Intelligent decision-making system for CAP treatment of kidney disease and skin complications with multi-parameter dynamic monitoring

Abstract

This invention relates to the fields of intelligent medical control and low-temperature plasma medical applications, specifically an intelligent decision-making system for CAP treatment of skin complications in kidney disease with multi-parameter dynamic monitoring. It includes: a multimodal vital sign acquisition module: acquiring multidimensional physiological characteristic signals of local tissues and simultaneously acquiring periodic metabolic interference data containing dialysis time-series information; a tolerance threshold inversion module: inverting the local tissue redox tolerance threshold using a tissue redox tolerance calculation model; a dynamic interference compensation module: generating a real-time safety boundary threshold; an adaptive strategy generation module: generating candidate treatment control strategies using a deep reinforcement learning algorithm; and a risk boundary control module: calculating the expected oxidative stress level, and if it is greater than or equal to the real-time safety boundary threshold, executing a blocking or downgrading operation; otherwise, outputting the final treatment instruction. This invention improves the precision of treatment dose control to the physiological response level, effectively avoiding the risk of iatrogenic skin necrosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical control and low-temperature plasma medical application technology, specifically to an intelligent decision-making system for CAP treatment of skin complications of kidney disease with multi-parameter dynamic monitoring. Background Technology

[0002] In the clinical treatment of skin complications of chronic kidney disease, plasma technology is often used to improve the microenvironment of the affected area and promote healing. The safety and effectiveness of the treatment largely depend on the precise matching of the treatment dose with the physiological tolerance of the local tissue.

[0003] In traditional methods, the control of treatment dosage mainly relies on fixed equipment preset parameters or the clinical experience of medical staff, usually assuming that the patient's skin tissue condition is relatively stable. However, due to the influence of the dialysis cycle, the metabolic environment and tissue physiological state of patients with chronic kidney disease will undergo drastic periodic fluctuations. Under the traditional static control mode, it is impossible to perceive and adapt to this dynamic change in tissue tolerance caused by dialysis, which makes it easy to cause iatrogenic skin damage due to dosage exceeding the limit during the patient's physiological vulnerable period.

[0004] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention discloses an intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease. Specifically, the technical solution of this invention is as follows:

[0006] The multimodal vital signs acquisition module is configured to perform non-invasive monitoring of local tissues of the target object, acquire multidimensional physiological characterization signals, and simultaneously acquire periodic metabolic interference data containing dialysis time sequence information.

[0007] The tolerance threshold inversion module is configured to invert the local tissue redox tolerance threshold based on multidimensional physiological characterization signals and using a preset tissue redox tolerance calculation model.

[0008] The dynamic interference compensation module is configured to identify the phase of the current metabolic cycle based on periodic metabolic interference data, and accordingly perform dynamic baseline correction on the local tissue redox tolerance threshold to generate a real-time safety boundary threshold.

[0009] The adaptive policy generation module is configured to use the real-time safety boundary threshold as a state constraint and use a deep reinforcement learning algorithm to optimize the relationship between treatment dose and tissue damage, thereby generating candidate treatment control strategies.

[0010] The risk boundary control module is configured to calculate the expected oxidative stress level corresponding to the candidate treatment control strategy and execute the safety judgment logic: if the expected oxidative stress level is greater than or equal to the real-time safety boundary threshold, then the blocking or downgrading operation is executed to generate the final treatment instruction; if the expected oxidative stress level is less than the real-time safety boundary threshold, then the candidate treatment control strategy is output as the final treatment instruction.

[0011] Optionally, methods for obtaining multidimensional physiological representation signals include:

[0012] Real-time acquisition of skin impedance spectrum data, microcirculation blood perfusion data, and local surface temperature data of the target area;

[0013] Frequency response analysis was performed on skin impedance spectroscopy data to extract the stratum corneum water content characteristics;

[0014] Temporal fluctuation analysis was performed on microcirculatory blood perfusion data to extract vasomotor response characteristics;

[0015] The characteristics of stratum corneum water content, vasomotor response, and local surface temperature are combined to form a multidimensional physiological characterization signal.

[0016] Optionally, methods for retrieving the local tissue redox tolerance threshold include:

[0017] A pre-built multilayer perceptron regression model is configured to take multidimensional physiological characterization signals as input and map them to a predefined redox buffer capacity feature space.

[0018] Calculate the vector projection of the multidimensional physiological representation signal in the feature space;

[0019] The magnitude of the vector projection is determined as the initial tolerance capacity at the current moment;

[0020] Based on the initial tolerance capacity and combined with the preset tissue cell viability decay data curve, the local tissue redox tolerance threshold is calculated.

[0021] Optionally, methods for dynamically correcting the redox tolerance threshold of local tissues include:

[0022] Obtain the timestamp of the most recent dialysis end from the periodic metabolic disturbance data;

[0023] Calculate the time difference between the current time and the timestamp of the most recent dialysis end to determine the specific time period during the current interdialysis interval;

[0024] Based on the specific time period, the corresponding toxin accumulation coefficient and water load coefficient are retrieved from the preset sawtooth metabolic fluctuation data model;

[0025] A nonlinear decay factor is constructed using the toxin accumulation coefficient and the water load coefficient.

[0026] The corrected real-time safety boundary threshold is obtained by multiplying the local tissue redox tolerance threshold by the nonlinear decay factor.

[0027] Optionally, methods for generating candidate treatment control strategies include:

[0028] Construct a reinforcement learning state space, which includes real-time safety boundary thresholds, current wound healing rate indicators, and multidimensional physiological representation signals.

[0029] Construct an action space, which includes plasma jet intensity, gas flow rate, and duration of a single action;

[0030] An asymmetric reward function is designed where a reward value positively correlated with the wound healing rate is given when the expected oxidative stress level is below the real-time safety boundary threshold; and an exponentially increasing penalty value is given when the expected oxidative stress level is above or equal to the real-time safety boundary threshold.

[0031] The current state is input into the pre-trained policy network, which outputs the action probability distribution in the action space, and then samples and generates candidate treatment control strategies based on the action probability distribution.

[0032] Optionally, methods for safety-tailoring candidate treatment control strategies include:

[0033] Based on the plasma jet intensity and single-action duration in the candidate treatment control strategy, the expected dose of exogenous reactive oxygen species is calculated.

[0034] The expected dose of exogenous reactive oxygen species is compared with the real-time safety boundary threshold.

[0035] If the dose of exogenous reactive oxygen species is less than the real-time safety boundary threshold, the candidate treatment control strategy will be directly output as the final treatment instruction.

