Precise monitoring and self-adaptive regulation and control system for immunological rejection in solid organ transplantation

Abstract

The invention provides a precise monitoring and self-adaptive regulation and control system for immunological rejection in solid organ transplantation. Relates to the field of artificial intelligence, and the system comprises a real-time data collection module which is used for collecting rejection-related biomarkers, organ function data and environment and treatment data in peripheral blood of a transplanted recipient in real time; the rejection risk prediction module is used for processing multi-dimensional time sequence characteristic data acquired in real time by adopting a long-short term memory network rejection risk prediction model based on layered boundary marginal loss function optimization, and outputting a rejection risk probability P; the intervention strategy generation module is used for generating an individualized intervention strategy through a mixed mask neighbor strategy optimization algorithm according to the current pathological state of the patient; and the intelligent decision regulation and control module is used for executing regulation and control decisions of precise regulation and control of the immunosuppressor and coordinated regulation and control of the organ microenvironment according to the rejection risk probability P and the individualized intervention strategy and performing real-time feedback. The rejection reaction recognition accuracy and intervention effectiveness are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a precise monitoring and adaptive regulation system for immune rejection response in solid organ transplantation. Background Technology

[0002] Solid organ transplantation is a crucial treatment for end-stage organ failure. However, postoperative immune rejection remains a key challenge affecting organ survival and long-term patient prognosis. Current clinical practice primarily relies on biochemical indicators, tissue biopsies, and empirical immunosuppression strategies for rejection assessment and intervention, but this approach has the following significant limitations: 1. Strong lag in assessment: Many rejection reactions occur after organ damage has already occurred, and changes in biochemical indicators lag behind; 2. Inefficient intervention strategies: The use of immunosuppressants is mostly empirically guided, lacking individualized dynamic regulation mechanisms; 3. Lack of predictive risk assessment: There is a lack of proactive predictive models for immune rejection trends, making timely intervention difficult; 4. Sparse and imbalanced labeled data: The distribution of labels in medical data is severely imbalanced, greatly affecting model performance; 5. Complex multi-factor intervention decisions: Intervention strategies often involve discrete and continuous variables and must strictly adhere to medical feasibility.

[0003] Therefore, there is an urgent need to build a high-frequency real-time monitoring, intelligent prediction and judgment and individualized closed-loop regulation system for the management of solid organ transplantation, so as to improve the accuracy of rejection identification and the effectiveness of intervention. Summary of the Invention

[0004] To address the aforementioned issues, this invention aims to tackle the challenges of predictive lag, inadequate intervention, labeling imbalance, and decision-making complexity in the risk management of rejection after solid organ transplantation. It constructs a comprehensive system for precise monitoring and adaptive regulation of rejection, integrating "high-frequency data perception, in-depth risk prediction, and intelligent intervention and control," thereby improving the accuracy of rejection identification and the effectiveness of intervention.

[0005] To achieve the above objectives, the following technical solution is adopted:

[0006] A precise monitoring and adaptive regulation system for immune rejection in solid organ transplantation includes:

[0007] The real-time data acquisition module is used to collect rejection-related biomarkers, organ function data, and environmental and treatment data in the peripheral blood of transplant recipients in real time.

[0008] The exclusion risk prediction module is used to process real-time collected multi-dimensional time-series feature data and output the exclusion risk probability P by adopting a long short-term memory network exclusion risk prediction model based on hierarchical boundary marginal loss function optimization.

[0009] The intervention strategy generation module is used to generate individualized intervention strategies containing discrete intervention actions and continuous control variables based on the patient's current pathological state through a hybrid mask nearest neighbor strategy optimization algorithm.

[0010] The intelligent decision-making and regulation module includes an immunosuppressant infusion unit driven by a PID controller and a temperature / pH linkage regulation unit. It is used to make regulatory decisions on precise regulation of immunosuppressants and coordinated regulation of organ microenvironment based on the rejection risk probability P and the individualized intervention strategy, and to provide real-time feedback.

[0011] Furthermore, the hierarchical boundary marginal loss function used in the exclusion risk prediction module optimizes the training of the long short-term memory network exclusion risk prediction model through the following mechanism:

[0012] Hierarchical dynamic boundary mechanism: The monitoring units are divided according to organ type or patient group, and each unit learns an adaptive dynamic boundary threshold to replace the fixed classification threshold and adapt to the characteristic differences of different transplantation scenarios.

[0013] Confidence-driven bias mechanism: Based on the dispersion of the predicted confidence of samples within each monitoring unit, the bias of positive and negative samples is dynamically calculated to enhance the model's ability to identify suspected exclusion samples whose predicted confidence is in the boundary region.

[0014] Marginal filtering mechanism: During training, high-confidence samples that have reached the saturation state are filtered out by preset marginal parameters, and training resources are concentrated on learning low-confidence boundary samples.

[0015] Furthermore, the hybrid mask nearest neighbor strategy optimization algorithm for generating discrete intervention actions includes:

[0016] Based on the patient's real-time pathological state, the original probability distribution of discrete actions is output through a neural network;

[0017] A medical knowledge-driven action masking mechanism is used to logically block actions that violate clinical contraindications or safety guidelines.

[0018] The unmasked actions are subjected to probability normalization to output discrete intervention action decisions that are clinically feasible.

