a given osmotic pressure driving force, taking into account the individualized peritoneal resistance of each patient.

[0049] It should be noted that the “moment” or “current moment” used in this invention refers to a processing cycle or sampling cycle of the system. For example, assuming that the processing cycle of the system, that is, the interval of a “moment”, is set to 5 seconds, this means that the system completes the entire process from data acquisition to calculation every 5 seconds.

[0050] It should be noted that the peritoneum is a tissue with a physiological structure, not an idealized, unobstructed sieve. When substances are exchanged through the peritoneum, they will inevitably encounter physiological resistance from the peritoneum itself. This resistance is an objective physical fact. Therefore, the peritoneal transport characteristic parameters cannot be zero.

[0051] It is important to understand that since the abdominal cavity is not a rigid container, the internal pressure increases nonlinearly with increasing volume. Generally, the closer the abdominal cavity is to its filling limit, the faster the pressure rises. Therefore, a volume-pressure state index is needed to quantify this nonlinear tightness.

[0052] In this embodiment, the initial perfusion volume of the dialysate is obtained; starting from the start of treatment, the time interval between each moment and the previous moment is recorded; the product of the instantaneous osmosis rate at each moment and the corresponding time interval is calculated as the volume increment; all volume increments before each moment are accumulated to obtain the volume change caused by osmosis from the start of treatment to each moment; the volume change at each moment is added to the initial perfusion volume to obtain the estimated value of the peritoneal volume; based on the estimated value of the peritoneal volume at each moment and the corresponding fluid pressure, the volume-pressure response coefficient is determined.

[0053] The initial perfusion volume is a known, fixed value set by the treatment prescription. The initial perfusion volume is the baseline starting point for all subsequent volume calculations, such as 2000 mL.

[0054] For example, if the system completes the entire process from data acquisition to calculation every 5 seconds, then the time interval is 5 seconds.

[0055] It should be noted that the instantaneous osmosis rate (usually in mL / min) is a rate, while the time interval (in min, which can be converted from seconds to minutes) is time. The product of the instantaneous osmosis rate and the time interval is the net increase in the volume of fluid in the abdominal cavity due to osmosis during the time interval.

[0056] The volume change represents the net increase in intraperitoneal fluid generated purely by ultrafiltration from the start of treatment to the present.

[0057] It should be noted that during the drainage phase, fluid will be drained outwards. Therefore, after adding the volume change at each moment to the initial perfusion volume, it is necessary to subtract the total volume of fluid drained from the start of the drainage phase to the current moment to obtain the estimated value of the abdominal cavity volume at the current moment of the drainage phase.

[0058] It should be noted that, at the same time, the real-time drainage rate is obtained from the flow sensor. Based on the real-time drainage rate, the total volume of liquid drained from the start of the drainage stage to the current moment is calculated in a similar way to the calculation of volume change. This is calculated by accumulating the product of the real-time drainage rate at each moment of the drainage stage and the corresponding time interval. That is, it is consistent with the integral logic of volume change, only the parameters used are different. This embodiment will not be described in detail.

[0059] To clearly understand the magnitude of pressure change caused by a unit change in abdominal cavity volume, i.e., to accurately determine the volume-pressure response coefficient, as an example, we statistically analyze the estimated abdominal cavity volume and corresponding fluid pressure at various time points to obtain a volume-pressure dataset. Based on the volume-pressure dataset, we establish the volume-pressure relationship curve for each individual using a curve fitting algorithm. We then calculate the slope of the volume-pressure relationship curve at each time point and use the statistical value of the slope as the volume-pressure response coefficient.

[0060] It should be noted that the corresponding fluid pressure refers to the pressure at the same moment as the estimated abdominal cavity volume.

[0061] Since only "stable pressure" that excludes external disturbances can form effective and causally related paired data points with the estimated "volume", using disturbed pressure data will severely distort the fitted curve. Therefore, in this invention, disturbed segments are not used in the fitting of the volume pressure curve. That is, when the abdominal cavity is subjected to obvious and instantaneous external or internal mechanical disturbances, the measured abdominal fluid pressure value is not used.