[0036] If the exogenous reactive oxygen species dose is greater than or equal to the real-time safety boundary threshold, the plasma jet intensity remains unchanged, and the duration of a single treatment is reduced until the recalculated exogenous reactive oxygen species dose is less than the real-time safety boundary threshold. The reduced strategy is then output as the final treatment instruction.

[0037] Optionally, performing blocking or degradation operations may also include non-monotonic decision logic:

[0038] Determine whether the real-time safety boundary threshold is lower than the preset critical crash threshold;

[0039] If the real-time safety boundary threshold is lower than the critical collapse threshold, the wound healing rate index is ignored, the plasma jet intensity parameter in the candidate treatment control strategy is forcibly set to zero, and a final treatment instruction to pause treatment and issue alarm data is generated.

[0040] If the real-time security boundary threshold is higher than or equal to the critical crash threshold, then the security trimming step continues.

[0041] Optionally, the system also includes an evolutionary analysis and model update module for:

[0042] After the final treatment instruction is executed, the changing trends of multidimensional physiological representation signals are continuously monitored;

[0043] Calculate the deviation between the actual tissue response and the expected response after treatment;

[0044] If the deviation exceeds the preset tolerance range, the periodic metabolic disturbance data and multidimensional physiological characterization signals at the current moment are extracted and stored in the experience playback pool.

[0045] The digital twin model of tissue redox tolerance was fine-tuned online using data from the experience replay pool to update the local tissue redox tolerance threshold inversion parameters at subsequent time points.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention constructs a sawtooth-shaped metabolic fluctuation model based on dialysis time sequence through a dynamic interference compensation module. It uses the toxin accumulation coefficient and water load coefficient to dynamically correct the local tissue redox tolerance threshold in real time, which can accurately identify the physiological vulnerability window of high toxin and low tolerance during the dialysis cycle. This mechanism effectively eliminates the physiological deception effect caused by the violent fluctuation of the physiological environment between dialysis sessions, and transforms the safety boundary from a static empirical value to a dynamic physiological benchmark. Thus, without reducing the effectiveness of treatment, it fundamentally avoids the risk of iatrogenic skin necrosis caused by conventional dose treatment when the patient's metabolic defense is weakest.

[0048] 2. This invention combines a tolerance threshold inversion module with multimodal vital sign acquisition technology, employs time-domain digital filtering and energy integration to extract vasomotor response features, and utilizes a multilayer perceptron to map multidimensional physiological signals onto a redox buffer capacity feature space for vector projection calculation. This method can quantify invisible tissue latent blood bars, i.e., initial tolerance capacity, and by converting the dimensionless feature projection modulus into a concentration threshold with clear physical meaning, it achieves deep perception of the skin barrier-circulation-metabolism coupling state of patients with chronic kidney disease, providing a geometrically interpretable physiological anchor for subsequent precise dose control.

[0049] 3. This invention utilizes the collaborative work of an adaptive strategy generation module and a risk boundary control module. It employs a deep reinforcement learning algorithm with an asymmetric reward function to find a strategy that maximizes the wound healing rate within the safety boundary and implements safety tailoring based on a physical dose model. In particular, when facing the risk of exceeding limits, a degradation strategy is adopted that maintains the jet intensity constant and reduces the duration of a single application. This ensures that the final dose of exogenous reactive oxygen species introduced is strictly below the real-time safety boundary and avoids physical failure problems such as plasma extinguishing or changes in active ingredients due to reduced intensity. This achieves the optimal match between treatment intensity and tissue tolerance.

[0050] 4. This invention introduces digital twin and experience playback mechanisms through evolution analysis and model update modules to continuously monitor the temperature rise slope and microcirculation blood flow variance after treatment, calculate the deviation between actual tissue response and expected dose to fine-tune model parameters online; at the same time, combined with the extreme risk collapse threshold circuit-breaking logic based on historical averages, the system not only has the self-evolutionary ability to cope with the physiological characteristic drift caused by the long-term development of the patient's disease, but also can forcibly block treatment when the tissue tolerance is detected to fall below the survival threshold, thereby constructing a complete safety closed loop covering microsecond-level control of a single treatment to long-term adaptation throughout the entire disease course. Attached Figure Description

[0051] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0052] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0054] Example 1:

[0055] Please see Figure 1 A multi-parameter dynamic monitoring intelligent decision-making system for CAP treatment of nephropathy and skin complications, including:

[0056] The multimodal vital signs acquisition module is configured to perform non-invasive monitoring of local tissues of the target object, acquire multidimensional physiological characterization signals, and simultaneously acquire periodic metabolic interference data containing dialysis time sequence information.

[0057] The tolerance threshold inversion module is configured to invert the local tissue redox tolerance threshold based on multidimensional physiological characterization signals and using a preset tissue redox tolerance calculation model.

[0058] The dynamic interference compensation module is configured to identify the phase of the current metabolic cycle based on periodic metabolic interference data, and accordingly perform dynamic baseline correction on the local tissue redox tolerance threshold to generate a real-time safety boundary threshold.

[0059] The adaptive policy generation module is configured to use the real-time safety boundary threshold as a state constraint and use a deep reinforcement learning algorithm to optimize the relationship between treatment dose and tissue damage, thereby generating candidate treatment control strategies.

[0060] The risk boundary control module is configured to calculate the expected oxidative stress level corresponding to the candidate treatment control strategy and execute the safety judgment logic: if the expected oxidative stress level is greater than or equal to the real-time safety boundary threshold, then the blocking or downgrading operation is executed to generate the final treatment instruction; if the expected oxidative stress level is less than the real-time safety boundary threshold, then the candidate treatment control strategy is output as the final treatment instruction.

[0061] In this embodiment, the system addresses the core pain point of high metabolic vulnerability and drastic fluctuations in physiological environment of the skin tissue of patients with chronic kidney disease during dialysis cycles by constructing a closed-loop control circuit.

[0062] The multimodal vital signs acquisition module is configured to perform non-invasive monitoring of local tissues of the target object, acquire multidimensional physiological characterization signals, and simultaneously acquire periodic metabolic interference data containing dialysis timing information. This module aims to construct a holographic physiological view of the target tissue, not only capturing the microenvironmental state of the affected area and surrounding normal skin through a contact probe, but also synchronizing key metabolic background parameters such as dialysis start time, duration, and ultrafiltration volume through the hospital information system interface.

[0063] The system calls the tolerance threshold inversion module, which is configured to invert the local tissue redox tolerance threshold based on multidimensional physiological characterization signals and using a preset tissue redox tolerance calculation model. This step quantifies the invisible tissue latent blood bar, that is, the benchmark value of the maximum oxidative stress that the local tissue can withstand at the current moment.

