A method and system for optimizing toxicity management of a car-t cell therapy

Abstract

The application provides a CAR-T cell therapy toxicity management optimization method and system, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring clinical data of a blood tumor patient receiving a CAR-T cell therapy; performing missing value interpolation and standardization processing on baseline characteristic information and process monitoring data to obtain a standardized data set; obtaining a calibrated virtual patient physiological model according to a preset tumor cell kinetics model, an immune cell kinetics model and a toxicity reaction model; taking the calibrated virtual patient physiological model as an interactive environment to obtain a correlation database; constructing a deep Q network intelligent agent with a maximum reward function as a target to obtain an optimized deep Q network intelligent agent; and inputting the process monitoring data into the optimized deep Q network intelligent agent to obtain a target strategy, so that the individualized CAR-T treatment strategy balancing the therapeutic effect and toxicity can be automatically output by inputting the real-time monitoring data of the patient.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for optimizing the management of toxicity in CAR-T cell therapy. Background Technology

[0002] CAR-T cell therapy has shown significant efficacy in treating hematological malignancies, but adverse reactions such as cytokine release syndrome and immune effector cell-related neurotoxicity syndrome limit its safety and widespread application. Therefore, developing technologies capable of predicting and precisely managing these toxicities is crucial for improving the safety of CAR-T therapy.

[0003] Several solutions aimed at optimizing toxicity management have emerged in existing technologies. For example, some studies manage neurotoxicity by monitoring multiple indicators such as tumor burden, cytokine levels, and CAR-T cell expansion, and then implementing interventions based on these indicators. In addition, there are attempts to use deep learning models to predict the probability of severe CRS.

[0004] However, these existing methods still have limitations. Monitoring and intervention methods based on fixed rules are ill-suited to individual patient differences and dynamic changes in their condition, and cannot provide quantifiable, personalized intervention plans. Furthermore, existing predictive models often focus on the independent prediction of single toxic events, failing to simultaneously and dynamically optimize treatment actions within a unified decision-making framework to balance efficacy and multiple toxicity risks. Therefore, existing technologies suffer from technical problems such as clinical decision lag and insufficient individualization and dynamic optimization capabilities. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for optimizing the management of CAR-T cell therapy toxicity, in order to solve the technical problems of lagging clinical decision-making and insufficient individualization and dynamic optimization capabilities in the existing technology.

[0006] In a first aspect, this application provides an optimized method for managing the toxicity of CAR-T cell therapy, comprising:

[0007] Acquire clinical data from hematologic malignancy patients receiving CAR-T cell therapy. The clinical data includes baseline characteristics, process monitoring data, efficacy evaluation data, toxicity event data, and treatment intervention data.

[0008] Missing values ​​were imputed and standardized on baseline feature information and process monitoring data to obtain a standardized dataset.

[0009] Based on the pre-set tumor cell dynamics model, immune cell dynamics model, and toxicity response model, a virtual patient physiological model is constructed using digital twin technology. The model parameters in the virtual patient physiological model are calibrated using the gradient descent algorithm with clinical data in the standardized dataset as a benchmark, resulting in a calibrated virtual patient physiological model.

[0010] Using the calibrated virtual patient physiological model as the interactive environment, the state space of the interactive environment is defined based on the static features in the baseline feature information and the dynamic features in the process monitoring data. The action space of the interactive environment is defined based on the infusion dose of CAR-T cells and the preset toxicity intervention strategy, so as to carry out simulation and deduction within the interactive environment and obtain the associated database.

[0011] A deep Q-network agent with the goal of maximizing the reward function is constructed. Training samples are extracted from the associated database through a priority experience replay mechanism. A dual-network training architecture consisting of a main network and a target network is used. The network parameters of the deep Q-network agent are iteratively optimized based on the training samples and the dual-network training architecture to obtain the optimized deep Q-network agent. The reward function is used to quantify the balance between tumor clearance and toxicity control.

[0012] The process monitoring data is input into the optimized deep Q-network agent to obtain the target policy.

[0013] Optionally, missing value imputation and standardization are performed on the baseline feature information and process monitoring data to obtain a standardized dataset, including:

[0014] Identify the first missing value in the baseline feature information and the second missing value in the process monitoring data;

[0015] Using multiple interpolation, the first imputation value of the first missing value in the baseline feature information is generated to form complete baseline feature information;

[0016] Using time series interpolation algorithms, second imputation values ​​are generated for the second missing values ​​in the process monitoring data to form complete process monitoring data;

[0017] Calculate the mean and standard deviation of each continuous indicator in the complete baseline feature information and complete process monitoring data;

[0018] Based on the mean and standard deviation, each continuous indicator is standardized to obtain a preliminary dataset;

[0019] The initial dataset is de-identified to obtain a standardized dataset.

[0020] Optionally, based on a pre-defined tumor cell kinetics model, immune cell kinetics model, and toxicity response model, a virtual patient physiological model is constructed using digital twin technology, including:

[0021] Based on the tumor burden, immune status data, and inflammatory factor indicators in the baseline feature information, the corresponding initial parameters of the preset tumor cell dynamics model, immune cell dynamics model, and toxicity response model are set respectively.

[0022] By using digital twin technology, tumor cell dynamics models, immune cell dynamics models, and toxicity response models with pre-set initial parameters are coupled and spliced ​​together to obtain a virtual patient physiological model. The virtual patient physiological model is used to simulate changes in tumor burden, CAR-T cell expansion, and toxic events in patients after receiving CAR-T cell therapy.

[0023] Optionally, the model parameters in the virtual patient physiological model are calibrated using the gradient descent algorithm based on clinical data in the standardized dataset, resulting in a calibrated virtual patient physiological model, including:

[0024] Clinical data from a standardized dataset is input into a virtual patient physiological model to obtain simulated clinical data;

[0025] Calculate the prediction error between simulated clinical data and clinical data;

[0026] Based on the prediction error, the model parameters in the virtual patient physiological model are calibrated using the gradient descent algorithm until the prediction error is less than the preset error threshold, thus obtaining the calibrated virtual patient physiological model.

[0027] Optionally, a deep Q-network agent is constructed with the objective of maximizing the reward function, including:

[0028] Based on the state space dimension of the state space, a network structure is designed, which includes an input layer, a hidden layer, and an output layer that matches the action space dimension of the action space.

[0029] Based on efficacy evaluation data and toxicity event data, a reward function is constructed, which includes a positive reward for tumor clearance and a negative penalty for toxicity.

[0030] By associating the network structure with the reward function, a deep Q-network agent is obtained.

[0031] Secondly, this application provides a CAR-T cell therapy toxicity management optimization system, comprising:

[0032] The acquisition module is used to acquire clinical data of hematologic malignancies patients receiving CAR-T cell therapy. The clinical data includes baseline characteristics, process monitoring data, efficacy evaluation data, toxicity event data, and treatment intervention data.

[0033] The module is used to impute missing values ​​and standardize baseline feature information and process monitoring data to obtain a standardized dataset.

[0034] The construction module is also used to construct a virtual patient physiological model based on a preset tumor cell dynamics model, immune cell dynamics model, and toxicity response model using digital twin technology. The gradient descent algorithm is used to calibrate the model parameters in the virtual patient physiological model based on clinical data in a standardized dataset to obtain a calibrated virtual patient physiological model.