[0062] It should be noted that the volume-pressure relationship curve reveals the quantitative relationship between the abdominal cavity volume and pressure for a given individual. For example, based on the volume-pressure dataset, a volume-pressure relationship curve that best reflects the distribution trend of data points in the dataset can be established through curve fitting algorithms (such as linear regression or low-order polynomial fitting). The specific methods for constructing curves through curve fitting algorithms are common technical means, and will not be elaborated in this embodiment.

[0063] In actual working scenarios, each data point (estimated abdominal cavity volume, fluid pressure) contains noise (sensor error, respiratory fluctuations, intestinal peristalsis, etc.). If an instantaneous slope is calculated based on only two adjacent data points, the slope will be very unstable, sometimes positive and sometimes negative, and cannot truly reflect the macroscopic elasticity of the abdominal cavity. Therefore, statistical methods can be used to analyze the most representative central trend value from all fitted local slopes, that is, the statistical value of the slope.

[0064] For example, to calculate the slope values ​​at multiple instants, sort all the instantaneous slopes by size, and take the middle value, then the middle value is the statistical value of the slope.

[0065] It should be noted that within the physiological range of peritoneal dialysis, an increase in the volume of the abdominal cavity inevitably leads to an increase in the intra-abdominal fluid pressure. This is a monotonically increasing relationship; therefore, without violating basic physical and physiological laws, the volume-pressure response coefficient is a positive number.

[0066] It's important to understand that a larger volume-pressure response coefficient means that a small increase in volume can lead to a large increase in pressure. This strongly suggests that the individual has low compliance or a rigid abdominal cavity, commonly seen in people with a history of abdominal surgery, those recovering from peritonitis, those with peritoneal adhesions, or certain predispositions. For these individuals, the risk of sudden pressure changes is extremely high, requiring the highest level of vigilance. Conversely, a smaller volume-pressure response coefficient (a flatter curve) means that the system can accommodate a larger increase in volume with only a slight increase in pressure. This strongly suggests that the individual has high compliance or a flexible abdominal cavity, commonly seen in obese patients or those with a loose abdominal wall. These individuals are more tolerant of changes in fluid volume and have a relatively lower risk of pressure changes.

[0067] In this embodiment, a preset abdominal cavity reference volume value is obtained; the square of the preset abdominal cavity reference volume value is calculated as the volume square value; the ratio of the estimated abdominal cavity volume value to the volume square value at each time point is calculated as the second ratio; the product of the second ratio and the volume-pressure response coefficient is calculated as the volume-pressure state index.

[0068] The preset abdominal cavity reference volume is a safety threshold used for risk assessment, initialized based on clinical knowledge and dynamically calibrated based on individual data. For example, in the absence of individual historical data, an initial value can be set based on group historical data. The system can have a built-in statistical regression model based on this data. The input to the statistical regression model is easily accessible anthropometric data (height, weight, gender), and the output is an estimated preset abdominal cavity reference volume. For example, assuming the input is an individual's height (e.g., 175cm), weight (e.g., 70kg), and gender (e.g., male), the system inputs these parameters into the built-in statistical regression model to obtain an abdominal cavity reference volume, such as 2800mL. The specific construction method of the statistical regression model is a technique well-known to those skilled in the art and will not be elaborated upon in this embodiment.

[0069] It should be noted that the abdominal cavity is a physical space: even in a completely fasted and emptied state, the adult abdominal cavity is an inherent body cavity with fixed geometric dimensions. Therefore, the preset reference volume of the abdominal cavity cannot be zero.

[0070] When calculating the second ratio, directly using a preset abdominal cavity reference volume as the denominator yields a linear value, representing the current abdominal cavity volume as a percentage of the reference limit. However, this linear relationship fails to accurately reflect the true physiological risk—as the abdominal cavity approaches its filling limit, the rate of pressure increase accelerates dramatically; the closer to the limit, the faster the pressure on the abdominal wall increases with each additional unit of pressure applied. Therefore, the square of the volume is introduced as the denominator.