[0064] Based on this, the dynamic interference compensation module is configured to identify the phase of the current metabolic cycle based on periodic metabolic interference data, and dynamically correct the local tissue redox tolerance threshold accordingly to generate a real-time safety boundary threshold. This module eliminates physiological deception caused by the dialysis cycle by calculating correction coefficients, and prevents the tissue tolerance barrier from being breached by using conventional standard doses when the patient is in a period of high toxin and low tolerance.

[0065] The adaptive policy generation module is configured to use the real-time safety boundary threshold as a state constraint and use a deep reinforcement learning algorithm to optimize the relationship between treatment dose and tissue damage, generating candidate treatment control strategies. This module searches for the optimal solution within the safety boundary and performs multi-objective optimization between treatment dose and potential tissue damage.

[0066] The risk boundary control module is configured to calculate the expected oxidative stress level corresponding to the candidate treatment control strategy and execute the safety judgment logic: in response to the expected oxidative stress level being greater than or equal to the real-time safety boundary threshold, a blocking or downgrading operation is performed to generate the final treatment instruction; in response to the expected oxidative stress level being less than the real-time safety boundary threshold, the candidate treatment control strategy is output as the final treatment instruction.

[0067] This embodiment, through the coordinated operation of the above modules, can detect in real time the severe physiological fluctuations caused by dialysis in patients with chronic kidney disease, and improve the control precision of treatment dosage from the experience level to the physiological response level, effectively avoiding the risk of iatrogenic skin necrosis.

[0068] Example 2:

[0069] Methods for obtaining multidimensional physiological representation signals include:

[0070] Real-time acquisition of skin impedance spectrum data, microcirculation blood perfusion data, and local surface temperature data of the target area;

[0071] Frequency response analysis was performed on skin impedance spectroscopy data to extract the stratum corneum water content characteristics;

[0072] Temporal fluctuation analysis was performed on microcirculatory blood perfusion data to extract vasomotor response characteristics;

[0073] Combining the characteristics of stratum corneum water content, vasomotor response, and local surface temperature data constitutes a multidimensional physiological characterization signal;

[0074] This embodiment is a further refinement of the method for obtaining multidimensional physiological characterization signals based on Embodiment 1. The method for obtaining multidimensional physiological characterization signals includes: real-time acquisition of skin impedance spectrum data, microcirculation blood perfusion data, and local surface temperature data of the target area; frequency response analysis of the skin impedance spectrum data to extract the stratum corneum water content characteristics.

[0075] Temporal fluctuation analysis was performed on microcirculation blood perfusion data to extract vasomotor response characteristics; the stratum corneum water content characteristics, vasomotor response characteristics and local surface temperature data were combined to form a multidimensional physiological characterization signal.

[0076] The system uses a multi-frequency bioelectrical impedance analyzer to collect skin impedance spectrum data in real time, laser speckle contrast imaging technology to collect microcirculation blood perfusion data, and an infrared thermal imaging sensor to collect local surface temperature data. The system performs a feature extraction step to analyze the frequency response of the skin impedance spectrum data and extract the stratum corneum water content characteristics.

[0077] Specifically, the impedance modulus ratio is extracted between the low-frequency value (set to 1kHz in this embodiment, reflecting extracellular fluid impedance) and the high-frequency value (set to 100kHz in this embodiment, reflecting total fluid impedance). To ensure the consistency and computability of the feature, this embodiment explicitly defines this feature. The calculation formula is: ,in This represents the complex impedance modulus at the corresponding frequency; since the higher the water content of the stratum corneum, the smaller the difference in impedance between high and low frequencies, this ratio... It will exhibit a highly nonlinear positive correlation with the hydration state of the stratum corneum;

[0078] Simultaneously, time-domain fluctuation analysis was performed on the microcirculation blood perfusion data to extract vasomotor response characteristics. Specifically, the system employs time-domain digital filtering technology to replace traditional frequency-domain transformation, strictly meeting the requirements of time-domain fluctuation analysis. The acquired raw microcirculation blood perfusion data is defined as a discrete-time series. The sampling frequency is set to For example, 20Hz; construct two sets of parallel fourth-order Butterworth bandpass filters, whose passband ranges are set to the endothelial cell metabolic activity frequency band and the neurogenic activity frequency band, respectively. In this embodiment, the endothelial cell metabolic activity frequency band is set to... The frequency band of neurogenic activity is set to The filtering process is implemented using a recursive difference equation, the formula of which is:

[0079]

[0080] Among them, the filter coefficients Based on the above cutoff frequency and sampling frequency The frequency band of endothelial cell metabolic activity was pre-calculated using the bilinear transform method; this was considered to be... It has a long cycle characteristic; the system takes a period of time to execute after startup and before officially outputting treatment commands. The system performs a pre-acquisition buffering step; during this period, the filter output does not participate in feature calculation to eliminate the transient response error of the filter; after buffering, the system uses a circular buffer for subsequent real-time streaming computation; the real-time acquired microcirculatory blood perfusion time-domain sequence is passed through the above filters to obtain the time-domain fluctuation components in a specific frequency band. and That is, the corresponding discrete sequence and Continuous representation;

[0081] The mean square value of each component is calculated using a sliding window as a representation of the energy in the time domain. and The unit of microcirculatory blood flow signal is the perfusion unit, which is defined as the product of the number of moving red blood cells and the average moving velocity per unit tissue volume. It is a relative physical quantity that reflects the microcirculatory perfusion.

[0082] To combine the energy of these two frequency bands into a single scalar feature and eliminate dimensional differences, this embodiment uses a logarithmic energy weighted sum to calculate the vasomotor response characteristics. To strictly adhere to the dimensionless principle of logarithmic operations, a unit reference energy is introduced here. The mean square value is normalized. This reference energy... The unit of measurement is the reference for the injected fluctuation energy. The formula is revised as follows:

[0083]

[0084] in, To prevent the logarithmic singularity from being a tiny constant, this eigenvalue It directly characterizes the overall activity of microvessels under autonomous regulation; it combines the characteristics of stratum corneum water content, vasomotor response, and local surface temperature to form a multidimensional physiological characterization signal.

[0085] In this embodiment, the formula for constructing multidimensional physiological representation signals is as follows:

[0086]

[0087] in, The characteristics of stratum corneum water content are derived from impedance spectrum frequency response analysis and are dimensionless.

[0088] Vasomotor response characteristics are derived from time-domain filtering and energy integration of blood perfusion data, and are dimensionless pure numbers after normalization and logarithmic processing.

[0089] Local surface temperature, derived from an infrared sensor, is measured in degrees Celsius.

