Artificial intelligence driven neoadjuvant therapy regimen optimization method and system for gastric antrum cancer

Abstract

The present application relates to the technical field of medical decision and artificial intelligence, and discloses an artificial intelligence driven method and system for optimizing a neoadjuvant therapy scheme for gastric antrum cancer, comprising: S1: obtaining clinical data of a patient with gastric antrum cancer and preprocessing; S2: screening core features and constructing compound features by using an iterative ReliefF algorithm; S3: constructing a prediction model based on a K nearest neighbor algorithm of Bayesian optimization, and taking the compound features as inputs of the prediction model to predict a treatment response, and outputting a tumor regression grade probability distribution of the patient; S4: taking the tumor regression grade probability distribution as a core component of a state space, inputting into a double-layer reinforcement learning framework based on a punishment mechanism to optimize a treatment decision, and outputting a set of personalized treatment strategy parameters; and S5: performing real-time scanning and checking on drug dosage parameters in the set of strategy parameters, and outputting a final approved personalized treatment scheme. The present application improves the precision and intelligence level of gastric antrum cancer treatment.

Description

Technical Field

[0001] This invention relates to the fields of medical decision-making and artificial intelligence technology, and in particular to an artificial intelligence-driven method and system for optimizing neoadjuvant treatment plans for gastric antral cancer. Background Technology

[0002] Gastric antral cancer is a common subtype of gastric cancer. Neoadjuvant therapy (such as preoperative chemotherapy) can effectively reduce tumor size, increase surgical resection rates, and improve long-term prognosis. However, the following challenges exist in clinical practice:

[0003] 1. Significant individual differences: Patients vary greatly in age, nutritional status, tumor stage, and molecular characteristics, leading to significant differences in the efficacy and safety of the same treatment regimen in different populations;

[0004] 2. Complex decision variables: Neoadjuvant therapy involves multiple variables such as the type and dosage of chemotherapy drugs, cycle arrangement, and timing of surgery, resulting in a large optimization space;

[0005] 3. Lack of intelligent tool support: Currently, treatment plan design still mainly relies on empirical judgment and lacks personalized decision-making support tools based on big data and machine learning.

[0006] Breakthroughs in artificial intelligence (AI) technology, particularly in feature selection, model optimization, and reinforcement learning, have made it possible to solve the aforementioned problems. By integrating clinical big data with machine learning algorithms, accurate predictive models can be established to assist in the development of individualized neoadjuvant therapy strategies, thereby improving efficacy, reducing toxic side effects, and achieving the goal of patient-centered treatment optimization. Summary of the Invention

[0007] This invention aims to propose an artificial intelligence-driven method and system for optimizing neoadjuvant treatment plans for gastric antral cancer. By introducing advanced feature screening, model optimization, and decision enhancement methods, it addresses the problems of insufficient personalization, coarse parameter configuration, and imperfect strategy evaluation mechanisms in current treatment plan design.

[0008] To achieve the above objectives, the following technical solution is adopted:

[0009] According to a first aspect of the present invention, an artificial intelligence-driven method for optimizing neoadjuvant therapy for gastric antral cancer is provided, comprising the following steps:

[0010] S1: Obtain and preprocess clinical data of patients with gastric antral cancer, including basic information, diagnostic data, treatment data, and follow-up data;

[0011] S2: The iterative ReliefF algorithm is used to screen core features and construct composite features, which include cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammatory indicators and lesion complexity features;

[0012] S3: Construct a prediction model based on the Bayesian-optimized K-Nearest Neighbor (KNN) algorithm, and use the composite features as input to the prediction model to predict the treatment response, outputting the patient's individualized tumor regression grade (TRG) probability distribution;

[0013] S4: The probability distribution of tumor regression level is used as the core component of the state space and input into a two-layer reinforcement learning framework based on a penalty mechanism to optimize treatment decisions and output a set of personalized treatment strategy parameters, which includes recommended chemotherapy drug dosage, treatment cycle duration, and surgical timing suggestions.

[0014] S5: Perform real-time scanning and verification of the drug dosage parameters in the set of personalized treatment strategy parameters, and output the finally approved and safe personalized treatment plan.

[0015] Furthermore, the preprocessing described in step S1 includes: processing missing values ​​using multiple imputation, correcting outliers by combining them with clinical reference ranges, encoding categorical data, and standardizing numerical data.