[0071] It is important to understand that the second ratio reflects the fullness of the abdominal cavity at the current moment. If the estimated abdominal cavity volume at a certain moment gradually approaches the preset reference abdominal cavity volume, the second ratio will increase sharply. In this case, the nonlinear risk can be amplified at that moment, and the danger signal that the abdominal cavity becomes more full with each increase in pressure can be amplified. If the volume-pressure response coefficient of an individual is larger at a certain moment, it means that for an individual with a stiff abdominal cavity, even if the fullness is not high, the volume-pressure state index of that individual may be larger, because the individual's abdominal cavity is more "fragile".

[0072] It is important to understand that the same action (such as turning over) may have little effect when the abdominal cavity is empty, but may cause drastic pressure fluctuations when the abdominal cavity is full. Therefore, the impact of changes in an individual's actions under the current tense state of the user's abdominal cavity should also be considered.

[0073] Inertial measurement data includes acceleration data.

[0074] In this embodiment, based on inertial measurement data, the attitude change at each moment during the treatment process is calculated using an attitude calculation algorithm; high-pass filtering is applied to the acceleration data to separate high-frequency components; norm calculation is performed on the attitude change, and normalization is applied to obtain the intensity of the body position change; the variance of the high-frequency components is normalized to obtain the high-frequency impact intensity; based on preset weights, the intensity of the body position change and the high-frequency impact intensity are weighted and fused to obtain the original disturbance intensity; the product of the original disturbance intensity and the volume-pressure state index is calculated as the action risk index.

[0075] It should be noted that, based on inertial measurement data, the Euler angles of the individual's abdomen in three-dimensional space are calculated using attitude calculation algorithms (such as complementary filtering, gradient descent, or Kalman filtering). Then, based on the Euler angles at consecutive time points, a vector is constructed, which is the attitude change.

[0076] For example, when an individual starts to roll over to the right from an initial supine position, the Euler angles can be represented as [roll angle 0°, pitch angle 0°, yaw angle 0°] (all angles are 0° when lying supine). After rolling over to the right, the Euler angles can be represented as [roll angle 10°, pitch angle 0°, yaw angle 0°]. The change in posture after this movement is [roll angle 10°, pitch angle 0°, yaw angle 0°], which is a vector. The Euler angles are used to quantify the patient's positional changes. The specific calculation of the Euler angles is a common technique, which will not be described in detail in this embodiment.

[0077] Since posture changes are three-dimensional vectors containing angular changes in three directions, directly using them would require processing each direction separately, leading to complex data dimensions and making it difficult to comprehensively assess the overall intensity of the movement. Therefore, by calculating the norm of posture changes, the changes in the three directions can be integrated into a single scalar value, namely the intensity of the postural change. The intensity of the postural change comprehensively reflects the overall amplitude of the posture change, and can be uniformly quantified regardless of whether the patient is turning over, sitting up, or twisting their waist.

[0078] It should be noted that the norm usually refers to the L2 norm, which is used to calculate the length of a vector.

[0079] In this embodiment, high-pass filtering is applied to the acceleration data in the inertial measurement data to separate the low-frequency signals generated by gravitational acceleration and slow body movement, while retaining the high-frequency signals generated by instantaneous impacts, i.e., the high-frequency components. The specific process of high-pass filtering will not be described in detail.

[0080] High-frequency components are signals representing instantaneous impacts (such as coughing, tremors, or slaps) separated from acceleration data through high-pass filtering. Variance is a statistical measure of data fluctuation; a larger variance indicates more severe signal fluctuations, meaning a greater impact intensity. Therefore, calculating the variance of high-frequency components can condense a segment of instantaneous impact signals into a single numerical value representing the intensity of the action. This value can then be normalized to a dimensionless high-frequency impact intensity.