[0090] Temperature change rate, derived from the time-domain derivative of temperature data, is expressed in degrees Celsius per second. This embodiment, by fusing impedance, blood flow, and temperature information, can indirectly reflect the barrier-circulation-metabolism coupling state of the patient's skin. In particular, by introducing vasomotor response characteristics analyzed by time-domain energy, it can keenly capture the loss of regulatory capacity caused by microvascular lesions, providing high-dimensional physical feature support for subsequent inversion of tissue tolerance.

[0091] Example 3:

[0092] Methods for inverting the redox tolerance threshold of local tissues include:

[0093] A pre-built multilayer perceptron regression model is configured to take multidimensional physiological characterization signals as input and map them to a predefined redox buffer capacity feature space.

[0094] Calculate the vector projection of multidimensional physiological representation signals in the feature space;

[0095] The magnitude of the vector projection is determined as the initial tolerance capacity at the current moment;

[0096] Based on the initial tolerance capacity and combined with the preset tissue cell viability decay data curve, the local tissue redox tolerance threshold is calculated.

[0097] This embodiment is a further refinement of the method for retrieving the redox tolerance threshold of local tissues based on Embodiment 2. Since the buffering capacity of tissue cells to oxidative stress cannot be directly measured online, this embodiment adopts a feature space projection strategy. The method for retrieving the redox tolerance threshold of local tissues includes: pre-constructing a multilayer perceptron regression model, configured to take multidimensional physiological characterization signals as input and map them to a predefined redox buffering capacity feature space.

[0098] To ensure the feasibility of the model, the specific topology of this multilayer perceptron is configured as follows: the input layer contains 4 neurons, corresponding to the 4 components of the multidimensional physiological representation signal; the first hidden layer contains 64 neurons, using the ReLU activation function; the second hidden layer contains 32 neurons, using the ReLU activation function; the output layer, i.e., the feature mapping layer, contains 8 neurons, using the Linear activation function; this model... Instead of directly outputting the final scalar, it acts as a feature extractor.

[0099] To imbue the feature space with explicit physiological meaning, the model is pre-trained using supervised contrastive loss based on a clinical dataset containing physiological signals and in vitro antioxidant measurements; the logic for constructing this mapping relationship is as follows:

[0100] The in vitro antioxidant values ​​used as labels here are derived from subcutaneous interstitial fluid samples obtained using microdialysis technology from the target area. When constructing the dataset, local ISF was simultaneously collected using implanted microprobes, and FRAP was measured immediately under anaerobic conditions to obtain the true chemical concentration values ​​that directly reflect the redox buffering capacity of the tissue's extracellular microenvironment. To address the issue of in vitro measurements, typically... To address the difference between pH 3.6 and the in vivo physiological environment, this embodiment introduces an in vivo-in vitro activity mapping function. Preprocess the raw FRAP values;

[0101] This function is derived based on a pre-conducted paired experiment of in vitro tissue homogenate and ISF, and its specific form is as follows:

[0102]

[0103] in, This refers to the tissue redox tolerance threshold label used for training. For temperature sensitivity coefficient, The pH correction factor; in this embodiment, the temperature sensitivity coefficient. The value was determined through in vitro pig skin experiments and taken as an example. Used to compensate for the effects of temperature changes on antioxidant enzyme activity; pH ​​correction factor Based on the Nernst equation correction term set as This corresponds to the potential calibration value at pH 7.4 under normal physiological conditions. Through this tag construction method based on microdialysis and environmental correction, a quantitative anchor point between non-invasive multidimensional physiological characterization signals and the actual biochemical tolerance in vivo is forcibly established, enabling the model to learn the concentration level of antioxidants hidden behind impedance, blood flow and temperature signals.

[0104] Specifically, constructing training sample pairs and its corresponding FRAP tag Set a similarity threshold ;like These are marked as positive sample pairs, i.e. Otherwise, it is a negative sample pair, i.e. Define the loss function ,in denoted as Euclidean distance in the feature space. The default boundary is defined; by minimizing this loss, high-tolerance samples and low-tolerance samples are separated along a specific principal axis in the feature space.

[0105] Specifically, the penultimate layer of the model, i.e., the fully connected layer, outputs a high-dimensional feature vector. This vector represents the coordinates of the multidimensional physiological characterization signal in the redox buffer capacity feature space; the vector projection of the multidimensional physiological characterization signal in the feature space is calculated; the magnitude of the vector projection is determined as the initial tolerance capacity at the current moment; the system has a preset health baseline unit vector. This vector is obtained by calculating the geometric center of the feature vector of healthy people's skin data in the feature space and then performing normalization.

[0106] To address potential issues such as unclear acquisition steps and circular dependencies during implementation, this embodiment provides a detailed explanation. The standardized construction process is as follows: Based on the aforementioned contrastive loss function, train and freeze the parameters of the multilayer perceptron model to ensure the feature extractor... The mapping relationship was fixed; a group of volunteers meeting the health criteria were selected, defined as: age 18-35 years, glomerular filtration rate (eGFR) > 90 mL / min / 1.73 m², and no lesions on the back skin. Their multidimensional physiological characteristic signals were collected. ;

[0107] Will Input a frozen MLP model to obtain the corresponding feature vector set. ; Calculate the geometric center of the set and normalize it, i.e. This process establishes As a pre-defined constant independent of the real-time inference process of patients, it solves the temporal dependency problem between model training and benchmark acquisition; compared with principal component analysis selecting the direction of maximum variance, using the geometric center can more accurately represent the core position of the ideal health state and avoid benchmark deviation caused by individual differences.

[0108] System Calculation exist Projection components in the direction The projection component Length of the module Defined as initial tolerance capacity Physically, it represents the effective buffer reserve of the current tissue relative to an ideal healthy state; based on the initial tolerance capacity and combined with the preset tissue cell viability decay data curve, the local tissue redox tolerance threshold is calculated; the calculation formula is as follows:

[0109]

[0110] in, Local tissue redox tolerance threshold, in moles per liter This is to ensure dimensional consistency when comparing with oxidative stress concentrations in the future.

[0111] Characteristic-concentration spatial mapping coefficient, in moles per liter. This coefficient is used to project the dimensionless characteristic modulus. This is converted into a physical concentration value; in this embodiment, the average antioxidant capacity of interstitial fluid in the skin tissue of healthy adults is selected as a physiological reference value and set as follows: Due to the health baseline vector Let be a unit vector, and let its projection magnitude in its own direction be... Therefore, set This serves as a scaling factor for mapping the feature space distance back to the chemical concentration space.

[0112] By using health baseline vector The corresponding ideal tolerable capacity is anchored to the physiological reference value for calibration.