[0019] Furthermore, the hybrid mask nearest neighbor strategy optimization algorithm generates continuous control variables including:

[0020] Based on the patient's real-time pathological state, the neural network activates and outputs non-negative, biologically meaningful shape parameters through the SoftPlus function. ;

[0021] Using the shape parameters Construct a Beta probability distribution model;

[0022] Continuous intervention values ​​are generated by sampling from the distribution and used as continuous control intensity coefficients for the intelligent decision-making and control module to adjust the drug infusion rate; wherein, the continuous control intensity coefficients reflect whether the reinforcement learning strategy suggests that the current control intensity is mild, standard, or enhanced.

[0023] Furthermore, the closed-loop optimization process of the hybrid mask nearest neighbor strategy optimization algorithm includes:

[0024] Execute the generated discrete actions and continuous control values, and observe whether a rejection reaction is induced;

[0025] Collect biomarker feedback signals and calculate the reward for improving rejection risk;

[0026] A proximal strategy is adopted to optimize the updated decision network parameters and improve the total reward function.

[0027] Furthermore, the medical knowledge-driven action masking mechanism specifically involves: predefining or calling in real-time a medical knowledge base containing organ-specific contraindications, drug interaction rules, and patient physiological state restrictions; matching the knowledge base rules based on the patient's current real-time monitoring data, and assigning a logical mask prohibiting the execution of intervention actions that violate the rules.

[0028] Furthermore, the intelligent decision-making and control module includes:

[0029] An immunosuppressant infusion unit is used to finely adjust the infusion rate of the immunosuppressant in real time using a PID control algorithm based on the rejection risk probability P and the individualized intervention strategy.

[0030] A temperature / pH linkage control unit is used to automatically initiate the linkage regulation of organ perfusion fluid temperature and pH when the detected rejection risk probability P reaches a preset high-risk threshold, including:

[0031] The temperature of the organ perfusion fluid is precisely lowered from the basic setting value to the preset low temperature setting value range;

[0032] The pH value of the perfusion solution is monitored in real time. When it deviates from the preset physiological neutral range, the buffer solution infusion is automatically started to precisely adjust the pH value to the target range.

[0033] Both temperature and pH controls are achieved through a closed-loop control system, enabling rapid response and high-precision maintenance.

[0034] Furthermore, the immunosuppressant infusion unit employs a PID control algorithm to precisely adjust the infusion rate of the immunosuppressant in real time, specifically including:

[0035] The target rejection risk probability value is set as the benchmark for the PID control algorithm;

[0036] Calculate the real-time error between the current predicted rejection risk probability P and the target risk threshold;

[0037] The real-time error is input into the PID controller, and the original output of the PID controller is multiplied by the real-time control intensity coefficient a(t) generated by the intelligent control decision module to obtain the final immunosuppressant infusion rate adjustment amount u(t).

[0038] Based on the immunosuppressant infusion rate adjustment amount u(t), a high-precision peristaltic pump is driven to adjust the infusion rate in a stepwise manner with a preset minimum adjustment granularity, so that the blood concentration of immunosuppressant in the patient is stably maintained within a predefined therapeutic window range.

[0039] Furthermore, the intelligent decision-making and control module includes:

[0040] When the exclusion risk probability P < 0.2, it is determined to be a low-risk state, and basic monitoring is maintained;

[0041] When 0.2 ≤ the rejection risk probability P < 0.5, it is determined to be a medium-risk state, and enhanced monitoring is initiated: the frequency of immune indicator detection is increased to once per hour;

[0042] When the rejection risk probability P ≥ 0.5, it is determined to be a high-risk state, triggering the adaptive control mechanism. The temperature / pH linkage control unit then performs the following operations:

[0043] The temperature of the perfusion fluid was reduced from 36.5°C to 36.0±0.1°C using a semiconductor cooling chip;

[0044] The pH was adjusted to 7.4 ± 0.05 using a buffer infusion pump, with a response time ≤ 30 seconds.

[0045] Secondly, the present invention also provides a method for precise monitoring and adaptive regulation of immune rejection in solid organ transplantation, comprising:

[0046] Real-time collection of rejection-related biomarkers, organ function data, and environmental and treatment data from the peripheral blood of transplant recipients;

[0047] A long short-term memory network exclusion risk prediction model based on hierarchical boundary marginal loss function optimization is adopted to process real-time collected multi-dimensional time series feature data and output the exclusion risk probability P.

[0048] Based on the patient's current pathological state, an individualized intervention strategy containing discrete intervention actions and continuous control variables is generated through a hybrid masked nearest neighbor strategy optimization algorithm.

[0049] Based on the rejection risk probability P and the individualized intervention strategy, the regulatory decision-making for precise regulation of immunosuppressants and synergistic regulation of organ microenvironment is executed and feedback is provided in real time.