[0081] It should be noted that the normalization process used in the calculation of body position change intensity and high frequency impact intensity is specifically a minimum-maximum normalization. The minimum and maximum values ​​can be global ranges obtained based on historical data statistics, or preset physiological reasonable ranges. For example, the variance corresponding to the maximum possible high frequency impact intensity is set to 1, and the minimum high frequency impact intensity is set to 0.

[0082] The process of obtaining the original disturbance intensity can be represented by the following formula: Original disturbance intensity = a × body position change intensity + b × high frequency impact intensity, where a represents the preset weight of body position change intensity; b represents the preset weight of high frequency impact intensity, and a + b equals 1.

[0083] It should be noted that the specific values ​​of a and b depend on the risk preference for different types of disturbances. For example, if one believes that coughing (an impactful action) poses a greater immediate threat to stress than slowly turning over (a postural change action), then b would be greater than a, and b could be 0.6, while a could be 0.4.

[0084] It is important to understand that the initial disturbance intensity reflects the severity of the external action itself without considering the state of the abdominal cavity. When the abdominal cavity is relaxed (the volume-pressure state index value is small), even if there is a large action (the initial disturbance intensity is large), the final risk index of the risk action will not be very high.

[0085] Because the same movement (such as turning over) has minimal impact when the abdominal cavity is empty, but can trigger dramatic pressure fluctuations when the abdominal cavity is full, the movement risk index is obtained by multiplying the original disturbance intensity (which combines the intensity of external movements involving changes in body position and high-frequency impacts) with the volume-pressure state index. If the abdominal cavity is relaxed (i.e., the volume-pressure state index is small), even a large movement will not result in a high movement risk index; if the abdominal cavity is tense (i.e., the volume-pressure state index is large), even a small movement will be amplified to a high-risk value.

[0086] In this embodiment, a preset minimum constant is obtained; the real-time drainage rate at each moment during the treatment process is added to the preset minimum constant to obtain the comprehensive drainage rate; the ratio of the preset reference drainage rate to the comprehensive drainage rate is calculated as a drainage efficiency index.

[0087] It should be noted that the preset minimum constant is a very small positive number, such as 0.001 mL / min, introduced to ensure computational stability and mathematical rigor.

[0088] It should be noted that the preset reference drainage rate is a theoretical benchmark value, representing the level that the drainage rate should reach under the condition of "ideal patency" with the current abdominal cavity volume and pressure. It can be analyzed from the average rate of stable drainage during the patency phase in the past treatments of an individual. For example, if the historical data of individual A shows that when drainage is smooth, the drainage rate is usually stable at around 120-130 mL / min, then the median of 125 mL / min can be set as the preset reference drainage rate for this treatment.

[0089] Because the calculated ratio is dimensionless and relative, it can accommodate differences in drainage rates among different individuals and at different stages. Therefore, the calculation of the drainage efficiency index adopts the form of a ratio between a preset reference drainage rate and the comprehensive drainage rate. This design is based on a simple and effective diagnostic logic: judging whether the drainage system is smooth by comparing "how fast it should flow" with "how fast it actually flows".

[0090] It is important to understand that the drainage efficiency index is a barometer for diagnosing whether the drainage system is functioning properly. If the actual drainage rate is much lower than expected, it strongly suggests that the drainage tube may be blocked, kinked, or wrapped with tissue. In this case, since the preset reference drainage rate is much greater than the denominator comprehensive drainage rate, the drainage efficiency index will be greater than 1.

[0091] The prediction unit 103 is used to determine the dynamic weights of each sensor data at each moment of the current working stage of the system by fusing instantaneous permeation rate, volume-pressure state index, action risk index and drainage efficiency index through a pre-trained correlation matrix; after weighting the sensor data with dynamic weights, it is input into the pre-trained time series training model to obtain the pressure prediction value.