[0016] Furthermore, in step S2, the use of the iterative ReliefF algorithm to filter core features includes:

[0017] The iterative ReliefF algorithm introduces a dynamic weight update mechanism, which iteratively corrects feature weights based on Manhattan distance. By comparing the feature distance differences between effective samples of the same class and ineffective samples of different classes, the feature weights are cyclically adjusted to screen discriminative features related to efficacy or adverse reactions.

[0018] Furthermore, the iterative ReliefF algorithm introduces a dynamic weight update mechanism. The specific execution process of iteratively correcting feature weights based on Manhattan distance includes:

[0019] Calculate the feature distance between the current sample and similar efficacy samples as the positive correlation weight increment;

[0020] Calculate the feature distance between the current sample and the heterogeneous efficacy sample as a negative correlation weight reduction;

[0021] The final feature weight ranking is obtained through multiple rounds of iterative convergence.

[0022] Furthermore, the construction process of the prediction model built by the Bayesian-optimized K-nearest neighbor algorithm includes: optimizing the KNN hyperparameters through a Gaussian process model and an expectation improvement (EI) strategy, specifically including:

[0023] Define the hyperparameter search space, including the range of values ​​for the number of neighbors K and the distance metric;

[0024] In the initialization phase, N0 sets of hyperparameter combinations are randomly selected to train the initial KNN model, and a Gaussian process model is fitted based on the comprehensive efficacy-safety score of the validation set.

[0025] In the iterative optimization phase, the expected improvement function is used to iteratively select hyperparameters, including: selecting the current potential optimal hyperparameter combination based on the expected improvement function value, training the KNN model based on the current potential optimal hyperparameter combination and evaluating the error on the validation set, updating the Gaussian process model, until the termination condition is met;

[0026] In the output phase, the best historical parameters are selected to construct the final prediction model, and the probability distribution of the patient's tumor regression grade (TRG) is calculated. The efficacy distribution of neighboring samples is weighted and statistically analyzed to obtain the predicted value of the current patient's treatment response probability.

[0027] Furthermore, step S4 includes:

[0028] Use the probability distribution vector as a forced input component of the state space;

[0029] A two-layer Markov decision framework is established, in which the upper layer sets the comprehensive efficacy objective function and optimizes the meta-parameters of the treatment strategy; the lower layer constructs a state space containing TRG probability distribution to simulate the transition of the patient's physiological state.

[0030] The bi-level optimization problem is transformed into a single-level problem by using a strategy suboptimal penalty mechanism.

[0031] By iteratively updating treatment strategy parameters in conjunction with clinical safety constraints, a personalized set of treatment strategy parameters is generated.

[0032] Furthermore, the two-layer reinforcement learning framework is constructed as follows:

[0033] Treatment strategy parameters are used as upper-level optimization variables, and individual adaptability parameters are used as lower-level optimization variables.

[0034] The lower-level Markov decision process includes a state space composed of tumor indicators, an action space composed of treatment operations, and a reward function composed of efficacy indicators.

[0035] The suboptimal penalty term for the strategy is constructed by the difference in value functions or Bellman error.

[0036] Furthermore, the strategy suboptimal penalty mechanism includes:

[0037] Value penalty mode: Calculates the cumulative difference between the current strategy and the theoretically optimal strategy;

[0038] Bellman penalty pattern: Construct policy uniqueness constraints through regularization terms.

[0039] Furthermore, the clinical safety constraints include: the maximum safe dose threshold for chemotherapy drugs, the lower limit of the treatment cycle interval, and the time window range for surgical procedures.

[0040] According to a second aspect of the present invention, an artificial intelligence-driven neoadjuvant therapy optimization system for gastric antral cancer is also provided, comprising:

[0041] The data acquisition and preprocessing module is used to acquire and preprocess clinical data of patients with gastric antral cancer. The clinical data includes basic information, diagnostic data, treatment data, and follow-up data.

[0042] The feature selection and construction module is used to screen core features and construct composite features using the iterative ReliefF algorithm. The composite features include cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammatory indicators, and lesion complexity features.

[0043] The treatment response prediction module constructs a prediction model based on the Bayesian-optimized K-Nearest Neighbor (KNN) algorithm, and uses the composite features as input to the prediction model to predict the treatment response, outputting a patient-specific tumor regression grade (TRG) probability distribution; wherein, the construction of the prediction model includes: optimizing the KNN hyperparameters through a Gaussian process model and an expectation improvement (EI) strategy;

[0044] The treatment decision optimization module is used to take the probability distribution of the tumor regression level as the core component of the state space and input it into a two-layer reinforcement learning framework based on a penalty mechanism to optimize treatment decisions and output personalized treatment plans.