[0050] Compared with the prior art, the present invention achieves the following beneficial effects:

[0051] 1. To enhance the model's ability to discriminate the risk of immune rejection, especially in scenarios with sparse and imbalanced labels, this invention proposes a Hierarchical Boundary Margin (HBM) loss function. This method introduces a learnable boundary to adapt to the differences in predicted distributions among different patient groups and organ types, replacing the traditional static threshold; it enhances the model's ability to identify low-confidence (i.e., clinically "suspected rejection") samples through a bias mechanism; it incorporates a marginal strategy to significantly reduce the dominant role of high-confidence samples in the training process, improving the model's robustness on boundary samples; and it is particularly suitable for complex data structures with highly dynamic immune indicators and difficult-to-quantify trends, enabling the sensitive capture of early rejection signals.

[0052] 2. To achieve individualized regulatory intervention strategies, this invention innovatively introduces a hybrid action modeling mechanism from reinforcement learning and designs a hybrid masked nearest neighbor strategy optimization algorithm (mHPPO). This algorithm can simultaneously handle discrete intervention behaviors (such as whether to add hormones or adjust the dosing frequency) and continuous control variables (such as immunosuppressant dosage and drip rate). It also introduces a medical knowledge-driven action masking mechanism to shield logically infeasible or contraindicated operations during the action sampling phase, ensuring the clinical feasibility and safety of the intervention recommendations. Furthermore, it updates the reinforcement learning reward based on real rejection feedback signals (such as marker resurgence and abnormal physiological indicators), forming an intelligent intervention strategy that dynamically adapts to the patient's condition.

[0053] 3. This invention addresses the toxic side effects or rejection risks caused by over- or under-regulation of immunosuppressants. Based on predictive modeling results (rejection risk probability P and regulatory action a), this invention introduces a proportional-integral-derivative (PID) control algorithm to adjust the immunosuppressant infusion rate. The predicted risk error is used as a control signal to generate a regulatory output u(t), which drives a high-precision peristaltic pump to adjust the dose at a minimum particle size of 0.05 ml / h. This achieves dynamic and stable control of blood drug concentration within the therapeutic window (e.g., maintaining cyclosporine A at 100-200 ng / ml), improving the responsiveness and individual precision of clinical interventions, and effectively preventing acute rejection or long-term toxic reactions.

[0054] 4. This invention proposes a mechanism for the coordinated regulation of organ perfusion fluid temperature and pH based on rejection risk prediction results. Under high-risk conditions, by lowering the perfusion temperature and precisely adjusting the pH, it actively optimizes the local microenvironment of the graft, synergistically suppresses immune activity, and improves the accuracy and effectiveness of non-pharmacological interventions. This mechanism is implemented through PID control and an automated buffer infusion system, possessing high responsiveness and precise regulation capabilities.

[0055] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0056] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0057] Figure 1 This is a schematic diagram of a module of a precise monitoring and adaptive regulation system for immune rejection in solid organ transplantation according to an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the workflow of a precise monitoring and adaptive regulation system for immune rejection in solid organ transplantation according to an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0061] Figure 1 This is a schematic diagram of a module of a precise monitoring and adaptive regulation system for immune rejection in solid organ transplantation according to an embodiment of the present invention; Figure 2This is a schematic diagram illustrating the workflow of a precise monitoring and adaptive regulation system for immune rejection in solid organ transplantation, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a precise monitoring and adaptive regulation system 100 for immune rejection in solid organ transplantation includes:

[0062] The real-time data acquisition module 110 is used to collect rejection-related biomarkers in the peripheral blood of transplant recipients in real time and monitor organ function parameters through the organ perfusion system.

[0063] 1.1 Multi-dimensional feature data acquisition

[0064] (1) Immune marker data: Rejection-related biomarkers in the peripheral blood of transplant recipients were collected in real time using microfluidic chips and biosensor arrays, including:

[0065] Cytokines (IL-2, IL-6, TNF-α) IFN- (Minimum detection frequency: once every 4 hours).

[0066] Antibody levels (anti-HLA antibodies, donor-specific antibodies DSA, minimum detection frequency 1 time / 12 hours);

[0067] Immune cell activity (T cell proliferation rate, NK cell toxicity, minimum detection frequency once every 24 hours).

[0068] In practical applications, the detection frequency can be dynamically increased according to the risk level (such as the high-risk strategy of module 140).

[0069] (2) Organ function data: Reusing monitoring parameters from the ex vivo organ perfusion system, including:

[0070] Temperature of organ perfusion fluid ( Accuracy), pH value (7.35-7.45 range), flow rate (real-time monitoring by ultrasonic measurement module);

[0071] Concentration of metabolites (lactic acid, creatinine, urea nitrogen, detected once every 4 hours).

[0072] (3) Environmental and treatment data:

[0073] Immunosuppressant blood concentrations (e.g., tacrolimus, cyclosporine, once every 8 hours);

[0074] Patient's vital signs (body temperature, blood pressure, heart rate, collected in real time);

[0075] Post-transplantation time (time series data in days).

[0076] 1.2 Data Preprocessing

[0077] Outlier handling: using The criteria remove sensor noise data and fill in missing values ​​using time series interpolation.

[0078] Feature standardization: Normalize all indicators to Interval.

[0079] Temporal feature construction: Using a 24-hour sliding window, extract statistical measures such as the mean, variance, and peak value of features within each window to form the temporal feature matrix input to the network model.

[0080] The exclusion risk prediction module 120 is used to process real-time collected multi-dimensional time series feature data and output the exclusion risk probability P by adopting a long short-term memory network exclusion risk prediction model based on hierarchical boundary marginal loss function optimization.