[0113] Initial tolerance capacity, derived from vector projection magnitude. Dimensionless;

[0114] Cell viability decay coefficient, the value of which is obtained from tissue cell viability decay data curves. The search revealed that the curve fits the non-linear relationship between the activity of cellular ROS scavenging enzymes and the decrease in dialysis age, for example... ;

[0115] This embodiment uses explicit vector projection calculations, rather than direct black-box regression, to ensure that the robustness assessment has clear geometric interpretation and can effectively distinguish between effective buffer components and noise components in the signal.

[0116] Example 4:

[0117] Methods for dynamically correcting the redox tolerance threshold of local tissues include:

[0118] Obtain the timestamp of the most recent dialysis end from the periodic metabolic disturbance data;

[0119] Calculate the time difference between the current time and the timestamp of the most recent dialysis end to determine the specific time period during the current interdialysis interval;

[0120] Based on specific time periods, the corresponding toxin accumulation coefficient and water load coefficient are retrieved from the preset sawtooth metabolic fluctuation data model.

[0121] A nonlinear attenuation factor was constructed using the toxin accumulation coefficient and the water load coefficient.

[0122] The corrected real-time safety boundary threshold is obtained by multiplying the local tissue redox tolerance threshold by a nonlinear decay factor.

[0123] This embodiment is a further refinement of the method for dynamically correcting the redox tolerance threshold of local tissues based on Embodiment 3. In response to the fluctuations in tolerance caused by toxin accumulation during dialysis cycles, this embodiment introduces a time-toxin coupling correction mechanism. The method for dynamically correcting the redox tolerance threshold of local tissues includes: obtaining the most recent dialysis end timestamp from the periodic metabolic interference data.

[0124] Calculate the time difference between the current time and the most recent dialysis end time stamp to determine the specific time period currently in the interdialysis interval; this step aims to locate the patient's current physiological clock phase; based on the specific time period, retrieve the corresponding toxin accumulation coefficient and water load coefficient from a pre-set sawtooth metabolic fluctuation data model; this model reflects the exponential or linear accumulation trend of uremic toxins in the body over time; construct a nonlinear decay factor using the toxin accumulation coefficient and water load coefficient; multiply the local tissue redox tolerance threshold by the nonlinear decay factor to obtain the corrected real-time safety boundary threshold; the correction calculation formula is as follows:

[0125]

[0126]

[0127] in, : Nonlinear decay factor, derived from time difference calculation, dimensionless;

[0128] The toxin accumulation sensitivity coefficient is derived from the fitting of patients' historical biochemical indicators, and the unit is the reciprocal of the hourly rate. The specific fitting method is as follows: Peak serum toxin concentrations before each dialysis session within the past month were selected to establish time-series data. Linear regression using the least squares method was then performed to obtain the slope of the regression line. ,unit: Set a normalized reference rate constant. In this embodiment, ,calculate As Value; explicitly stated here. With Same physical dimensions, to ensure Only the reciprocal of the time dimension is retained, thus making the item It is a dimensionless numerical value, which satisfies the dimension consistency requirement of the addition formula;

[0129] : Nonlinear weighting coefficient for water load, derived from interdialysis weight gain rate data, in reciprocal units per square hour. The fitting process involves collecting continuous monitoring data of patients' weight between dialysis periods and fitting the data to generate a quadratic growth curve. Take the coefficient of the quadratic term The unit is Multiply by the preset quality normalization coefficient ,Right now As Value; explicitly stated here. A physical quantity with the inverse of mass is used to eliminate the mass dimension from weight gain data, ensuring... The term is converted into a dimensionless value to prevent dimensional errors caused by directly adding the physical mass to the dimensionless constant 1.

[0130] coefficient The value is derived from the reference standardization process in the biological allometric growth law; its value corresponds to the mass of the standard reference body. The reciprocal of, that is The biophysical significance of introducing this coefficient lies in representing the individualized acceleration of weight gain. Normalized to specific load intensity relative to standard physiological capacity; this makes It is no longer merely a mathematical elimination coefficient, but rather a characterization of the water load impact intensity under unit physiological carrying capacity, thus giving the formula terms... A clear pathophysiological explanation is that it reflects the cumulative effect of water retention relative to the standard human tolerance limit;

[0131] The time elapsed since the last dialysis session, derived from the system clock, is expressed in hours.

[0132] This embodiment innovatively uses time as a key dimension for measuring physiological vulnerability. By introducing a sawtooth metabolic fluctuation model, the system can identify the extremely dangerous physiological window at the end of the dialysis interval and automatically lower the safety threshold to prevent high-intensity treatment when the patient's physiological defenses are weakest.

[0133] Example 5:

[0134] Methods for generating candidate treatment control strategies include:

[0135] Construct a reinforcement learning state space, which includes real-time safety boundary thresholds, current wound healing rate indicators, and multidimensional physiological representation signals.

[0136] Construct an action space, which includes plasma jet intensity, gas flow rate, and duration of a single action;

[0137] An asymmetric reward function is designed to provide a reward value positively correlated with the wound healing rate when the expected oxidative stress level is below the real-time safety boundary threshold; and to provide an exponentially increasing penalty value when the expected oxidative stress level is above or equal to the real-time safety boundary threshold.

[0138] The current state is input into the pre-trained policy network, which outputs the action probability distribution in the action space, and generates candidate treatment control strategies based on the action probability distribution.

[0139] This embodiment is a further specification of the method for generating candidate treatment control strategies based on embodiment 4. This embodiment uses the proximal policy optimization algorithm in deep reinforcement learning to solve the multi-objective conflict problem. The method for generating candidate treatment control strategies includes: constructing a reinforcement learning state space, which includes a real-time safety boundary threshold, a current wound healing rate index, and multi-dimensional physiological representation signals.

[0140] Simultaneously, an action space is constructed, which includes plasma jet intensity, gas flow rate, and single-action duration. Based on this, an asymmetric reward function is designed, where a reward value positively correlated with the wound healing rate is given when the expected oxidative stress level is lower than the real-time safety boundary threshold; and an exponentially increasing penalty value is given when the expected oxidative stress level is higher than or equal to the real-time safety boundary threshold. This asymmetric design forces the algorithm to have risk-averse characteristics. The current state is input into a pre-trained policy network, which outputs the action probability distribution in the action space, and candidate treatment control strategies are generated based on the action probability distribution.