[0092] It's important to understand that while all sensor data are treated equally at all times, during the retention period, drainage flow data is zero and meaningless; treating it equally would become noise. When an individual is stationary, inertial measurement data has very low value. Therefore, through dynamic weighting, the system can implement an "attention mechanism." For example, when movement risk indicators are high, the system automatically assigns high weight to inertial measurement data, paying close attention to changes in body position. When drainage efficiency indicators are abnormally high, the system assigns high weight to pressure and flow sensor data, as they are key evidence for diagnosing obstruction at this moment. When volume-pressure status indicators are high (abdominal tightness), the system comprehensively increases the sensitivity of all data.

[0093] In this embodiment, a binary gating vector is generated based on the current working stage of the system. This binary gating vector is then used to select effective indicators from instantaneous permeation rate, volume-pressure state indicators, action risk indicators, and drainage efficiency indicators, forming an effective feature vector. The binary gating vector is used to disable sensor data channels whose physical meaning is unclear or invalid at a specific working stage. The effective feature vector is then input into a pre-trained correlation matrix to obtain the dynamic weights of each sensor data channel. The pre-trained correlation matrix is ​​obtained through end-to-end training.

[0094] For example, at different treatment stages, the system intelligently "ignores" irrelevant data and focuses only on key information. Suppose we construct a vector from the four key state indicators at the current moment: instantaneous permeability rate, volume-pressure state indicator, action risk indicator, and drainage efficiency indicator. The treatment stages can be divided into the abdominal retention stage, drainage stage, and perfusion stage. Then, in the abdominal retention stage: the binary gating vector is [1, 1, 1, 0]. The drainage efficiency indicator is disabled in this stage (value is 0) because there is no drainage at this stage, making this indicator meaningless; forcibly calculating it would only introduce noise. Therefore, the effective feature vector is [instantaneous permeability rate, volume-pressure state indicator, and drainage efficiency indicator]. [Target, Action Risk Indicator]; In the drainage stage: the binary gating vector is [0, 1, 1, 1], and the instantaneous permeation rate is disabled in this stage (value is 0) because the small amount of new fluid generated by permeation is not the focus during active drainage, and the drainage rate is the dominant factor. Therefore, the effective feature vector is [volume-pressure state indicator, action risk indicator, drainage efficiency indicator]; In the infusion stage: the binary gating vector is [1, 1, 1, 0], and the drainage efficiency indicator is disabled in this stage (value is 0) because there is no drainage at all in this stage. Therefore, the effective feature vector is [instantaneous permeation rate, volume-pressure state indicator, action risk indicator].

[0095] The system's main working stages include the perfusion stage, the abdominal retention stage, and the drainage stage. Since there is no need to calculate core indicators in the treatment preparation and termination stages, the dynamic weight allocation logic will not be used for the time being.

[0096] In this embodiment, dynamic weights are weighted and fused with the corresponding original sensor data to form weighted feature data; the weighted feature data is then input into a pre-trained time series training model to obtain the pressure prediction value for future time moments.

[0097] In this context, it can be understood that weighted fusion refers to multiplying dynamic weights with the corresponding original sensor data.

[0098] It should be noted that the pre-trained time series prediction model can be an LSTM model. The specific training process will not be described in this embodiment. For example, a weighted feature data sequence over a period of time (not just data from a single moment) is input into the pre-trained time series prediction model. The dynamic patterns contained in this weighted data sequence are analyzed, and finally a pressure prediction value for a future moment is output. For example, it is predicted that the pressure will rise to 12.5 mmHg in 5 seconds.

[0099] The control unit 104 is used to adjust the operating parameters of the peritoneal dialysis machine based on the predicted pressure value.

[0100] Among them, the system can automatically send control commands to the peritoneal dialysis machine by comparing the predicted pressure value with the preset safe pressure threshold and based on the comparison results.