[0081] 2.1 Construction of Exclusion Risk Prediction Model

[0082] To address the time-series characteristics of preprocessed multidimensional immune rejection monitoring, a rejection risk prediction model was constructed using a Long Short-Term Memory (LSTM) network. The input to the rejection risk prediction model is a multidimensional... The feature matrix is ​​T=24 (corresponding to a 24-hour window, linear interpolation is used to complete the time series for indicators with insufficient detection frequency), and F≥20 (covering key indicators such as cytokines, antibody levels, immune cell activity, organ perfusion parameters, environmental and treatment information).

[0083] 2.2 Training of the Exclusion Risk Prediction Model

[0084] In order to more effectively identify potential rejection risk labels, the training process adopts the hierarchical boundary margin (HBM) loss function. The hierarchical boundary margin loss function optimizes the training process of the long short-term memory network rejection risk prediction model through the following mechanisms: (1) hierarchical dynamic boundary mechanism: the monitoring units are divided according to organ type or patient group, and each unit learns an adaptive dynamic boundary threshold to replace the fixed classification threshold to adapt to the characteristic differences of different transplantation scenarios; (2) confidence-driven bias mechanism: based on the dispersion of the predicted confidence of samples in each monitoring unit, the bias of positive and negative samples is dynamically calculated to enhance the model's ability to identify suspected rejection samples with predicted confidence in the boundary region; (3) marginal filtering mechanism: during the training process, high confidence samples with predicted confidence reaching saturation are filtered out by preset marginal parameters, and training resources are concentrated on learning low confidence boundary samples.

[0085] Specifically, this function incorporates learnable boundaries. Bias and the margin This strategy demonstrates significant advantages in addressing the label imbalance and static threshold issues encountered in immune rejection discrimination. Its overall loss function is defined as follows:

[0086]

[0087] A collection of all monitoring units (e.g., grouped by organ type or patient) used for hierarchical modeling; A single monitoring unit represents a subset of tags (such as the rejection subtypes corresponding to liver transplantation or kidney transplantation). The total number of cells, used to average the loss of all cells; The logit value output by LSTM represents the prediction confidence of positive (if there is a risk of rejection) or negative (if the possibility of rejection is extremely low) labels in this cell within the current time window. Learnable boundary thresholds, dynamic thresholds learned based on organ function fluctuations and immune dynamic features, can replace fixed judgment values ​​(such as 0.5) and improve the model's adaptability across different organs and individuals; , : These are the positive and negative sample bias terms, respectively. The training focus is dynamically adjusted based on the standard deviation of the logit value to enhance the learning ability for "hard-to-discriminate" samples near the boundary (such as early rejection signals). The calculation method is as follows:

[0088] Positive sample bias term: ;

[0089] Negative sample bias term: ;

[0090] Hyperparameters; The standard deviation of the logit value reflects the dispersion of the label distribution. It can be in the range of [0.1, 0.5], with 0.3 being the preferred value.

[0091] , Marginally filtered subset of labels is used to mitigate the influence of high-confidence labels on the dominant training effect, retaining only the low-confidence samples. The specific definition is as follows:

[0092]

[0093]

[0094] m: Marginal hyperparameter, controlling the filtration intensity (take 0.1 or 0.01). This represents the normalized prediction probability of the current label. This is the sigmoid function.

[0095] Positive sample items passed Increase attention to low-confidence positive samples (suspected exclusion) to avoid high-confidence samples dominating training; negative sample items: through Improve the ability to identify low-confidence negative samples (non-rejection) and reduce false negatives. Each monitoring unit Independent learning boundaries And bias terms to adapt to the dynamic differences in immunity among different organs / patient groups.

[0096] The rejection risk probability P predicted by the LSTM network is:

[0097]

[0098] Where X: is the preprocessed multimodal temporal feature input matrix; : Fully connected weight matrix; : Bias term; : Perform time series modeling using an LSTM network and output the hidden state at the last time step; The Sigmoid function is used to map the output to... The interval represents the probability of immune rejection.

[0099] This model exhibits higher sensitivity and robustness in identifying early immune rejection risk signals during solid organ transplantation (such as rapid fluctuations in cytokine levels and elevated lactate in the perfusion fluid), providing a reliable basis for subsequent intervention and regulation.

[0100] The intervention strategy generation module 130 is used to generate an individualized intervention strategy containing discrete intervention actions and continuous control variables based on the patient's current pathological state through a hybrid mask nearest neighbor strategy optimization algorithm.

[0101] To achieve precise regulation and intervention recommendations for immune rejection after solid organ transplantation, a hybrid masked nearest neighbor strategy optimization algorithm (mHPPO) is introduced in the intervention strategy generation module 130. This algorithm integrates the hybrid action space modeling mechanism in reinforcement learning with the feasibility masking mechanism driven by medical rules. It can simultaneously handle discrete selection behaviors (such as whether to administer immunosuppressants or whether to add auxiliary examinations) and continuous parameter control (such as drug dosage and administration rate) in postoperative intervention strategies, thereby achieving efficient and reasonable individualized regulation.