[0035] The definition module is used to define the state space of the interactive environment based on the static features in the baseline feature information and the dynamic features in the process monitoring data, and to define the action space of the interactive environment based on the infusion dose of CAR-T cells and the preset toxicity intervention strategy, so as to perform simulation and deduction within the interactive environment and obtain the associated database.

[0036] The building module is also used to construct a deep Q-network agent with the goal of maximizing the reward function. Training samples are extracted from the associated database through a priority experience replay mechanism. Based on the dual-network training architecture consisting of the main network and the target network, the network parameters of the deep Q-network agent are iteratively optimized according to the training samples and the dual-network training architecture to obtain the optimized deep Q-network agent. The reward function is used to quantify the balance between tumor clearance and toxicity control.

[0037] The input module takes process monitoring data and inputs it into the optimized deep Q-network agent to obtain the target policy.

[0038] Thirdly, this application provides an electronic device, comprising:

[0039] Memory, used to store computer programs;

[0040] A processor is used to perform computer programs to implement the steps of an optimized method for toxicity management of CAR-T cell therapy as described in the first aspect above.

[0041] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the CAR-T cell therapy toxicity management optimization method as described in the first aspect above.

[0042] This application provides a method for optimizing CAR-T cell therapy toxicity management. By acquiring and processing multi-dimensional clinical data, it solves the problem of inconsistent model inputs caused by scattered and inconsistent data formats in existing technologies. It provides a core technical means to address the lack of personalized and dynamic prediction tools in existing technologies. It overcomes the limitation of obtaining massive decision-making samples through experiments in the real world, creating conditions for training data-driven high-order decision-making models. It realizes a shift from passive response to proactive optimization, and from experience-driven to data- and model-driven collaborative decision-making, improving the systematicness and foresight of clinical decision-making.

[0043] Furthermore, based on the tumor burden, immune status data, and inflammatory factor indicators in the patient's baseline characteristics, initial parameters were set for the tumor cell dynamics model, immune cell dynamics model, and toxicity response model, respectively. Digital twin technology was then used to couple and splice the three models to form an individualized model that can simulate changes in tumor burden, cell expansion, and toxic events after CAR-T therapy. This solves the key technical obstacle that existing single or static models cannot simultaneously and continuously simulate the complex interlocking process of tumor clearance, immune expansion, and toxicity during CAR-T therapy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating an optimized method for CAR-T cell therapy toxicity management provided in this application embodiment;

[0046] Figure 2 A training flowchart for a deep Q-network agent is provided in an embodiment of this application;

[0047] Figure 3 A schematic diagram of a CAR-T cell therapy toxicity management optimization system provided in this application embodiment;

[0048] Figure 4 A schematic diagram of another CAR-T cell therapy toxicity management optimization system provided in this application embodiment;

[0049] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0050] In view of the problems of clinical decision lag, insufficient individualization and dynamic optimization capabilities in existing CAR-T toxicity management protocols, especially the difficulty of adapting fixed rule methods to individual differences and the limitation of existing predictive models lacking a unified decision framework to simultaneously optimize treatment actions, a closed-loop system from data to decision was designed.

[0051] Specifically, a standardized, high-quality dataset is constructed by integrating multi-dimensional clinical data; innovatively, digital twin technology is used to integrate tumor, immune, and toxicity response mechanism models to build and calibrate a virtual patient physiological model that can simulate the dynamic physiological responses of individual patients, overcoming the shortcomings of static rules; based on this model, a quantified state and action space is defined, and a relational database covering complex decision-making scenarios is generated through large-scale simulation and deduction; in this environment, a deep Q-network agent is trained with the reward objective of balancing efficacy and toxicity, thereby synchronously and dynamically optimizing treatment strategies within a unified framework; finally, the system inputs real-time patient data into the trained agent, thereby achieving individualized and proactive optimization management of the treatment process.

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0053] The core of this application is to provide an optimized method for CAR-T cell therapy toxicity management, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0054] Step 101: Obtain clinical data from hematologic malignancy patients receiving CAR-T cell therapy. Clinical data includes baseline characteristics, process monitoring data, efficacy evaluation data, toxicity event data, and treatment intervention data.

[0055] In this step, clinical data refers to the sum of all medically relevant information generated for hematologic malignancy patients during CAR-T cell therapy. Baseline characteristics refer to characteristic data collected before treatment begins to describe the patient's initial state, which may include demographic information, tumor characteristics, immune status indicators, and underlying disease conditions. Process monitoring data refers to dynamic indicators reflecting the patient's real-time physiological state and treatment response, continuously collected at preset time frequencies after treatment begins. Efficacy evaluation data refers to indicators used to quantitatively assess treatment effectiveness, typically including changes in tumor burden obtained through imaging examinations and efficacy levels determined according to criteria. Toxicity event data refers to adverse reaction information directly related to CAR-T therapy recorded during treatment, primarily including the occurrence time, severity grade, and response to interventions for cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS). Treatment intervention data refers to records of actual medical measures taken during treatment, including the infusion dose of CAR-T cell products and the types, dosages, and timing of administration of various drugs used to manage toxicity.

[0056] In this embodiment, baseline characteristic information is first collected when a patient is admitted to the hospital for CAR-T therapy. This includes recording the patient's age, gender, and weight; determining the tumor type and gene mutation status through pathological examination; assessing the tumor burden using imaging methods; detecting immune status indicators such as the proportion of CD4-positive T cells using flow cytometry; and collecting routine biochemical indicators such as liver and kidney function to assess underlying disease conditions. All of these constitute a personalized patient baseline profile.

[0057] Next, starting from the date of CAR-T cell infusion, high-frequency process monitoring data collection began. For example, blood samples were collected every six hours to detect the levels of key inflammatory factors such as interleukin IL-6; the number of CAR-T cells in peripheral blood was quantified daily using flow cytometry or quantitative pCR technology; and the patient's vital signs, such as body temperature and blood pressure, were continuously monitored. These data formed a dynamic time-series dataset.

[0058] Then, efficacy evaluation data are collected at specific time points after treatment. For example, at the first, third, and sixth month after infusion, changes in tumor metabolic activity are assessed by pET-CT imaging, and clinicians determine whether the response is complete remission, partial remission, stable disease, or disease progression according to the efficacy evaluation criteria for solid tumors.

[0059] Simultaneously, toxicity event data were meticulously recorded throughout the treatment and monitoring period. If a patient developed symptoms such as fever, CRS classification was performed according to standards; if neuropsychiatric symptoms appeared, ICANS classification was performed.

[0060] Detailed records were kept of the timing and duration of each toxicity event, and the decline in relevant cytokine levels was monitored after intervention with drugs such as tocilizumab to assess the response. Finally, complete records of treatment intervention data were maintained. This included the total dose of CAR-T cells actually infused, as well as the name, dosage, and specific time of administration of each intervention drug used during the treatment of toxicity.