[0141] Specifically, the policy network in this embodiment is built based on the Actor-Critic architecture. The Actor network, which generates the policy, includes an input layer and two fully connected hidden layers with 64 and 32 neurons respectively. The activation functions are Tanh and Gaussian distribution parameters of the output actions. The Critic network, which evaluates the value, shares some parameters with the Actor network and is used to calculate the value function of the current state. ;

[0142] To ensure the consistency of input data distribution and network convergence, a standardized state vector construction step is performed before inputting the policy network: This involves setting the real-time safety boundary threshold... Normalization is performed by dividing by a preset historical peak value, and the multidimensional physiological characteristic signal is then processed. Z-score standardization involves subtracting the mean and dividing by the standard deviation. Specifically, it uses the global mean vector calculated from the training set used during the pre-training phase of the policy network. With global standard deviation vector The calculation formula is: ,in This represents element-wise division and the wound healing rate index. The data is truncated and mapped to the interval [-1, 1], and then concatenated into a one-dimensional feature vector. Input network; reward function The design logic is as follows:

[0143]

[0144] in, : Wound healing rate index; the specific method for obtaining this index is as follows: the system controls the camera to acquire high-definition visible light images of the affected area, uses a pre-trained U-Net semantic segmentation network to extract the wound area mask, and calculates the current wound pixel area. Read the wound area recorded in the previous treatment cycle. If this is the first treatment, then... Calculate the relative healing rate ;like If the wound expands, it will be truncated to 0 or a negative value as a penalty;

[0145] The expected oxidative stress level is derived from model predictions and is expressed in moles per liter. In this embodiment, the model prediction specifically uses a physical dose estimation formula. Perform calculations, where The dose conversion coefficient, whose physical meaning is the rate of reactive oxygen species concentration generation per unit intensity per unit time;

[0146] In this embodiment, The determination method is as follows: measure the molar flow rate of reactive oxygen species generated by the equipment under unit intensity. And estimate the effective penetration volume of the plasma jet in skin tissue. Calculate the ratio, in units of L. In this embodiment, the value is 0.05, which is used to convert the device parameters into a concentration response within the tissue.

[0147] The jet intensity during the current action. The duration of a single action is used to simulate the therapeutic dose in a reinforcement learning training environment.

[0148] Real-time security boundary thresholds are derived from the dynamic correction module;

[0149] : Penalty sensitivity normalization coefficient, in units of liters per mole (L / mol), used to eliminate the dimensional influence of the exponential term and set the penalty gradient. In this embodiment, it is set to... ;

[0150] The algorithm adjusts the weighting factor. Here, the term "factor" is used to explicitly define its attributes as a hyperparameter of the algorithm. This is intended to strictly distinguish it from the toxin accumulation sensitivity coefficient and the water load nonlinear weighting coefficient, which have specific biophysical meanings and dimensions, and to prevent conceptual confusion during reading.

[0151] Used to regulate positive rewards Used to adjust the penalty for exceeding the boundary, and Much larger In this embodiment, to ensure the strong dominance of security constraints, the following is adopted: ,Pick ;

[0152] This embodiment uses an asymmetric reward function to enable the reinforcement learning agent to actively explore parameter combinations that maximize the healing rate within the safety boundary. However, once it approaches the safety threshold, its exploration behavior is suppressed by exponential penalties, thus ensuring the conservatism and safety of the strategy at the algorithm level.

[0153] Example 6:

[0154] Methods for safety-tailoring candidate treatment control strategies include:

[0155] Based on the plasma jet intensity and single-action duration in the candidate treatment control strategy, the expected dose of exogenous reactive oxygen species is calculated.

[0156] The expected dose of exogenous reactive oxygen species is compared with the real-time safety boundary threshold.

[0157] If the dose of exogenous reactive oxygen species is less than the real-time safety boundary threshold, the candidate treatment control strategy will be directly output as the final treatment instruction.

[0158] If the exogenous reactive oxygen species dose is greater than or equal to the real-time safety boundary threshold, the plasma jet intensity remains unchanged, and the duration of a single treatment is reduced until the recalculated exogenous reactive oxygen species dose is less than the real-time safety boundary threshold. The reduced strategy is then output as the final treatment instruction.

[0159] This embodiment further specifies the safety-tailoring method for candidate treatment control strategies based on Embodiment 5; this is the last hard constraint defense before system output; the method for safety-tailoring candidate treatment control strategies includes: calculating the expected introduced exogenous reactive oxygen species dose based on the plasma jet intensity and single-action duration in the candidate treatment control strategy; this calculation is based on a preset physical dose model, and it is clear that the expected introduced exogenous reactive oxygen species dose is the expected oxidative stress level. In this embodiment, to emphasize its property as a controlled physical quantity, it is symbolized as The expected dose of exogenous reactive oxygen species is compared with the real-time safety boundary threshold.

[0160] Based on this, the system makes the following decisions: if the exogenous reactive oxygen species (ROS) dose is less than the real-time safety boundary threshold, the candidate treatment control strategy is directly output as the final treatment command; if the exogenous ROS dose is greater than or equal to the real-time safety boundary threshold, the plasma jet intensity remains unchanged, and the duration of a single treatment is reduced until the recalculated exogenous ROS dose is less than the real-time safety boundary threshold, at which point the reduced strategy is output as the final treatment command; the reduction formula is as follows:

[0161]

[0162] in, : Reduced duration of a single action, in seconds;

[0163] The original duration of a single action is derived from the policy network output and is in seconds.

[0164] Real-time safety boundary threshold, derived from the dynamic interference compensation module, is expressed in moles per liter;

[0165] The expected volumetric molar concentration of exogenous reactive oxygen species introduced is equivalent to the expected level of oxidative stress. To ensure that it matches the threshold The dimensions are consistent, and here it is defined as the amount of substance per unit volume. The specific calculation formula is as follows: The unit is moles per liter;

[0166] in, The intensity of the plasma jet. The dose conversion coefficient, which includes a tissue action volume normalization factor, is set to 0.05 in this embodiment, and its physical dimensions are set such that the calculation results... The unit of concentration is used here. The dose conversion factor defined in the foregoing embodiments; The calibration method is as follows: Measure the molar flow rate of reactive oxygen species generated by the equipment under unit intensity. And estimate the effective penetration volume of the plasma jet in skin tissue. Unit: L, Calculate the ratio get;

[0167] Safety margin coefficient, a preset constant, dimensionless, is set to 0.9 in this embodiment to retain a 10% safety buffer;

[0168] This embodiment ensures that no matter how the AI ​​algorithm explores, the final dose to the patient is always physically limited to a safe range, and adopts a strategy of reducing the dose while maintaining the strength.

[0169] Performing blocking or degradation operations also includes non-monotonic decision logic:

[0170] Determine whether the real-time safety boundary threshold is lower than the preset critical crash threshold;

[0171] If the real-time safety boundary threshold is lower than the critical collapse threshold, the wound healing rate index is ignored, the plasma jet intensity parameter in the candidate treatment control strategy is forcibly set to zero, and a final treatment instruction is generated to pause treatment and issue alarm data.

[0172] If the real-time safety boundary threshold is higher than or equal to the critical crash threshold, then continue with the safety trimming step.