[0101] It should be noted that the preset safe pressure threshold is not a fixed value, but rather dynamically adjusted based on the individual's real-time status. The system typically compensates for the safe pressure threshold based on real-time body position information (such as lying down, sitting up, or standing) calculated from inertial measurement data. For example, when the system detects that the patient changes from lying down to sitting, the safe threshold will be lowered accordingly to account for the increase in intra-abdominal pressure baseline caused by the change in body position. At the same time, the system fine-tunes the threshold based on the real-time calculated volume-pressure state index. When the volume-pressure state index shows that the abdominal cavity is in a state of high distension and low compliance (i.e., the volume-pressure state index value is large), the system will adopt a more conservative safe threshold, such as a general threshold based on clinical consensus. When the intra-abdominal pressure continuously exceeds 18-20 cmH2O (approximately equal to 13.5 mmHg), the individual's discomfort increases significantly, and 13.5 mmHg can be set as the preset safe pressure threshold.

[0102] For example, if the system predicts that the intraperitoneal pressure will exceed the safety threshold within the next 2 minutes if the current perfusion flow rate is maintained, the perfusion flow rate can be reduced: the system sends a command to the dialysis machine to reduce the flow rate from, for example, 200 mL / min to 100 mL / min, or to pause perfusion: if the predicted risk of a pressure spike is high and urgent, the system will issue a command to pause perfusion, temporarily and completely stopping the inflow of fluid, allowing the intraperitoneal pressure a buffer and stabilization period.

[0103] Furthermore, the control unit collaborates with the system status perception unit and prediction unit to form a closed-loop management system: when the action risk indicator or drainage efficiency indicator indicates abnormality, the system not only adjusts the weight of the pressure prediction model accordingly, but also triggers graded alarms and initiates corresponding intervention plans (such as adjusting the flow rate or pausing perfusion). At the same time, it continuously monitors the changes in physiological parameters after intervention to verify the effect, thereby realizing a complete clinical treatment path from status perception and risk warning to precise intervention and feedback verification.

[0104] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0105] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A high-precision automatic pressure control system for peritoneal dialysis based on machine learning, characterized in that, The system includes: The acquisition unit is used to acquire multi-source sensor data and initial glucose concentration of dialysate in real time during peritoneal treatment. The sensor data includes peritoneal fluid pressure, retention time, drainage rate, individual body temperature, and inertial measurement data. The state sensing unit is used to determine the estimated glucose concentration in real time based on the initial glucose concentration and retention time; determine the instantaneous osmosis rate in real time based on the estimated glucose concentration and body temperature; determine the volume-pressure response coefficient based on the instantaneous osmosis rate and fluid pressure, and determine the volume-pressure state index in combination with the preset reference abdominal cavity volume; determine the action risk index in real time based on inertial measurement data and volume-pressure state index; and determine the drainage efficiency index in the drainage stage based on the real-time drainage rate and the preset reference drainage rate. The prediction unit is used to determine the dynamic weights of each sensor data at each moment of the current working stage of the system by fusing instantaneous permeation rate, volume-pressure state index, action risk index and drainage efficiency index through a pre-trained correlation matrix. After weighting the sensor data with dynamic weights, the data is input into the pre-trained time series training model to obtain the pressure prediction value. The control unit is used to adjust the operating parameters of the peritoneal dialysis machine based on the predicted pressure value.

2. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 1, characterized in that, The process for determining the estimated glucose concentration includes: Based on the initial glucose concentration of the dialysate and the current retention time of the dialysate, the estimated glucose concentration at the current moment is calculated using a preset glucose concentration decay model. The glucose concentration decay model decreases exponentially with the increase of retention time.

3. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 2, characterized in that, The instantaneous permeation rate determination process includes: Obtain individual peritoneal transport characteristics, intraperitoneal reference body temperature, and blood reference blood glucose concentration; Calculate the difference between the initial glucose concentration and the reference blood glucose concentration at each time point during the treatment process, and use it as the concentration difference. The absolute difference between the individual's body temperature and the abdominal reference body temperature at each moment during treatment is calculated as the temperature difference; the product of the preset empirical coefficient and the temperature difference is calculated as the temperature compensation coefficient at each moment. The ratio of the concentration difference to the peritoneal transport characteristic parameter is calculated as the first ratio; the product of the first ratio and the temperature compensation coefficient is calculated as the instantaneous permeation rate at each time point.

4. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 1, characterized in that, The process for determining the volumetric-pressure response coefficient includes: Obtain the initial perfusion volume of the dialysate; Starting from the beginning of treatment, the time interval between each moment and the previous moment is recorded; the product of the instantaneous osmosis rate at each moment and the corresponding time interval is calculated as the volume increment; all volume increments before each moment are accumulated to obtain the volume change caused by osmosis from the beginning of treatment to each moment. The volume change at each time point is added to the initial perfusion volume to obtain the estimated abdominal cavity volume. Based on the estimated abdominal cavity volume and the corresponding fluid pressure at each time point, the volume-pressure response coefficient is determined.

5. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 4, characterized in that, The determination of the volume-pressure response coefficient based on the estimated abdominal cavity volume and fluid pressure at various time points includes: The estimated abdominal cavity volume and corresponding fluid pressure at each time point were statistically analyzed to obtain a volume-pressure dataset. Based on the volume-pressure dataset, a curve fitting algorithm is used to establish the volume-pressure relationship curve for each individual. Calculate the slope of the volume-pressure relationship curve at each time point, and use the statistical value of the slope as the volume-pressure response coefficient.

6. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 5, characterized in that, The process for determining the volumetric-pressure state index includes: Obtain the preset reference abdominal cavity volume; Calculate the square of the preset abdominal cavity reference volume and use it as the volume square value; Calculate the ratio of the estimated abdominal cavity volume to the square of the volume at each time point, and use it as the second ratio; The product of the second ratio and the volumetric-pressure response coefficient is calculated and used as the volumetric-pressure state index.

7. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 1, characterized in that, The inertial measurement data includes acceleration data; the process for determining the motion risk index includes: Based on inertial measurement data, the attitude change at each moment during the treatment process is calculated using an attitude calculation algorithm; High-pass filtering is applied to the acceleration data to separate the high-frequency components; The norm of the posture change is calculated and normalized to obtain the intensity of the posture change; the variance of the high-frequency component is normalized to obtain the high-frequency impact intensity. Based on preset weights, the intensity of body position change and the intensity of high-frequency impact are weighted and fused to obtain the original disturbance intensity; Calculate the product of the original disturbance intensity and the volume-pressure state index as the action risk index.

8. The high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 1, characterized in that, The process for determining the drainage effectiveness index includes: Obtain the preset minimum constant; The comprehensive drainage rate is obtained by adding the real-time drainage rate at each moment during the treatment process to a preset minimum constant. The ratio of the preset reference diversion rate to the comprehensive diversion rate is calculated and used as a diversion efficiency indicator.

9. A high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 1, characterized in that, The dynamic weights of each sensor data point are determined by fusing instantaneous penetration rate, volume-pressure state index, action risk index, and drainage efficiency index through a pre-trained correlation matrix, including: Based on the current working stage of the system, a binary gating vector is generated. The binary gating vector is used to select effective indicators from instantaneous permeation rate, volume-pressure state indicators, action risk indicators and drainage efficiency indicators to form an effective feature vector. The binary gating vector is used to disable sensor data channels that have unclear physical meaning or are invalid in a specific working stage. The effective feature vectors are input into the pre-trained correlation matrix to obtain the dynamic weights of each sensor data channel. The pre-trained correlation matrix is ​​learned through end-to-end training.

10. A high-precision automatic pressure control system for peritoneal dialysis based on machine learning according to claim 9, characterized in that, The step of weighting the sensor data using dynamic weights and then inputting it into a pre-trained time-series training model to obtain pressure prediction values ​​includes: The dynamic weights are then weighted and fused with the corresponding original sensor data to form weighted feature data; Weighted feature data is input into a pre-trained time series training model to obtain stress prediction values ​​for future time periods.

Images