[0061] For example, suppose that at Hospital A, hematologic malignancy patient p1 is preparing to receive CD19-targeted CAR-T cell therapy. Before treatment, her baseline characteristics are collected: age 55 years, diagnosed with stage IV diffuse large B-cell lymphoma, baseline pET-CT assessment of tumor burden as X, flow cytometry analysis of her CD4-positive T cell percentage as Y%, and creatinine clearance rate as Z mL / min.

[0062] Intensive monitoring began after CAR-T cell infusion. For example, on day 3 post-infusion, monitoring data showed an IL-6 concentration of Mpg / mL, a CAR-T cell count of N cells / µL, and a patient temperature of 38.5 degrees Celsius. The first efficacy assessment was performed on day 30 post-infusion; pET-CT showed a significant decrease in metabolic activity at the target lesion compared to baseline, and the efficacy evaluation data indicated partial remission.

[0063] On day 5 post-infusion, the patient developed a high fever, and the toxic event data was recorded as grade 2 CRS. Tocilizumab 800mg was subsequently administered as an intervention, and this treatment intervention data was recorded. Six hours after administration, the IL-6 concentration decreased by more than 50%, and this response was recorded in the toxic event data. All the clinical data obtained from patient p1, after standardization in subsequent steps, will be used to construct a personalized digital twin model representing this patient and will serve as the data foundation for training the intelligent decision-making model.

[0064] Step 102: Impute missing values ​​and standardize the baseline feature information and process monitoring data to obtain a standardized dataset.

[0065] In this step, the standardized dataset refers to a dataset that has been formatted by missing value imputation and scaling, and is uniform and can be directly used for model training.

[0066] Step 201: Identify the first missing value in the baseline feature information and the second missing value in the process monitoring data.

[0067] In this step, the first missing value refers to an unrecorded or invalid data entry in the baseline feature information. The second missing value refers to data in the process monitoring data sequence that was not successfully collected or recorded at the expected time point.

[0068] In this embodiment, firstly, the acquired baseline feature information is scanned to check whether each predefined field contains blanks or is marked as invalid. For example, in the baseline information of patient p2, the gene mutation status field is empty, which is identified as a first missing value. Next, the temporal integrity of the process monitoring data arranged in chronological order is checked. For example, in the process monitoring data of patient p3, IL-6 testing should be performed every six hours according to a preset frequency, but the record at the 36th hour after infusion is missing; the data at this position is identified as a second missing value.

[0069] Step 202: Using multiple interpolation, generate the first imputation value for the first missing value in the baseline feature information to form complete baseline feature information, and use time series interpolation algorithm to generate the second imputation value for the second missing value in the process monitoring data to form complete process monitoring data.

[0070] In this step, multiple imputation is a method that generates multiple reasonable substitute values ​​based on a statistical model to fill in missing data. Time series interpolation algorithms are methods that estimate missing values ​​in time series data based on the trends of neighboring known data points.

[0071] In this embodiment, multiple imputation is used to process the identified first missing value. For example, for the p2 missing gene mutation status of a patient, a model is built based on the patient's known tumor type, age, and other characteristics, combined with the complete data patterns of other patients, and finally a most likely imputation status is generated: mutant type A. After all the first missing values ​​are imputed, they constitute complete baseline feature information. The time series interpolation algorithm is used to process the identified second missing value. Specifically, the time series linear interpolation method is used, and its calculation formula is shown in formula (1):

[0072]

[0073] Where F represents the imputed value, tm is the time point where the missing value is located, ta and tb are the nearest valid observation time points before and after it, and Ca and Cb are the corresponding observed values. For example, for the IL-6 value at hour 36 of a patient with p3 deficiency, given that the value at hour 30 (Ca) is 80 pg / mL and the value at hour 42 (Cb) is 150 pg / mL, substituting these values ​​into the formula yields an imputed value F of approximately 113 pg / mL. After imputed, all second missing values ​​constitute complete process monitoring data.

[0074] Step 203: Calculate the mean and standard deviation of each continuous indicator in the complete baseline feature information and complete process monitoring data.

[0075] In this step, continuous indicators refer to indicators whose values ​​can change continuously and have practical measurement significance, such as tumor burden, IL-6 concentration, and age.

[0076] In this embodiment, firstly, all continuous indicators are selected from the complete baseline characteristics and complete process monitoring data of all patients. Then, for each selected continuous indicator, its mean M and standard deviation S are independently calculated across all patient samples. For example, assume that the mean Mt of the baseline tumor burden values ​​for all patients is 4500 and the standard deviation St is 1200; and the mean Mi of the baseline IL-6 concentration for all patients is 15 pg / mL and the standard deviation Si is 8 pg / mL. This process generates a pair of parameters (M, S) for each indicator.

[0077] Step 204: Based on the mean and standard deviation, standardize each continuous indicator to obtain a preliminary dataset, and then de-identify the preliminary dataset to obtain a standardized dataset.

[0078] In this embodiment, standardization is first performed. For the raw value x of each continuous index in the complete data, the Z-score standardization formula is applied using the corresponding mean M and standard deviation S: Calculate the standardized value z.

[0079] For example, assuming patient p1's baseline tumor burden X is 5000, using its corresponding parameters Mt=4500 and St=1200, the standardized value z≈0.42 is calculated. After this transformation, all indicators yield the initial dataset. Then, the initial dataset undergoes de-identification processing. This includes directly removing direct identifiers such as patient names and ID numbers. Simultaneously, quasi-identifiers such as age and consultation date are generalized or perturbed; for example, exact age is converted to an age range. Finally, a fully de-identified standardized dataset is generated, which can be used for subsequent model building and training.

[0080] Step 103: Based on the preset tumor cell dynamics model, immune cell dynamics model, and toxicity response model, construct a virtual patient physiological model using digital twin technology. Use the gradient descent algorithm to calibrate the model parameters in the virtual patient physiological model using clinical data in the standardized dataset as a benchmark, and obtain the calibrated virtual patient physiological model.

[0081] In this step, digital twin technology refers to a technique that integrates multi-source data and physical models to construct a digital model in virtual space that maps, interacts with, and simulates a physical object in real time. Model parameters refer to adjustable internal variables in the model used to control the model's behavior and output results, such as tumor cell killing rate constants and cytokine release rate constants. A virtual patient physiological model refers to a computational system that integrates tumor cell dynamics models, immune cell dynamics models, and toxicity response models using digital twin technology, capable of simulating the physiological dynamics of a specific individual.

[0082] Step 301: Based on the tumor burden, immune status data, and inflammatory factor indicators in the baseline feature information, set the corresponding initial parameters for the preset tumor cell dynamics model, immune cell dynamics model, and toxicity response model.

[0083] In this step, the initial parameters of the model refer to a set of basic values ​​that need to be assigned to initiate the operation of the preset mechanistic model. These values ​​link the general framework of the model with the initial physiological state of a specific patient.