[0173] This embodiment introduces non-monotonic decision logic; the execution of blocking or degradation operations also includes non-monotonic decision logic: determining whether the real-time security boundary threshold is lower than a preset critical crash threshold; the critical crash threshold here... It is not an arbitrary empirical value, but a physical boundary that characterizes the irreversible necrosis of tissue;

[0174] In this embodiment, The data is obtained as follows: Based on the patient's interdialysis monitoring data over the past 3 months, the historical average of the safety boundary threshold is calculated. and will Set as 15% of This setting is based on the fact that when the tissue's tolerance falls below 15% of the normal baseline, the stability of the cell membrane potential cannot be maintained, and any exogenous stimulus may induce an avalanche effect.

[0175] The system executes the following branch logic: In response to the real-time safety boundary threshold being lower than the critical collapse threshold, the wound healing rate index is ignored, the plasma jet intensity parameter in the candidate treatment control strategy is forcibly set to zero, and a final treatment instruction is generated to pause treatment and issue an alarm. This step simulates the refusal judgment of an experienced physician. At this time, the system determines that the patient is in a metabolic storm period and must be given priority for systemic medical intervention rather than local physical therapy. In response to the real-time safety boundary threshold being higher than or equal to the critical collapse threshold, the safety pruning step continues to be executed.

[0176] This embodiment establishes a dynamic circuit breaker mechanism based on statistical historical benchmarks, which effectively avoids serious medical risks.

[0177] The system also includes an evolutionary analysis and model update module, used for:

[0178] After the final treatment instruction is executed, the changing trends of multidimensional physiological representation signals are continuously monitored;

[0179] Calculate the deviation between the actual tissue response and the expected response after treatment;

[0180] If the deviation exceeds the preset tolerance range, the periodic metabolic disturbance data and multidimensional physiological characterization signals at the current moment are extracted and stored in the experience playback pool.

[0181] The digital twin model of tissue redox tolerance was fine-tuned online using data from the empirical replay pool to update the local tissue redox tolerance threshold inversion parameters at subsequent time points.

[0182] This embodiment is a further improvement on the system based on embodiment 1, adding self-evolution capability; it also includes an evolution analysis and model update module, which is used to continuously monitor the changing trend of multidimensional physiological characterization signals after the final treatment instruction is executed; and calculate the deviation between the actual tissue response and the expected response after treatment;

[0183] If the deviation exceeds the preset tolerance range, the periodic metabolic disturbance data and multidimensional physiological characterization signals at the current moment are extracted and stored in the experience replay pool; the data in the experience replay pool are used to fine-tune the digital twin model of tissue redox tolerance online to update the local tissue redox tolerance threshold inversion parameters at subsequent moments.

[0184] To make the actual tissue response calculable, this embodiment constructs an explicit mathematical mapping; the slope of the local surface temperature rise within 15 minutes after treatment is extracted. variance of microcirculatory blood flow signal Construct reaction feature vector The actual tolerance capacity consumed is calculated using a pre-defined stimulus-response phenomenological model. Its formula is:

[0185]

[0186] in, It is a preset weight vector. This is the bias term; this parameter was calibrated through the following in vitro and in vivo experiments: 30 SD rats were selected to construct a dorsal skin model, divided into 5 groups, and treated with different gradients of CAP doses. The physical feature vectors after each treatment were recorded simultaneously. That is, temperature slope With blood flow variance ;

[0187] Immediately after treatment, a biopsy was performed. The ratio of necrotic tissue volume was determined using TTC staining, and the ROS concentration in the interstitial fluid was measured using ELISA. The weighted normalized sum of these two measurements was defined as the experimental true value of the actual tolerable volume consumed. ; Use the multiple linear regression algorithm on the dataset Perform fitting and solve the regression equation The coefficients; under the experimental conditions of this embodiment, regression analysis shows the goodness of fit of the model. The calibration calculation yielded the following results. , ;

[0188] It should be noted that, in order to ensure the above formula holds true in terms of physical dimensions, the weight vector... The elements in the are not pure scalars, but rather conversion factors carrying physical units; specifically, corresponding to Units are The first weighted component has The unit corresponding to Units are The second weight component has The units; this dimensional design ensures that the physical characterization signal and the chemical concentration unit are consistent. The legitimate mapping between them avoids numerical calculations with unclear physical meaning;

[0189] against The question of whether using a relative unit leads to ambiguity in physical meaning is addressed in this embodiment, which states that in this system, This does not refer to a general dimensionless value, but is strictly defined as an instrument characteristic reading within this specific hardware system; the unit of the weighted component. In essence, it constitutes the device-tissue conversion operator, whose physical meaning is the tissue ROS concentration increment corresponding to the variance of blood flow fluctuation detected by each unit of instrument;

[0190] Through the aforementioned rigorously controlled SD rat biopsy calibration experiment, this operator anchored the relative fluctuation value of optics to the absolute chemical concentration value, thereby constructing a closed-loop and self-consistent dimensional system within the system, ensuring... Compared with the predicted value exist Precise comparability across dimensions overcomes the limitations of universal unit definitions; the so-called deviation is... ,in This refers to the tolerance depletion predicted by the pre-treatment model, specifically the expected dose of exogenous reactive oxygen species (ROS) calculated above and corresponding to the final execution command. That is, the expected level of oxidative stress. This serves as a theoretical baseline for tissue tolerance and consumption.

[0191] In response to deviations exceeding a preset tolerance range, the periodic metabolic disturbance data and multidimensional physiological characterization signals at the current moment are extracted and stored in the experience playback pool; simultaneously, the actual tissue response characteristics calculated in the aforementioned steps are... Expected dose And the local tissue redox tolerance threshold output by the tolerance threshold inversion module at the current moment. As a supervisory label, synchronous associated storage is used to construct complete state-reaction sample pairs to solve the technical problem that storing only input data leads to a lack of target values ​​for subsequent model fine-tuning;

[0192] The digital twin model of tissue redox tolerance was fine-tuned online using data from the empirical replay pool to update the local tissue redox tolerance threshold inversion parameters at subsequent time points; the fine-tuning process did not directly use bias. Instead, it constructs a revised monitoring signal to update the logic in a closed loop: based on the inverse proportionality assumption, it explicitly defines... The local tissue redox tolerance threshold extracted from the empirical replay pool at the corresponding time point. That is, the original output value of the model before dynamic baseline correction, and the calculated tolerance threshold label. ,in The relaxation factor is set to 0.5;

[0193] To establish a complete gradient backpropagation path in the computation graph, the system performs a re-inference step: extracting historical input states from the empirical replay pool. The value is then input into a multilayer perceptron model with the current weights for forward propagation to calculate the tolerance threshold estimate under the current weights. ;

[0194] Define loss function This construction ensures that the loss function is differentiable with respect to the network weights; the gradient is calculated using the stochastic gradient descent (SGD) algorithm. It also backpropagates to update the connection weights in the multilayer perceptron, thereby enabling the model to adapt to the physiological characteristic drift caused by the progression of the patient's disease.