[0045] The safety verification and adjustment module is used to scan and verify the drug dosage parameters in the personalized treatment strategy parameter set in real time, and output the finally approved and safe personalized treatment plan.

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

[0047] 1. This invention designs an iterative ReliefF feature selection algorithm: Based on the traditional ReliefF, a dynamic weight update mechanism is introduced. The Manhattan distance is used to iteratively correct the differences between sample neighborhoods, dynamically adjust the feature weights, effectively reduce the interference of redundant information, highlight discriminative features that are highly correlated with efficacy or adverse reactions, and extract key inputs from high-dimensional clinical variables.

[0048] 2. This invention designs a Bayesian-optimized K-Nearest Neighbor (KNN) classification algorithm: based on a Gaussian process model and an expectation improvement (EI) strategy, the key hyperparameters of the KNN model (such as the number of neighbors K and the distance measurement method) are adaptively optimized to improve the model's generalization ability and performance stability in predicting individualized treatment plans.

[0049] 3. This invention designs a two-layer reinforcement learning algorithm based on a penalty mechanism: the complex problem of "optimal treatment strategy search" is modeled as a constrained two-layer optimization problem, where the lower-layer strategy models patient state transitions through a Markov Decision Process (MDP), and the upper-layer objective models the overall therapeutic effect. By introducing a penalty function, the two-layer optimization is transformed into a single-layer problem, significantly reducing the difficulty of solving it.

[0050] 4. This invention designs a value penalty gradient calculation method: based on the difference in value functions between the current strategy and the optimal strategy, a quantitative index of "strategy suboptimality" is defined, and its gradient is calculated for model updates. This method can accurately measure the benefit gap of the current treatment strategy and guide the optimization direction.

[0051] 5. This invention designs a Bellman penalty gradient mechanism: for scenarios where the policy is not unique, a regularized objective function based on Bellman error is introduced, and a strongly convex structure is constructed to ensure the uniqueness of policy optimality, thereby achieving efficient and differentiable policy learning and optimization.

[0052] In summary, this invention constructs an intelligent treatment optimization framework oriented towards individual differences by combining three key steps: feature selection, model training, and strategy optimization. It possesses good interpretability, flexibility, and scalability, and can effectively guide clinicians in formulating personalized neoadjuvant therapy strategies, thereby improving the precision and intelligence of gastric antral cancer treatment.

[0053] 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

[0054] 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:

[0055] Figure 1 This is a flowchart illustrating an artificial intelligence-driven method for optimizing neoadjuvant therapy for gastric antral cancer according to an embodiment of the present invention.

[0056] Figure 2This is a flowchart of the Bayesian optimized KNN prediction model construction process according to an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the reinforcement learning optimization process according to an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the architecture of an artificial intelligence-driven neoadjuvant therapy optimization system for gastric antral cancer 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] The neoadjuvant therapy for gastric antral cancer of the present invention is a treatment method such as chemotherapy, targeted therapy or radiotherapy performed on patients with gastric antral cancer before radical surgery. Its purpose is to shrink the tumor, reduce the stage, and create favorable conditions for subsequent surgery.

[0062] Figure 1 This is a flowchart illustrating an artificial intelligence-driven method for optimizing neoadjuvant therapy for gastric antral cancer, according to an embodiment of the present invention. Figure 1 As shown, an AI-driven method for optimizing neoadjuvant therapy for gastric antral cancer includes the following steps:

[0063] S1: Obtain and preprocess clinical data from patients with gastric antral cancer. The clinical data includes basic information, diagnostic data, treatment data, and follow-up data.

[0064] Step S1.1: Clinical Data Collection

[0065] Data related to clinical research on gastric antral cancer from 2015 to 2025 in tertiary hospitals were collected using the hospital's electronic medical record (EMR) system, pathology database, picture archiving and communication system (PACS), and follow-up management system, covering a multi-center collaborative research cohort. Patients pathologically diagnosed with gastric antral cancer (according to the WHO classification of digestive system tumors) and possessing complete medical and follow-up records were included. Exclusion criteria included patients with other malignant tumors or those with more than 30% missing medical records.

[0066] The data includes:

[0067] Basic information includes age (accurate to the year), gender, ethnicity, BMI, smoking history (in packs per year), drinking history (recorded for years and amount of alcohol consumed), and comorbidities such as hypertension and diabetes.