[0084] In this embodiment, firstly, baseline feature information of a specific patient is extracted from a standardized dataset. Then, appropriate initial parameters are assigned to three preset models based on this information. Specifically, the specific value of the tumor burden assessed by the patient's baseline imaging is set as the initial number of tumor cells in the tumor cell dynamics model. For example, if patient p1's baseline pET-CT assessment of tumor burden is X, then let... The patient's baseline immune status data, such as the proportion of CD4-positive T cells (R), is input into the immune cell kinetics model to initialize the immune microenvironment. Simultaneously, the patient's baseline inflammatory cytokine markers, such as the concentration of IL-6, are input. This serves as the starting biomarker level for the toxicity response model. In this way, each general mechanistic model is given an individualized initial state representing the starting point of treatment for that patient.

[0085] Step 302: Using digital twin technology, the tumor cell dynamics model, immune cell dynamics model, and toxicity response model with initial parameters set are coupled and spliced ​​to obtain a virtual patient physiological model. The virtual patient physiological model is used to simulate changes in tumor burden, CAR-T cell expansion, and toxic events in patients after receiving CAR-T cell therapy.

[0086] In this embodiment, after setting the initial parameters, digital twin technology is used as an integration framework. Based on the physiological and pathological mechanisms of CAR-T therapy, mathematical relationships are established between three models. For example, the CAR-T cell count output by the immune model is simultaneously used as the input for the killing term of the tumor model and the stimulation term of the toxicity model, integrating the tumor cell dynamics model, the immune cell dynamics model, and the toxicity response model into a unified virtual patient physiological model. This process establishes input-output relationships between models based on the physiological and pathological mechanisms of CAR-T therapy. For example, based on the patient p1 model, the CAR-T cell expansion data simulated by the immune cell dynamics model is used. This will serve as an input to the killing term in the tumor cell kinetics model, and will also affect the release level of cytokines in the toxicity response model.

[0087] Through this dynamic data exchange and feedback mechanism, the three sub-models are integrated into a unified virtual patient physiological model capable of simulating the interactive chain reaction from tumor to immunity to toxicity. This model can receive treatment action commands and output simulated tumor burden change curves, CAR-T cell expansion dynamics, and the probability and severity of toxic events.

[0088] Step 311: Input the clinical data from the standardized dataset into the virtual patient physiological model to obtain simulated clinical data.

[0089] In this embodiment, to calibrate the model, time-series clinical data of the patient after the start of treatment in a standardized dataset are used as input to drive the model. For example, the actual CAR-T cell count sequences of patient p1 at multiple time points after CAR-T cell infusion are used. and IL-6 concentration sequence The data are input sequentially into the virtual patient physiological model constructed in step 302. Based on its internal mechanisms, the model runs and calculates the corresponding simulation results using the initial state and these inputs. These results are simulated CAR-T cell counting sequences. and simulated IL-6 concentration sequence These are collectively referred to as simulated clinical data. By comparing simulated data with real clinical data, the predictive accuracy of the model can be evaluated.

[0090] Step 312: Calculate the prediction error between simulated clinical data and clinical data.

[0091] In this step, prediction error is a scalar that quantifies the difference between the model's simulated output and the actual observed values, and is used to measure the model's accuracy.

[0092] In this embodiment, after obtaining simulated clinical data, it is necessary to quantitatively calculate the difference between it and the real clinical data used as a benchmark. A common method is to calculate the mean squared error. Specifically, for the CAR-T cell count of patient p1, the real value sequence is R, the simulated value sequence is S, and the sequence length is n. The prediction error E can be calculated using formula (2):

[0093]

[0094] Where j iterates from 1 to n, and These represent the actual value and the simulated value at the j-th time point, respectively. This error value E comprehensively reflects the overall deviation level of the model's prediction of this indicator throughout the entire simulation period. Similarly, the prediction errors for other indicators such as IL-6 concentration can be calculated.

[0095] Step 313: Based on the prediction error, use the gradient descent algorithm to calibrate the model parameters in the virtual patient physiological model until the prediction error is less than the preset error threshold, and obtain the calibrated virtual patient physiological model.

[0096] In this embodiment, after calculating the prediction error E, the model parameters are automatically calibrated using the gradient descent algorithm. The gradient descent algorithm determines the fastest direction to reduce the error by calculating the gradient g of the prediction error E relative to each adjustable parameter θ of the model. Then, along the negative direction of this gradient descent, at a step size α called the learning rate, according to the update rule: Fine-tune the parameters.

[0097] Taking patient p1 as an example, after adjusting all parameters in one round, the prediction error is repeatedly calculated using the updated parameters. This process is repeated cyclically, with each iteration aiming to reduce the prediction error E. The prediction error E is maintained below a pre-set error threshold. When the iteration stops, the internal parameters of the virtual patient physiological model have been optimized to better reflect the individualized response characteristics of the patient, thus obtaining the final calibrated virtual patient physiological model.

[0098] Step 104: Using the calibrated virtual patient physiological model as the interactive environment, define the state space of the interactive environment based on the static features in the baseline feature information and the dynamic features in the process monitoring data, and define the action space of the interactive environment based on the infusion dose of CAR-T cells and the preset toxicity intervention strategy, so as to perform simulation and deduction within the interactive environment and obtain the associated database.

[0099] In this step, the interaction environment refers to the object with which the agent interacts within the reinforcement learning framework. Static features are attributes extracted from the patient's baseline characteristics that remain largely unchanged during treatment. Dynamic features are indicators extracted from process monitoring data that change over time and reflect the current treatment response. The state space is a mathematical description of all possible states of the interaction environment, where each element, or state, is a feature vector composed of static and dynamic features. The pre-defined toxicity intervention strategy refers to a specific drug intervention plan predefined in the action space for managing toxicity, which may include options such as different doses of tocilizumab monotherapy, tocilizumab combined with different doses of glucocorticoids, etc. The action space refers to the set of all executable operations of the agent in a given environment.

[0100] In this embodiment, the interaction environment and its core elements for reinforcement learning are first defined. A calibrated virtual patient physiological model is established as the interaction environment for the agent to learn and make decisions. Next, the state space of this environment is defined, consisting of both static and dynamic features: static features are extracted directly from the patient's baseline features; dynamic features are extracted from process monitoring data within the most recent time window, such as using a sliding window of data from the past three days to capture trend changes. Simultaneously, the action space is defined, consisting of a series of pre-set optional operations, mainly including different levels of CAR-T cell infusion dosage regimens and various pre-set toxicity intervention strategies.

[0101] Then, large-scale simulations are performed within a defined environment to generate data. Starting from the patient's initial state, an action is randomly or strategically selected in the action space, such as choosing a standard dose of CAR-T cell infusion combined with a pre-defined tocilizumab intervention strategy. This action is input into a virtual patient physiological model, which calculates the new state after a period of time, such as a decrease in tumor burden but the appearance of a certain level of toxicity, and provides an immediate reward according to pre-defined rules. The state, action, reward, and next state of this interaction are fully recorded as a data trajectory.

[0102] This process is repeated tens of thousands or even millions of times to explore various combinations of states and actions, eventually accumulating into a massive database of associations covering all kinds of clinical possibilities.

[0103] Step 105: Construct a deep Q-network agent with the goal of maximizing the reward function. Extract training samples from the associated database through a priority experience replay mechanism. Based on a dual-network training architecture consisting of a main network and a target network, iteratively optimize the network parameters of the deep Q-network agent according to the training samples and the dual-network training architecture to obtain the optimized deep Q-network agent. The reward function is used to quantify the balance between tumor clearance and toxicity control.