[0102] 3.1 Discrete Intervention Action Decision

[0103] For discrete-type medical interventions, the Hybrid Masked Nearest Neighbor Optimization (mHPPO) algorithm introduces a medical feasibility masking mechanism. This mechanism constructs a medical knowledge base rule system based on key indicators such as organ type, patient immune status, and physiological parameters. Interventions that are not feasible in the current state are masked, preventing irrational actions from being missampled and thus improving the safety and effectiveness of the strategy. Specifically, the medical knowledge-driven action masking mechanism involves: predefining or calling upon a medical knowledge base containing organ-specific contraindications, drug interaction rules, and patient physiological state limitations; matching the knowledge base rules to the patient's current real-time monitoring data; and assigning a logical mask prohibiting the execution of interventions that violate the rules.

[0104] Furthermore, the Hybrid Masked Nearest Neighbor Optimization (mHPPO) algorithm generates discrete intervention actions by: based on the patient's real-time pathological state, outputting the original probability distribution of discrete actions through a neural network; applying a medical knowledge-driven action masking mechanism to logically mask actions that violate clinical contraindications or safety guidelines; and performing probability normalization on the unmasked actions to output discrete intervention action decisions that meet clinical feasibility. The core formulas involved in this process are as follows:

[0105]

[0106] The log probability of the original intervention action output by the neural network; : By combining medical rule masks with adjusted action probabilities, unreasonable operations are assigned... ; : The final sampling probability of each intervention action under the current pathological state vector s; s: the current pathological state vector, i.e., the real-time pathological state, such as the concentration of immune factors, white blood cell count, liver and kidney function indicators, etc. on day t after transplantation. ,in, Characterizing immune status: CD4+ / CD8+ T cell ratio, IFN-γ concentration, and anti-HLA antibody titer (normalized to [0,1]); Characterizing organ function: lactate value of perfusion fluid, bile secretion rate (liver transplantation), and urine output (kidney transplantation). Characterizing treatment status: current blood concentration of immunosuppressants, and markers of hormone use; Characterizing temporal features: number of days post-transplantation, number of days since the last rejection event; Let be the dimension of the state vector s, for example =15, and the specific feature selection can be determined by ranking the importance of random forest; Activation function.

[0107] 3.2 Decision-making for continuous intervention actions

[0108] For continuous intervention variables (such as the dosage or duration of cyclosporine A), the uncertainty and boundary conditions are modeled using a Beta distribution.

[0109] Furthermore, the process of generating continuous control variables using the Hybrid Masked Nearest Neighbor Optimization (mHPPO) algorithm is as follows: based on the patient's real-time pathological state, the neural network activates and outputs non-negative, biologically meaningful shape parameters through the SoftPlus function. Using shape parameters Construct a Beta probability distribution model; sample from the distribution to generate continuous intervention values. Specifically: the shape parameters are output by a neural network. The SoftPlus activation function is used to ensure that the parameters are non-negative and have biological significance. The sampling formula is as follows:

[0110]

[0111] The generated continuous intervention value, and, as the continuous regulation intensity coefficient of the immunosuppressant infusion unit 141 for adjusting the drug infusion rate (described later), is... ,and It is generated by weighting the rate of change of the dynamic change of the exclusion risk probability P: λ is the sensitivity adjustment coefficient (adjustable hyperparameter (0.5~2.0), preferably 0.8). To exclude the instantaneous rate of change of the probability of risk, It represents the absolute value of the rate of change of the probability of excluding risk. Used to model the probability distribution of drug administration parameter uncertainties; A neural network with parameter θ, using the current pathological state vector s and discrete action encoding vectors. (i.e., discrete actions, such as "whether to stop medication") is the input, and the output is... ; : Controls the shape of the beta distribution, reflecting the preference and range of fluctuation in dose selection, with a value ≥1; Activation function, mathematical expression is .

[0112] In Hybrid Masked Nearest Neighbor Policy Optimization (mHPPO), the continuous action network takes discrete actions as conditional inputs and implicitly learns the constraints between actions through the neural network. Discrete actions (such as "whether to stop medication") establish dependencies with continuous actions (such as drug dosage).

[0113] (1) Discrete action coding:

[0114] or Where k∈{0,1,…,K}: discrete action index; K is the number of discrete action types, for example, K=5. k=0: discontinue medication (emergency intervention, such as in case of severe infection); k=1: maintain the current regimen; k=2: add adjuvant drugs (such as adding mycophenolate mofetil); k=3: hormone pulse therapy (methylprednisolone 500mg×3 days); k=4: reduce the dose (50% of the current dose). This represents discrete intervention actions, implemented using one-hot encoding or embedding.

[0115] (2) Neural network construction:

[0116] Neural network structures include:

[0117] Input layer: , ⊕ represents the vector concatenation operation, This represents the dimension of the state vector s. Discrete action coding The dimension;

[0118] Hidden layer: 2 fully connected layers, 128 nodes per layer, ReLU activated;

[0119] Output layer: ].

[0120] (3) Network output decomposition:

[0121]

[0122] and These are the original shape parameters output by the neural network;

[0123] For example, when Online learning during "discontinuation of medication":

[0124] α→1, β→∞, Beta(1,∞)≈δ(0), forced dose ≈0.

[0125] when = "Impact therapy" during online learning:

[0126] α>>1, β≈1, right-skewed distribution, tendency to generate high doses.