[0068] Diagnostic data: In terms of histological type, adenocarcinoma types are subdivided, such as papillary adenocarcinoma, tubular adenocarcinoma, mucinous adenocarcinoma, etc., and Lauren classification (intestinal type or diffuse type) is clarified; the diagnostic criteria for pyloric obstruction include the degree of pyloric stenosis under gastroscopy (divided into mild, moderate, and severe) and the passage time of gastrointestinal contrast agent; staging examination is performed by CT / MRI to determine TNM staging (T1-T4, N0-N3, M0-M1), and the distribution of metabolically active foci shown by PET-CT is recorded.

[0069] Treatment data: Surgical information includes the specific surgical technique of the partitioned gastrojejunostomy (such as Billroth II or Roux-en-Y anastomosis), operation time, intraoperative blood loss, and the number of lymph nodes dissected (distinguishing between D1 and D2 procedures); Neoadjuvant therapy includes the duration of chemotherapy regimens (such as FLOT, SOX), targeted drugs (such as trastuzumab), and the radiation dose (unit: Gy) and fractionation method; Postoperative adjuvant therapy records include the start time of chemotherapy, number of cycles, dose adjustments, and adverse reaction grading (refer to CTCAE 5.0 standards).

[0070] Follow-up data: Disease progression includes recurrence time, metastatic sites (such as liver, lung, peritoneum, etc.) and ECOG score at the time of progression; survival data includes overall survival (OS), disease-free survival (DFS) and status at the last follow-up (survival, death or loss to follow-up); adverse event records include the occurrence time and management of postoperative complications (such as anastomotic leakage, gastroparesis) and treatment-related toxicities (such as myelosuppression level).

[0071] Step S1.2: Clinical data preprocessing

[0072] Missing value handling: For continuous variables (such as age and BMI), multiple imputation (MICE) is used to build a predictive model based on relevant variables such as gender and diagnostic stage; for categorical variables (such as surgical method), if the missing rate is <5%, it is marked as "unknown", otherwise it is filled in according to the mode of patients in the same stage; when key indicators (such as TNM stage) are missing, the original pathology report or imaging data is traced back, and if it cannot be filled in, the sample is removed (the removal rate is controlled within 5%).

[0073] Outlier detection: Numerical data (such as white blood cell count) are analyzed using a 3-fold standard deviation method (Z-score > 3 or < -3) combined with clinical reference ranges (such as the normal range of white blood cell count 4-10 × 10⁻⁶). 9 / L) Identification, abnormal values ​​are corrected after review by clinicians (such as extreme values ​​caused by laboratory instrument errors) or marked as "suspicious values".

[0074] Categorical data coding: Ordinal variables (such as ECOG scores of 0-4) are directly converted to numerical values; unordered variables (histological types) are coded using one-hot encoding; multi-category variables are coded using labels and the mapping rules are specified.

[0075] Numerical data standardization: Blood routine indicators (such as hemoglobin) are standardized using Z-scores to facilitate comparison of indicators with different dimensions; survival time (months) retains the original value, but the right skewness effect is reduced through logarithmic transformation.

[0076] Step S1 systematically collects multi-center, multi-modal clinical data on gastric antral cancer and preprocesses it through a standardized process (including multiple imputation, outlier handling, coding, and standardization) to provide high-quality, standardized input data for model construction.

[0077] S2: The iterative ReliefF algorithm is used to screen core features and construct composite features, which include cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammatory indicators and lesion complexity features;

[0078] To construct an accurate and interpretable personalized neoadjuvant therapy optimization model, it is necessary to extract key features closely related to efficacy and safety from massive clinical data. This step S2 focuses on feature selection and feature construction to ensure that the input features are both representative and medically interpretable, thereby improving the model's generalization ability and clinical applicability.

[0079] Step S2.1: Core Feature Selection

[0080] This invention employs the iterative ReliefF algorithm to screen core variables related to adjuvant therapy for gastric antral cancer from preprocessed high-dimensional clinical features. By incorporating class information from neighboring samples, this algorithm evaluates the ability of features to distinguish different treatment responses or adverse events, making it particularly suitable for nonlinear, high-dimensional, and mixed (numerical + classification) feature scenarios.

[0081] The iterative ReliefF algorithm introduces a dynamic weight update mechanism based on the traditional ReliefF algorithm. It iteratively adjusts feature weights based on Manhattan distance, weakening redundant information and highlighting discriminative features. By comparing the feature distance differences between effective samples of the same class and ineffective samples of different classes, the feature weights are iteratively adjusted to screen discriminative features related to efficacy or adverse reactions.