[0104] In this step, the Deep Q-Network agent is an artificial intelligence program based on deep neural networks that learns optimal decision-making strategies by maximizing long-term cumulative rewards. The reward function is a function that maps the consequences of the agent's actions to a scalar reward value, used to quantify the balance between tumor clearance and toxicity control. The priority experience replay mechanism is a method of sampling data from an association database for training, assigning different sampling probabilities to samples based on their learning value. The dual-network training architecture refers to simultaneously maintaining two neural networks with identical structures but asynchronous parameter updates: the main network and the target network, to stabilize the training process. The main network is the network in the Deep Q-Network agent that updates parameters in real time and is used for action selection and value prediction. The target network is the network in the Deep Q-Network agent whose parameters are updated with a lag and is used to calculate the stable target value.

[0105] Step 501: Based on the state space dimension of the state space, design the network structure, which includes an input layer, a hidden layer, and an output layer that matches the action space dimension of the action space.

[0106] In this step, the state space dimension refers to the total number of features included in the state vector, which determines the size of the network's input layer. The action space dimension refers to the total number of all selectable actions in the action space, which determines the size of the network's output layer. Here, network structure specifically refers to the organization of the neuron layers within a deep Q-network.

[0107] In this embodiment of the application, firstly, the state space dimension is determined. and action space dimension . This is equal to the number of features defined in the state space. For example, if the state space consists of 20 features such as age, tumor type, current tumor burden, and IL6 concentration, then... . This is equal to the number of independent action options defined in the action space. For example, if the action space includes four CAR-T infusion doses and three preset drug intervention strategies, then... It is 12.

[0108] Next, the network structure is designed based on dimensionality. The number of neurons in the input layer is set to... This is used to receive the complete state vector. The number of neurons in the output layer is set to... Each neuron's output value corresponds to a Q-value for an optional action. Between the input and output layers, at least one hidden layer is configured to learn complex features. For example, a hidden layer H1 with 128 neurons using the ReLU activation function can be designed, along with a hidden layer H2 with 64 neurons using the ReLU activation function. This ultimately forms the basic computational graph of the deep Q-network.

[0109] Step 502: Based on efficacy evaluation data and toxicity event data, construct a reward function, which includes a positive reward for tumor clearance and a negative penalty for toxicity.

[0110] In this step, positive rewards are incentive signals for the desired treatment outcome, while negative punishments are inhibitory signals for adverse treatment outcomes.

[0111] In this embodiment, the reward function R is designed directly based on the treatment goals defined by efficacy evaluation data and toxicity event data. The function generates a scalar reward value r for each step of the simulated or actual treatment evaluation.

[0112] The reward value r is calculated by weighting the overall therapeutic effect reward components. and toxic punishment portion The basic form of the calculation is shown in formula (3):

[0113]

[0114] in, and These are preset weighting coefficients used to balance the importance of efficacy and safety in decision-making. The criteria are set based on efficacy evaluation data; for example, when the tumor burden decreases by more than 10%, When the therapeutic effect reaches complete remission hour, When the therapeutic effect is partial remission, ; Based on toxic event data settings, for example, when a Level 1 toxicity event occurs, For each level increase in toxicity, Increase penalty value ,For example By adjusting and This can guide the agent to learn different risk preference strategies.

[0115] Step 503: Correlate the network structure and reward function to obtain the deep Q-network agent.

[0116] In this embodiment, the network structure and reward function are logically linked to form a complete deep Q-network agent. The core function of this agent is to calculate, through the network, the expected long-term reward for each selectable action a for any input patient state s. Simultaneously, the agent relies on a reward function to evaluate the actual effect of its decisions, that is, to calculate the immediate reward *r* based on the therapeutic and toxic effects produced after the action is performed. For example, if the agent selects an action for a certain state in a virtual environment, and this action, after being simulated by a virtual patient physiological model, produces partial relief and grade 1 toxicity, the immediate reward is calculated according to the reward function rules. .

[0117] The learning objective of an intelligent agent is to continuously adjust network parameters through subsequent training so that the Q-value predicted by the network can accurately reflect the expected long-term cumulative reward obtained from performing the action, thereby learning to make optimal decisions.

[0118] Step 511: Assign a corresponding priority weight to each sample based on the instantaneous reward size of each sample in the associated database.

[0119] In this step, the priority weight is a scalar value assigned to each sample in the associated database to indicate the relative importance of that sample during training.

[0120] In this embodiment, a priority weight p is calculated for each sample data in the associated database, using a method based on the absolute value of temporal difference error. The priority weight p is positively correlated with the learning value of the sample. A commonly used calculation method is shown in formula (4).

[0121] Here, δ represents the temporal difference error of the sample. For newly added samples to the experience pool, the temporal difference error δ is unknown, so the absolute value of its immediate reward |r| can be used as an estimate of its initial priority p. In subsequent training, whenever this sample is used to update the network, its priority is updated using the calculated actual δ value; e is a small normal number used to ensure that all samples have a chance to be sampled. For example, a record of achieving complete remission, i.e., the reward... The samples that record a normal state transition, i.e., the reward, will have a very high initial priority; while a record that records a normal state transition, i.e., the reward... The samples that are prioritized are given lower priority. This ensures that key successes or failures are reviewed more frequently during training.

[0122] Step 512: Based on the priority weights, extract training samples from the associated database through the priority experience replay mechanism.

[0123] In this embodiment, the selection of training samples is guided by a priority weight assigned to each sample. Samples with higher priority have a greater chance of being selected for training. Specifically, a weighted random sampling method, such as a roulette-like algorithm, is used to extract a small batch of samples from the associated database based on the priority weight distribution of all samples. This allows samples that record key treatment experiences, such as those showing significant therapeutic effects or serious toxicity, to be used more frequently for the agent's learning and updating than ordinary state transition samples.

[0124] For example, a mini-batch of 32 samples is drawn from a database containing millions of samples, prioritizing them according to probability. This process ensures that high-priority samples, such as those resulting in breakthrough therapeutic effects or serious toxicities, are more likely to be selected for network updates, thereby improving learning efficiency.

[0125] Step 513: Input the state data from the training samples into the main network to obtain the first Q value, and input the next state data from the training samples into the target network to obtain the second Q value. The state data includes static features and dynamic features.

[0126] In this embodiment of the application, a sample is taken from the training sample set, the content of which includes the current state s, the action a performed, the immediate reward r obtained, and the next state. First, the state s is input into the main network. The main network performs forward propagation calculations and obtains a set of Q-values ​​at its output layer, where the Q-value corresponding to action a is denoted as . This is called the first Q-value. For example, for a specific patient state s, the main network calculates the Q-value for the action of selecting infusion dose D2. The value is 2.3. Simultaneously, the next state in the sample is... The input is fed into the target network. The target network performs a forward propagation and selects the largest Q-value among all actions at its output layer, denoted as . This is called the second Q-value. For example, for the next state... The target network calculated the maximum Q value among all possible actions to be 8.1.

[0127] Step 514: Calculate the target Q value based on the immediate reward and second Q value in the training samples.