[0127] (4) Parameter sharing efficiency:

[0128] ,

[0129] This represents the gradient of the shape parameter α with respect to the network parameters. This represents the gradient of the shape parameter β with respect to the network parameters. Discrete action gradients can be backpropagated to the shared network.

[0130] 3.3 Overall Algorithm Flow

[0131] Furthermore, the closed-loop optimization process of the Hybrid Masked Nearest Neighbor Policy Optimization (mHPPO) algorithm includes: executing the generated discrete actions and continuous control values, and observing whether a rejection response is induced; collecting biomarker feedback signals and calculating the reward for improving rejection risk; updating the decision network parameters using the nearest neighbor policy optimization, and optimizing the total reward function. More specifically, the overall decision optimization process of the mHPPO algorithm is as follows:

[0132] Postoperative condition monitoring: such as the proportion of immune cell subtypes, the level of inflammatory factors, and clinical scores;

[0133] Intervention action generation:

[0134] Discrete decision-making: such as whether to add hormone pulse therapy or whether to discontinue medication, filtering out non-executable strategies through masking;

[0135] Continuous adjustment: such as the dosage and administration time of a specific drug, sampled using a Beta distribution.

[0136] Environmental interaction and effect feedback: Implement the strategy in simulated or real postoperative data and observe whether it induces rejection reaction;

[0137] Reward Design: Rewards are calculated based on indicators such as whether the rejection reaction is successfully delayed and the level of immune attack markers is reduced;

[0138] Reward function design:

[0139]

[0140] Immune reward: The decrease in inflammatory factors over the past 24 hours;

[0141] Organ function reward: Improved organ function is a positive reward;

[0142] Toxic punishment: Exceeding the standard will result in punishment;

[0143] Policy parameter update: The policy network parameters are updated using the Proximal Policy Optimization (PPO) algorithm. Optimize the total reward function. PPO update process:

[0144] 1. Trajectory Collection: Using the current strategy Execute actions and store the quadruple in real time. };in, Current status (immune indicators, organ function, etc.); : The intervention action performed (discrete action); : Regulation intensity coefficient (continuous value); : Instant reward value (calculated based on the improvement of rejection risk);

[0145] 2. Dominance estimation: Calculating the dominance value using the time-difference method: γ is the future reward discount factor, with the preferred γ=0.99;

[0146] It is a state-value function, initialized as a zero matrix, and updated using a moving average:

[0147] β is the weighting coefficient, and β = 0.1.

[0148] 3. Batch Update and Loss Calculation: After accumulating 2000 sets of data, the loss is calculated for each set based on a probability ratio pruning mechanism.

[0149]

[0150] Where θ represents the updated parameter vector of the policy network, i.e., the new parameters obtained by gradient descent; This represents the parameter vector before the policy network update, i.e., the old policy parameters used to generate the current batch of data;

[0151] 4. Parameter Update: Adam Optimizer Maximization The learning rate is set to 0.0003;

[0152] 5. Policy Synchronization: After updating the policy network parameters based on the current batch of data, the updated parameter vector θ is assigned to the baseline policy parameter vector. ,Right now: ←θ;

[0153] This operation ensures that the strategy is more accurate in the next round of data collection. The computational baseline remains consistent with the current optimization strategy;

[0154] 6. Periodic Reset: Clear stored data after every 10 updates.

[0155] In the intervention strategy generation module 130, the structured modeling and optimization design through the Hybrid Masked Nearest Neighbor Optimization (mHPPO) algorithm not only ensures the medical rationality and safety of the intervention strategy, but also improves the responsiveness and personalization of the recommended immune rejection intervention strategy.

[0156] The intelligent decision-making and regulation module 140 includes an immunosuppressant infusion unit driven by a PID controller and a temperature / pH linkage regulation unit, which is used to make regulatory decisions on precise regulation of immunosuppressants and coordinated regulation of organ microenvironment based on the rejection risk probability P and individualized intervention strategies, and provide real-time feedback.

[0157] Furthermore, the intelligent decision-making and control module 140 includes:

[0158] The immunosuppressant infusion unit 141 is used to finely adjust the infusion rate of the immunosuppressant in real time based on the rejection risk probability P and the individualized intervention strategy using a PID control algorithm.

[0159] To prevent excessive or insufficient immunosuppression after solid organ transplantation, this invention uses a proportional-integral-derivative (PID) control algorithm to finely adjust the infusion rate of immunosuppressants in real time, based on predictive modeling results (the rejection risk probability P obtained by the rejection risk prediction module 120 and the regulation action a obtained by the intervention strategy generation module 130).