[0195] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-parameter dynamic monitoring intelligent decision-making system for CAP treatment of skin complications in kidney disease, characterized in that, include: The multimodal vital signs acquisition module is configured to perform non-invasive monitoring of local tissues of the target object, acquire multidimensional physiological characterization signals, and simultaneously acquire periodic metabolic interference data containing dialysis time sequence information. The tolerance threshold inversion module is configured to invert the local tissue redox tolerance threshold based on multidimensional physiological characterization signals and using a preset tissue redox tolerance calculation model; the tissue redox tolerance calculation model is as follows: ,in, This refers to the local tissue redox tolerance threshold. These are the feature-concentration space mapping coefficients. Initial tolerance capacity, This is the cell viability decay coefficient; The dynamic interference compensation module is configured to identify the phase of the current metabolic cycle based on periodic metabolic interference data, and accordingly perform dynamic baseline correction on the local tissue redox tolerance threshold to generate a real-time safety boundary threshold. The adaptive policy generation module is configured to use the real-time safety boundary threshold as a state constraint and use a deep reinforcement learning algorithm to optimize the relationship between treatment dose and tissue damage, thereby generating candidate treatment control strategies. The risk boundary control module is configured to calculate the expected oxidative stress level corresponding to the candidate treatment control strategy and execute the safety judgment logic: if the expected oxidative stress level is greater than or equal to the real-time safety boundary threshold, then the blocking or downgrading operation is executed to generate the final treatment instruction; if the expected oxidative stress level is less than the real-time safety boundary threshold, then the candidate treatment control strategy is output as the final treatment instruction. Methods for dynamically correcting the redox tolerance threshold of local tissues include: Obtain the timestamp of the most recent dialysis end from the periodic metabolic disturbance data; Calculate the time difference between the current time and the timestamp of the most recent dialysis end to determine the specific time period during the current interdialysis interval; Based on the specific time period, the corresponding toxin accumulation coefficient and water load coefficient are retrieved from the preset sawtooth metabolic fluctuation data model; the sawtooth metabolic fluctuation data model reflects the exponential or linear accumulation trend of uremic toxins in the body over time. A nonlinear attenuation factor is constructed using the toxin accumulation coefficient and the water load coefficient; the calculation formula for the nonlinear attenuation factor is as follows: ,in, It is a nonlinear decay factor. This represents the toxin accumulation coefficient. This is the water load factor. This refers to the time difference between specific periods of the current interdialysis interval; The corrected real-time safety boundary threshold is obtained by multiplying the local tissue redox tolerance threshold by the nonlinear decay factor.

2. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 1, characterized in that, Methods for obtaining multidimensional physiological representation signals include: Real-time acquisition of skin impedance spectrum data, microcirculation blood perfusion data, and local surface temperature data of the target area; Frequency response analysis was performed on skin impedance spectroscopy data to extract the stratum corneum water content characteristics; Temporal fluctuation analysis was performed on microcirculatory blood perfusion data to extract vasomotor response characteristics; The characteristics of stratum corneum water content, vasomotor response, and local surface temperature are combined to form a multidimensional physiological characterization signal.

3. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 2, characterized in that, Methods for inverting the redox tolerance threshold of local tissues include: A pre-built multilayer perceptron regression model is configured to take multidimensional physiological characterization signals as input and map them to a predefined redox buffer capacity feature space. Calculate the vector projection of the multidimensional physiological representation signal in the feature space; The magnitude of the vector projection is determined as the initial tolerance capacity at the current moment; Based on the initial tolerance capacity and combined with the preset tissue cell viability decay data curve, the local tissue redox tolerance threshold is calculated.

4. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 3, characterized in that, Methods for generating candidate treatment control strategies include: Construct a reinforcement learning state space, which includes real-time safety boundary thresholds, current wound healing rate indicators, and multidimensional physiological representation signals. Construct an action space, which includes plasma jet intensity, gas flow rate, and duration of a single action; An asymmetric reward function is designed where a reward value positively correlated with the wound healing rate is given when the expected oxidative stress level is below the real-time safety boundary threshold; and an exponentially increasing penalty value is given when the expected oxidative stress level is above or equal to the real-time safety boundary threshold. The current state is input into the pre-trained policy network, which outputs the action probability distribution in the action space, and then samples and generates candidate treatment control strategies based on the action probability distribution.

5. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 4, characterized in that, Methods for safety-tailoring candidate treatment control strategies include: Based on the plasma jet intensity and single-action duration in the candidate treatment control strategy, the expected dose of exogenous reactive oxygen species is calculated. The expected dose of exogenous reactive oxygen species is compared with the real-time safety boundary threshold. If the dose of exogenous reactive oxygen species is less than the real-time safety boundary threshold, the candidate treatment control strategy will be directly output as the final treatment instruction. If the exogenous reactive oxygen species dose is greater than or equal to the real-time safety boundary threshold, the plasma jet intensity remains unchanged, and the duration of a single treatment is reduced until the recalculated exogenous reactive oxygen species dose is less than the real-time safety boundary threshold. The reduced strategy is then output as the final treatment instruction.

6. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 5, characterized in that, Performing blocking or degradation operations also includes non-monotonic decision logic: Determine whether the real-time safety boundary threshold is lower than the preset critical crash threshold; If the real-time safety boundary threshold is lower than the critical collapse threshold, the wound healing rate index is ignored, the plasma jet intensity parameter in the candidate treatment control strategy is forcibly set to zero, and a final treatment instruction to pause treatment and issue alarm data is generated. If the real-time security boundary threshold is higher than or equal to the critical crash threshold, then the security trimming step continues.

7. The intelligent decision-making system for multi-parameter dynamic monitoring of skin complications in CAP treatment of kidney disease according to claim 1, characterized in that, It also includes an evolution analysis and model update module, used for: After the final treatment instruction is executed, the changing trends of multidimensional physiological representation signals are continuously monitored; Calculate the deviation between the actual tissue response and the expected response after treatment; If the deviation exceeds the preset tolerance range, the periodic metabolic disturbance data and multidimensional physiological characterization signals at the current moment are extracted and stored in the experience playback pool. The digital twin model of tissue redox tolerance was fine-tuned online using data from the experience replay pool to update the local tissue redox tolerance threshold inversion parameters at subsequent time points.

Images