[0128] In this embodiment of the application, the instant reward r and the second Q value are obtained. Then, the target Q-value Y is calculated using the temporal difference objective formula of reinforcement learning. Specifically, first, a discount factor γ is obtained, which is a preset constant between 0 and 1 used to measure the importance of future rewards. The product of this discount factor γ and the second Q-value Qn is then added to the immediate reward r to obtain the target Q-value Y.

[0129] For example, if the immediate reward for this sample is r=+1, the second Q-value Qn=8.1, and γ=0.9, then the calculation process is as follows: γ×Qn=7.29, and this product is then added to the immediate reward to obtain the target Q-value Y=8.29. The target Q-value Y represents an improved estimate of the total reward that should be obtained by performing action a in state s, based on the optimal estimate of the current target network.

[0130] Step 515: Based on the dual-network training architecture, according to the difference between the first Q value and the target Q value, the network parameters of the main network are updated a preset number of times using the backpropagation algorithm. The updated network parameters are copied to the target network and returned. According to the priority weights, training samples are extracted from the associated database through the priority experience replay mechanism until the preset number of training rounds are met or the preset performance requirements are achieved, thus obtaining the optimized deep Q network agent.

[0131] In this embodiment of the application, firstly, the predicted value is calculated. The loss L between the target value Y and the mean squared error is usually expressed as the mean squared error. Calculation, based on , The calculated value of L is 35.9. Then, using the backpropagation algorithm, the gradient of the loss L with respect to all parameters θ of the main network is calculated. Next, using an optimizer such as Adam, the parameters of the main network are updated in the reverse direction of the gradient according to the learning rate α. This process can be described as follows: The mini-batch samples extracted in this iteration are used to calculate the average gradient of the loss function with respect to the network parameters, and the parameters of the main network are updated once based on the average gradient. Then, a dual-network training architecture is used to update the target network. The parameters of the target network are... It is not updated in real time. Typically, after N major network parameter updates are completed, the major network parameters are updated... Copy to the target network, that is, make This delayed update strategy helps stabilize the training process.

[0132] Finally, the sampling, computation, and update loop is repeated. Training termination conditions can be reaching a preset total number of iterations, or the agent's performance metrics in the validation environment, such as average round reward, reaching stability. When the termination condition is met, the current main network is the optimized deep Q-network agent.

[0133] like Figure 2 As shown, Figure 2 This application provides a training flowchart for a deep Q-network agent, illustrating the interaction and collaborative optimization mechanism between the core components, the main network and the target network, during the training process.

[0134] The training process begins by extracting samples from an associated database. The state and next state from these samples are input into the main network and the target network, respectively, to obtain the first Q-value and the second Q-value. The target network parameters are periodically copied from the main network to maintain the stability of the value assessment. The system then combines the immediate reward with the second Q-value to calculate the target Q-value, and calculates the difference between it and the first Q-value as the loss. The main network parameters are then updated through backpropagation.

[0135] This diagram visually illustrates how the dual-network training architecture solves the training stability problem through the collaborative work of differentiated roles, which is the core technology for agents to learn strategies to balance efficacy and toxicity.

[0136] Step 106: Input the process monitoring data into the optimized deep Q network agent to obtain the target policy.

[0137] In this step, the target strategy refers to the final recommended plan, which includes specific treatment actions and expected results, generated by the optimized deep Q-network agent based on the current patient to be decided and verified by the virtual patient physiological model.

[0138] Step 601: Based on the process monitoring data and baseline feature information of hematologic malignancy patients, construct a state vector corresponding to the state space.

[0139] In this step, the state vector is an ordered numerical array that encodes all relevant feature information of the patient at a specific point in time according to the definition of the state space. It is a mathematical representation that quantitatively describes the patient's current condition.

[0140] In this embodiment, when clinical decision-making is required for a specific hematologic malignancy patient, the latest data of the patient is first integrated. The latest process monitoring data of the patient at the time of decision is extracted from the hospital information system, along with existing baseline characteristic information. Then, these data are assembled into a state vector strictly according to the feature order, type, and format specified by the state space. Specifically, the first part of the state vector corresponds to static features in the state space, such as directly filling in baseline information such as the patient's age and tumor type; the latter part corresponds to dynamic features, such as filling in the most recently measured CAR-T cell count, the current IL6 concentration, and the trend values ​​of biomarkers calculated based on recent data.

[0141] Taking patient P1 as an example, at a certain moment when a decision needs to be made, the latest integrated data is as follows: static features include: age 55 years, tumor type is diffuse large B-cell lymphoma (coded as number 1); dynamic features include: current tumor burden reduced to 2800, CAR-T cell count is 22 cells / μL, and IL6 concentration is 38 pg / mL. According to the state space definition order, the current state vector constructed for patient P1 is [55, 1, 2800, 22, 38].

[0142] Step 602: Input the state vector into the optimized deep Q-network agent to calculate the Q value of each candidate action in the action space, and take the candidate action with the highest Q value as the target action.

[0143] In this embodiment, the state vector is input into a fully trained and optimized deep Q-network agent. The agent performs one forward propagation computation within the network. The state vector undergoes nonlinear transformations through the input and hidden layers, ultimately generating an array at the output layer. Each element in the array represents the Q-value of the corresponding action, indicating the expected long-term benefit of choosing that action in that state.

[0144] After obtaining the Q-values ​​of all candidate actions, the action index corresponding to the maximum value is found through numerical comparison. Based on this index, the specific action description is retrieved from a predefined action space mapping table, and this action is selected as the target action. For example, for patient P1's current state vector [55,1,2800,22,38], the agent outputs the Q-values ​​of all candidate actions, with the highest value of 9.2 corresponding to index number 7. Querying the action mapping table, index 7 corresponds to the action of infusion dosage. Tocilizumab 4 mg per kilogram was prescribed, with the intention of using it if the IL6 concentration exceeded 50 pg / mL. This regimen was thus designated as Target Action A.

[0145] Step 603: Input the state vector and target action into the calibrated virtual patient physiological model, perform forward simulation, and obtain the efficacy assessment and toxicity risk assessment of the target action.

[0146] In this embodiment, to verify the potential effects of the target action and assess its risks, the current patient's state vector and the selected target action A are input into the calibrated virtual patient physiological model based on the individual patient or a virtual patient with similar characteristics. The model uses the state vector as the initial state and the target action as the input operation command, and performs forward simulation in virtual spacetime.

[0147] The model operates based on its internally coupled tumor, immune, and toxicological dynamics equations, simulating physiological evolution over a future period. After the simulation, the model outputs key simulation results, including efficacy assessments and toxicity risk assessments. Efficacy assessments typically include the predicted rate of change in tumor burden over the simulated time period, while toxicity risk assessments include the predicted highest possible toxicity level and its probability of occurrence.

[0148] For example, taking patient P1 as an example, after inputting the state vector [55,1,2800,22,38] and the target action A into the corresponding virtual patient physiological model for deduction, the model outputs a therapeutic effect assessment. Toxic risk assessment for a 40% predicted decrease in tumor burden within the next week. The probability of a first-degree CRS is about 15 percent, and the probability of a second-degree or higher CRS is less than 5 percent.