[0160] Furthermore, the process of using a PID control algorithm to finely adjust the infusion rate of immunosuppressants in real time includes: setting a target rejection risk probability value as the benchmark for the PID control algorithm; calculating the real-time error between the current predicted rejection risk probability P and the target risk threshold; inputting the real-time error into the PID controller, multiplying the original output of the PID controller by the real-time control intensity coefficient a(t) generated by the intelligent control decision module to obtain the final immunosuppressant infusion rate adjustment amount u(t); and driving a high-precision peristaltic pump to adjust the infusion rate in a stepwise manner with a preset minimum adjustment granularity according to the immunosuppressant infusion rate adjustment amount u(t), so that the blood concentration of immunosuppressants in the patient is stably maintained within a predefined therapeutic window range. Through the above process, the algorithm can dynamically generate an infusion adjustment value based on the error between the current rejection risk and the set target range, as shown in the following formula:

[0161]

[0162] in, Adjustment to the immunosuppressant infusion rate at time t (unit: Positive values ​​indicate an increase in infusion rate, while negative values ​​indicate a decrease. : Proportional gain coefficient, used to adjust for the current rejection error Quickly adjust the infusion rate; Integral gain coefficient, used to correct long-term risk bias by accumulating historical errors; The differential gain coefficient measures the rate of change of error and suppresses system oscillations caused by rapid changes in repulsion risk. The aforementioned PID parameters can be determined using the Ziegler-Nichols tuning method. Typical values ​​are... =0.5, =0.2, =0.1. : Indicates the current prediction exclusion risk error value, A positive value indicates that the current risk of rejection is higher than the target, and more suppression is needed; a negative value indicates oversuppression and excessive risk; P: the probability of current rejection reaction predicted by the model, the higher the value, the more likely rejection is to occur; Target risk threshold, adjustable based on dynamic target mechanism. Continuous intervention values ​​output from intervention strategy generation module 130 Here, it serves as the continuous control intensity coefficient, and if =1, then it is completely controlled by PID; if <1 indicates that reinforcement learning is recommended to be "mildly adjusted" at the current stage; if A value greater than 1 indicates that the policy network determines that greater intervention is needed. Among these,

[0163]

[0164] Controller output A peristaltic pump controlled by a high-precision stepper motor is used to discretely adjust the infusion rate of immunosuppressants with a minimum adjustment accuracy of 0.05 ml / h. This method ensures that the blood drug concentration is dynamically maintained within the therapeutic window (e.g., cyclosporine A in the range of 100-200 ng / ml), reducing the risk of postoperative acute rejection and long-term toxicity.

[0165] In some embodiments, to prevent a sudden drop in concentration or overdose due to complete drug discontinuation, an infusion rate constraint mechanism is implemented, such that... , This is the minimum permissible infusion rate, for example, 0.1 ml / h. This is the maximum permissible infusion rate, for example, 5 ml / h. The immunosuppressant infusion unit 141 adjusts the infusion rate according to the final infusion rate, driving a high-precision peristaltic pump to adjust the infusion rate in a stepwise manner with a preset minimum adjustment granularity, so that the blood concentration of immunosuppressant in the patient is stably maintained within a predefined therapeutic window range.

[0166] The temperature / pH linkage control unit 142 is used to automatically initiate the linkage regulation of organ perfusion fluid temperature and pH value when the detected rejection risk probability P reaches a preset high-risk threshold. This includes: precisely lowering the organ perfusion fluid temperature from the basic set value to a preset low-temperature set value range; synchronously monitoring the perfusion fluid pH value in real time, and automatically initiating buffer infusion to precisely adjust the pH value to the target range when it deviates from the preset physiological neutral range; both temperature and pH regulation achieve rapid response and high-precision maintenance through a closed-loop control system.

[0167] Specifically, in this embodiment, the system automatically calls the exclusion risk prediction network model every 2 hours, inputs the latest time-series feature data, outputs the exclusion risk probability P, and sets three levels of risk thresholds:

[0168] Low risk: P < 0.2, maintain basic monitoring frequency;

[0169] Medium risk: 0.2≤P<0.5, initiate enhanced surveillance (increase the frequency of immune indicator testing to once / hour).

[0170] High risk: P ≥ 0.5, triggering the adaptive control strategy of the intelligent decision-making and control module 140 based on the prediction results, and the temperature / pH linkage control unit 142 performs coordinated adjustment of temperature and pH, specifically as follows:

[0171] Temperature control: Adjusting the temperature of the organ perfusion fluid from the conventional... Fine-tuning to (Reducing immune cell activity), rapid response (temperature difference) is achieved through compressor PID control (segmented control logic). (Single compressor PID control starts at the same time).

[0172] pH adjustment: The pH value of the perfusion solution is monitored in real time by a pH meter. When it deviates from the range of 7.4±0.1, the buffer infusion pump is started (synchronously controlled with the second peristaltic pump), with an adjustment accuracy of ±0.05.

[0173] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0174] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0175] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A precision monitoring and adaptive regulation system for immune rejection of solid organ transplantation, characterized in that, The method comprises the following steps: a real-time data acquisition module for real-time acquisition of rejection-related biomarkers, organ function data, and environmental and treatment data in peripheral blood of transplant recipients; a rejection risk prediction module for processing multi-dimensional time series feature data collected in real time using a long short-term memory network rejection risk prediction model optimized based on a hierarchical boundary marginal loss function, and outputting a rejection risk probability P; an intervention strategy generation module for generating an individualized intervention strategy containing discrete intervention actions and continuous control variables based on the current pathological state of the patient through a hybrid mask nearest neighbor strategy optimization algorithm; an intelligent decision-making regulation module containing a PID controller-driven immunosuppressant infusion unit and a temperature / pH linkage regulation unit for executing regulation decisions for precise immunosuppressant regulation and organ microenvironment synergistic regulation according to the rejection risk probability P and the individualized intervention strategy and providing real-time feedback.