[0149] Step 604: Combine the target action, efficacy assessment, and toxicity risk assessment to obtain the target strategy.

[0150] In this embodiment, the three core outputs obtained from the preceding steps—the specific treatment plan, the efficacy prediction of the plan, and the safety warning—are standardized, combined, and encapsulated to generate a complete and highly readable target strategy report. This report directly answers the core questions in clinical decision-making: what to do, what the effects are, and what the risks are.

[0151] For example, taking patient P1 as an example, the generated target strategy will integrate target action A and efficacy assessment. and toxicity risk assessment The following recommendations are made: CAR-T infusion is recommended for patients with P1. The tocilizumab intervention threshold was set at 50 pg / mL per kilogram of tumor tissue. The protocol simulated a projected decrease in tumor burden of approximately 40% over the next week with a low risk of severe toxicity; therefore, it was recommended to follow this protocol and maintain routine monitoring. This structured strategy is the final output that directly supports clinical decision-making.

[0152] This application's embodiments address the inconsistency in model input caused by fragmented and inconsistent data formats in existing technologies by acquiring and managing multi-dimensional clinical data. It provides a core technical means to solve the problem of the lack of personalized and dynamic prediction tools in existing technologies. It overcomes the limitation of obtaining massive decision-making samples through experiments in the real world, creating conditions for training data-driven high-order decision-making models. It realizes a shift from passive response to proactive optimization, and from experience-driven to data- and model-driven collaborative decision-making, improving the systematicness and foresight of clinical decision-making.

[0153] likeFigure 3 As shown, Figure 3 A schematic diagram of a specific embodiment of a CAR-T cell therapy toxicity management optimization system provided in this application includes:

[0154] Data receiving and processing module: As the input and preprocessing center, this module is responsible for collecting multi-dimensional clinical data from patients and generating high-quality, standardized datasets with consistent format through missing value imputation, standardization, and de-identification.

[0155] Patient Digital Twin Module: As the core mechanism model, this module uses digital twin technology to couple tumor, immune, and toxicity dynamics models, and calibrates the model parameters based on individual patient data, thereby constructing a personalized virtual physiological model that can simulate the dynamic changes in tumor burden, cell expansion, and toxic events after CAR-T therapy in the patient.

[0156] Learning Environment Module: This module transforms the calibrated virtual patient physiological model into a reinforcement learning training environment. It explicitly defines the state space describing the patient's condition, the space of available treatments and interventions, and generates an association database for training agents through large-scale simulations.

[0157] Agent Module: As the core of the system's decision learning, this module constructs a deep Q-network with the goal of maximizing rewards. It learns from an association database through a priority experience replay mechanism and utilizes a dual-network architecture for stable training, ultimately learning to output a treatment strategy that optimally balances efficacy and toxicity in a simulated environment.

[0158] Clinical Decision Support Module: As the final application output of the system, this module applies the trained agent to real-time clinical scenarios. It generates recommended treatment plans based on the latest patient data and uses a digital twin model to predict efficacy and risk, ultimately producing a structured target strategy report to provide doctors with personalized decision support.

[0159] Figure 4 A schematic diagram of a specific embodiment of another CAR-T cell therapy toxicity management optimization system provided in this application is shown below. Figure 4 The system may include:

[0160] The acquisition module 21 is used to acquire clinical data of hematologic malignancies patients receiving CAR-T cell therapy. The clinical data includes baseline characteristics, process monitoring data, efficacy evaluation data, toxicity event data, and treatment intervention data.

[0161] Module 22 is used to impute missing values ​​and standardize baseline feature information and process monitoring data to obtain a standardized dataset;

[0162] The construction module 22 is also used to construct a virtual patient physiological model based on a preset tumor cell dynamics model, immune cell dynamics model and toxicity response model using digital twin technology, and to calibrate the model parameters in the virtual patient physiological model using a gradient descent algorithm with clinical data in a standardized dataset as a benchmark, so as to obtain a calibrated virtual patient physiological model.

[0163] The definition module 23 is used to define the state space of the interactive environment based on the static features in the baseline feature information and the dynamic features in the process monitoring data, and to define the action space of the interactive environment based on the infusion dose of CAR-T cells and the preset toxicity intervention strategy, so as to perform simulation and deduction within the interactive environment and obtain the associated database.

[0164] Module 22 is also used to construct a deep Q-network agent with the goal of maximizing the reward function. Training samples are extracted from the associated database through a priority experience replay mechanism. Based on the dual-network training architecture consisting of the main network and the target network, the network parameters of the deep Q-network agent are iteratively optimized according to the training samples and the dual-network training architecture to obtain the optimized deep Q-network agent. The reward function is used to quantify the balance between tumor clearance and toxicity control.

[0165] Input module 24 inputs process monitoring data into the optimized deep Q-network agent to obtain the target policy.

[0166] This application provides a CAR-T cell therapy toxicity management optimization system to implement the aforementioned CAR-T cell therapy toxicity management optimization method. Therefore, the specific implementation of the CAR-T cell therapy toxicity management optimization system can be found in the embodiment section of the CAR-T cell therapy toxicity management optimization method above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0167] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0168] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the steps of the CAR-T cell therapy toxicity management optimization method as described above when performing the computer program.

[0169] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0170] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0171] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0172] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and is operable to perform operations described with reference to the method according to the first aspect of this disclosure when the software is executed (e.g., by one or more processors).

[0173] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the CAR-T cell therapy toxicity management optimization methods in the above embodiments.

[0174] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0175] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0176] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGp) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LpC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (pCI) bus, a pCI-Express (pCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0177] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for optimizing toxicity management of CAR-T cell therapy.

[0178] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0179] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the CAR-T cell therapy toxicity management optimization method described above.

[0180] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0181] The above provides a detailed description of the CAR-T cell therapy toxicity management optimization method and system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method of CAR-T cell therapy toxicity management optimization, characterized in that, The method comprises the following steps: Obtaining clinical data of a blood tumor patient receiving CAR-T cell therapy, wherein the clinical data comprises baseline characteristic information, process monitoring data, efficacy evaluation data, toxicity event data and treatment intervention data; Performing missing value imputation and standardization processing on the baseline characteristic information and the process monitoring data to obtain a standardized data set; According to a preset tumor cell kinetics model, an immune cell kinetics model and a toxicity reaction model, a virtual patient physiological model is constructed by using digital twinning technology, and a gradient descent algorithm is used to calibrate the model parameters in the virtual patient physiological model based on the clinical data in the standardized data set, thereby obtaining a calibrated virtual patient physiological model; The calibrated virtual patient physiological model is used as an interactive environment, the state space of the interactive environment is defined according to the static characteristics in the baseline characteristic information and the dynamic characteristics in the process monitoring data, and the action space of the interactive environment is defined according to the infusion dose of CAR-T cells and a preset toxicity intervention strategy, so that simulation deduction is performed in the interactive environment, thereby obtaining an association database; A deep Q network intelligent agent aiming at maximizing a reward function is constructed, training samples are extracted from the association database by using a priority experience replay mechanism, a double-network training architecture composed of a main network and a target network is used, and the network parameters of the deep Q network intelligent agent are iteratively optimized according to the training samples and the double-network training architecture, thereby obtaining an optimized deep Q network intelligent agent, wherein the reward function is used to quantify the balance between tumor clearance and toxicity control; The process monitoring data is input into the optimized deep Q network intelligent agent, thereby obtaining a target strategy; According to a preset tumor cell kinetics model, an immune cell kinetics model and a toxicity reaction model, a virtual patient physiological model is constructed by using digital twinning technology, comprising: According to the tumor burden, immune status data and inflammation factor indicators in the baseline characteristic information, corresponding model initial parameters are set for the preset tumor cell kinetics model, immune cell kinetics model and toxicity reaction model, respectively; The tumor cell kinetics model, the immune cell kinetics model and the toxicity reaction model with the set initial parameters are coupled and spliced by using digital twinning technology, thereby obtaining a virtual patient physiological model, wherein the virtual patient physiological model is used to simulate the changes of tumor burden, CAR-T cell expansion and toxicity event occurrence of a patient after receiving CAR-T cell therapy.

2. The method of claim 1, wherein, The baseline characteristic information and the process monitoring data are subjected to missing value imputation and standardization processing to obtain a standardized data set, comprising: Identifying first missing values in the baseline characteristic information and second missing values in the process monitoring data; A first imputation value of the first missing values in the baseline characteristic information is generated by using a multiple imputation method to constitute complete baseline characteristic information, and a second imputation value of the second missing values in the process monitoring data is generated by using a time series interpolation algorithm to constitute complete process monitoring data; calculating the mean and the standard deviation of each continuous index in the complete baseline feature information and the complete process monitoring data; based on the mean and the standard deviation, performing standardization processing on each continuous index to obtain a preliminary data set, and performing de-identification processing on the preliminary data set to obtain a standardized data set.

3. The method of claim 1, wherein, calibrating model parameters in the virtual patient physiological model based on clinical data in the standardized data set by using a gradient descent algorithm to obtain a calibrated virtual patient physiological model, including: inputting the clinical data in the standardized data set into the virtual patient physiological model to obtain simulated clinical data; calculating the prediction error between the simulated clinical data and the clinical data; calibrating the model parameters in the virtual patient physiological model based on the prediction error by using a gradient descent algorithm until the prediction error is less than a preset error threshold to obtain a calibrated virtual patient physiological model.

4. The method of claim 1, wherein, constructing a deep Q network agent with the goal of maximizing a reward function, including: designing a network structure based on the state space dimension of the state space, the network structure including an input layer, a hidden layer, and an output layer matching the action space dimension of the action space; constructing a reward function based on the efficacy evaluation data and the toxicity event data, the reward function including positive rewards for tumor clearance and negative penalties for toxicity occurrence; associating the network structure and the reward function to obtain a deep Q network agent.

5. The method of claim 1, wherein, extracting training samples from the association database through a priority experience replay mechanism, based on a double-network training architecture composed of a main network and a target network, and iteratively optimizing network parameters of the deep Q network agent according to the training samples and the double-network training architecture to obtain an optimized deep Q network agent, including: allocating a corresponding priority weight to each sample according to the immediate reward size of each sample in the association database; extracting training samples from the association database through a priority experience replay mechanism according to the priority weight; inputting state data in the training samples into the main network to obtain a first Q value, and inputting next state data in the training samples into the target network to obtain a second Q value, the state data including the static features and the dynamic features; calculating a target Q value according to the immediate reward in the training samples and the second Q value; based on the double-network training architecture, updating the network parameters of the main network a preset number of times according to the difference between the first Q value and the target Q value by using a backpropagation algorithm, copying the updated network parameters into the target network, and returning to extract training samples from the association database through a priority experience replay mechanism according to the priority weight, until a preset number of training rounds are met or a preset performance requirement is reached, to obtain an optimized deep Q network agent.

6. The method of claim 1, wherein, the candidate actions in the action space are composed of the infusion dose of CAR-T cells and a preset toxicity intervention strategy; inputting the process monitoring data into the optimized deep Q network agent to obtain a target strategy, comprising: constructing a state vector corresponding to the state space according to the process monitoring data and the baseline feature information of the hematological tumor patient; inputting the state vector into the optimized deep Q network agent to calculate each candidate action in the action space, obtaining the Q value of each candidate action, and taking the candidate action corresponding to the highest Q value as the target action; inputting the state vector and the target action into the calibrated virtual patient physiological model to perform forward simulation deduction, obtaining efficacy evaluation and toxicity risk evaluation of the target action; combining the target action, the efficacy evaluation and the toxicity risk evaluation to obtain a target strategy.

7. A CAR-T cell therapy toxicity management optimization system, comprising: comprising: an acquisition module configured to acquire clinical data of a hematological tumor patient receiving CAR-T cell therapy, the clinical data including baseline feature information, process monitoring data, efficacy evaluation data, toxicity event data and treatment intervention data; a construction module configured to perform missing value imputation and standardization processing on the baseline feature information and the process monitoring data to obtain a standardized data set; the construction module is further configured to construct a virtual patient physiological model according to a preset tumor cell dynamics model, an immune cell dynamics model and a toxicity reaction model using digital twin technology, and calibrate model parameters in the virtual patient physiological model using a gradient descent algorithm based on clinical data in the standardized data set to obtain a calibrated virtual patient physiological model; a definition module configured to define a state space of the interactive environment according to static features in the baseline feature information and dynamic features in the process monitoring data, and define an action space of the interactive environment according to an infusion dose of CAR-T cells and a preset toxicity intervention strategy, to perform simulation deduction in the interactive environment to obtain an association database; the construction module is further configured to construct a deep Q network agent with a maximum reward function as the target, extract training samples from the association database through a priority experience replay mechanism, construct a double network training architecture based on a main network and a target network, and iteratively optimize network parameters of the deep Q network agent based on the training samples and the double network training architecture to obtain an optimized deep Q network agent, wherein the reward function is used to quantify the balance between tumor clearance and toxicity control; an input module configured to input the process monitoring data into the optimized deep Q network agent to obtain a target strategy; constructing a virtual patient physiological model according to a preset tumor cell dynamics model, an immune cell dynamics model and a toxicity reaction model using digital twin technology, comprising: setting corresponding model initial parameters for the preset tumor cell dynamics model, immune cell dynamics model and toxicity reaction model according to tumor burden, immune status data and inflammatory factor indicators in the baseline feature information; The tumor cell dynamics model, the immune cell dynamics model and the toxicity reaction model with initial parameters set are coupled and spliced by using a digital twin technology to obtain a virtual patient physiological model, which is used to simulate changes in tumor load, CAR-T cell expansion and occurrence of toxicity events of a patient after receiving CAR-T cell therapy.

8. An electronic device, comprising: Comprise: a memory for storing a computer program; a processor for implementing the steps of the CAR-T cell therapy toxicity management optimization method according to any one of claims 1 to 6 when the computer program is executed.

9. A computer-readable storage medium, characterized in that, The computer program is stored in the computer readable storage medium, and the computer program can implement the CAR-T cell therapy toxicity management optimization method according to any one of claims 1 to 6 when executed by the processor.

Images