2. The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 1, wherein, The hierarchical boundary marginal loss function used by the rejection risk prediction module optimizes the training of the long short-term memory network rejection risk prediction model through the following mechanisms: a hierarchical dynamic boundary mechanism: dividing monitoring units according to organ type or patient group, learning adaptive dynamic boundary thresholds for each unit, and replacing fixed classification thresholds to adapt to the characteristic differences of different transplantation scenarios; a confidence-driven bias mechanism: dynamically calculating the bias of positive and negative samples based on the dispersion of the prediction confidence of samples in each monitoring unit, and enhancing the model's ability to identify suspected rejection samples with boundary region prediction confidence; a marginal filtering mechanism: in the training process, high-confidence samples with saturated prediction confidence are filtered out through a pre-set marginal parameter, and the training resources are concentrated on learning low-confidence boundary samples. 3.The system of claim 2, wherein, Wherein, the hybrid mask nearest neighbor strategy optimization algorithm for generating discrete intervention actions comprises: outputting the original probability distribution of discrete actions through a neural network based on the real-time pathological state of the patient; applying a medical knowledge-driven action mask mechanism to logically shield actions that violate clinical contraindications or safety criteria; performing probability normalization on unshielded actions to output discrete intervention action decisions that meet clinical feasibility. 4.The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 3, wherein, Wherein, the hybrid mask nearest neighbor strategy optimization algorithm for generating continuous control variables comprises: Based on the real-time pathological state of the patient, the non-negative biologically meaningful shape parameters are output by the neural network through SoftPlus function activation ; Utilizing the shape parameter constructing a beta probability distribution model; sampling from a distribution to generate continuous intervention values as continuous regulation intensity coefficients for adjusting the infusion rate of the intelligent decision-making regulation module; wherein the continuous regulation intensity coefficient reflects the recommendation of the reinforcement learning strategy for the current regulation intensity as mild, standard, or enhanced.

5. The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 4, wherein, Wherein, the closed-loop optimization process of the hybrid mask nearest neighbor strategy optimization algorithm comprises: execute the generated discrete actions and continuous regulation values, and observe whether they induce rejection reactions; collect biomarker feedback signals, calculate rejection risk improvement rewards; update decision network parameters using proximal policy optimization to optimize the total return function. 6.The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 3, wherein, Wherein, the medical knowledge-driven action mask mechanism specifically refers to: predefining or calling a medical knowledge base containing organ-specific contraindications, drug interaction rules, and patient physiological state restrictions in real time; matching the knowledge base rules according to the patient's current real-time monitoring data, and assigning a logical mask to the intervention actions that violate the rules to prohibit execution. 7.The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 1, wherein, Wherein, The intelligent decision-making regulation module comprises: An immunosuppressant infusion unit for real-time fine adjustment of the infusion rate of immunosuppressants based on the rejection risk probability P and the individualized intervention strategy, using a PID control algorithm; A temperature / pH linkage control unit for automatically starting linkage regulation of the temperature and pH of the organ perfusion fluid when the rejection risk probability P is detected to reach a preset high-risk threshold, comprising: Precise reduction of the temperature of the organ perfusion fluid from a basic set value to a preset low-temperature set value range; Synchronous real-time monitoring of the pH of the perfusion fluid, and automatic starting of buffer infusion to accurately adjust the pH to the target range when it deviates from the preset physiological neutral range; Both the temperature and pH regulation are realized by a closed-loop control system to achieve rapid response and high-precision maintenance. 8.The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 7, wherein, The immunosuppressant infusion unit uses a PID control algorithm to perform real-time fine adjustment of the infusion rate of immunosuppressants, specifically comprising: Setting the target rejection risk probability value as the reference of the PID control algorithm; Calculating the real-time error of the current predicted rejection risk probability P and the target risk threshold; Inputting the real-time error into the PID controller, and multiplying the original output of the PID controller by the real-time regulation intensity coefficient a(t) generated by the intelligent regulation decision-making module to obtain the final immunosuppressant infusion rate adjustment amount u(t); According to the immunosuppressant infusion rate adjustment amount u(t), driving a high-precision peristaltic pump to adjust the infusion rate in steps with a preset minimum adjustment granularity, so that the blood drug concentration of immunosuppressants in the patient's body is stably maintained within the predefined therapeutic window range. 9.The system for precision monitoring and adaptive regulation of immune rejection response of solid organ transplantation according to claim 8, wherein, The intelligent decision-making regulation module comprises: When the rejection risk probability P <0.2, it is determined to be a low-risk state, and the basic monitoring is maintained; When 0.2≤the rejection risk probability P <0.5, it is determined to be a medium-risk state, and intensive monitoring is started: the immune index detection frequency is increased to 1 time / hour; When the rejection risk probability P≥0.5, it is determined to be a high-risk state, triggering an adaptive regulation mechanism, and the temperature / pH linkage control unit performs the following operations: Reduce the perfusion fluid temperature from 36.5°C to 36.0±0.1°C by semiconductor refrigeration sheet; Adjust the pH to 7.4±0.05 by buffer infusion pump, with a response time ≤30 seconds.

Images