[0082] Furthermore, the iterative ReliefF algorithm introduces a dynamic weight update mechanism. The specific execution process, based on the Manhattan distance to iteratively correct feature weights, includes: calculating the feature distance between the current sample and similar efficacy samples as the positive correlation weight increment; calculating the feature distance between the current sample and dissimilar efficacy samples as the negative correlation weight decrement; and obtaining the final feature weight ranking through multiple iterations. More specifically, its iterative update formula is as follows:

[0083]

[0084] : No. In the next iteration, the features The weights; η: learning rate, which controls the speed of weight updates (typically 0.1-0.5); K: number of nearest neighbor samples; x: current sample; The nearest neighbor that belongs to the same class as x (e.g., effective treatment); The nearest neighbor of x that is a different class (e.g., treatment is ineffective or a serious adverse event occurs). :feature The Manhattan distance between samples x and y.

[0085] The iterative ReliefF algorithm terminates its iteration when the change in feature weights is less than a set threshold. (like The maximum number of iterations (T) is reached, or the maximum number of iterations (e.g., 50) is achieved. The impact of each feature on model performance is evaluated using 10-fold cross-validation. Ultimately, the original dimensionality of the clinical variables is reduced to the core feature set, significantly reducing computational complexity while retaining sufficient medical discriminative information.

[0086] Step S2.2: Construction of composite features

[0087] After completing the key feature selection, clinical domain knowledge is further incorporated to construct composite features, thereby enhancing the algorithm's ability to model complex medical events. Typical feature construction methods include:

[0088] Cumulative drug exposure characteristic: The cumulative dose characteristic is constructed by multiplying the number of adjuvant therapy cycles by the specific drug dose in each cycle, which measures the individual's total exposure to chemotherapy.

[0089] Postoperative recovery risk characteristics: The risk of postoperative complications is comprehensively assessed by integrating variables such as preoperative nutritional risk index (NRS2002 score), intraoperative operation time, and blood loss.

[0090] Comprehensive inflammatory markers: Combining neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), etc., to form a "systemic inflammation score" characteristic, used to assess an individual's immune response status;

[0091] Complexity characteristics of lesions: Combining information such as lesion diameter, number, and extent of invasion from imaging studies to quantify the degree of tumor burden.

[0092] Through the aforementioned feature engineering process, the iterative ReliefF algorithm, employing a dynamic weight update mechanism, filters out the core features most relevant to efficacy and safety from high-dimensional data. These features are then combined with prior clinical knowledge to construct composite features, effectively enhancing the discriminative power of the features and the medical interpretability of the model. The model input not only contains highly informative raw variables but also embeds multi-dimensional medical knowledge, laying a solid foundation for personalized treatment strategy recommendations in step S3.

[0093] S3: Construct a prediction model based on the Bayesian-optimized K-Nearest Neighbor (KNN) algorithm, and use the composite features as input to the prediction model to predict the treatment response, outputting the patient's individualized tumor regression grade (TRG) probability distribution;

[0094] To achieve personalized optimization of neoadjuvant therapy regimens for patients with gastric antral cancer, step S3 constructs a model architecture centered on the Bayesian-optimized K-nearest neighbor (KNN) algorithm, such as... Figure 2 The diagram shows the flowchart of the Bayesian optimized KNN prediction model construction according to an embodiment of the present invention. The model input consists of clinical features, i.e., composite features (such as pathological stage, nutritional status, biochemical indicators, previous treatment history, etc.), which have been screened and constructed through feature engineering. The output consists of individualized treatment strategy suggestions (such as chemotherapy regimen selection, cycle duration optimization, and whether to perform surgery earlier) and its efficacy and safety prediction results (such as tumor regression rate, TRG grade, and adverse reaction risk).

[0095] To improve the generalization ability and prediction performance of the KNN model, a Bayesian optimization algorithm is introduced to globally optimize its key hyperparameters (including the number of neighbors K, distance metric, etc.). The core of Bayesian optimization lies in the acquisition function. The design is used to select the next most worthwhile point to try in an unevaluated hyperparameter space. This step employs an expected improvement (EI) strategy, the objective of which is to maximize the expected improvement to the current best result, as shown in the following formula:

[0096]

[0097] The acquisition function is used to select the next most worthwhile point to try in the unevaluated hyperparameter space. . , which represents the standardized distance between the current predicted mean and the optimal value. : These are the predicted mean and standard deviation of the Gaussian process model at point H, respectively. : The current best observation (i.e., the minimum error). : The cumulative distribution function of the standard normal distribution. : The probability density function of the standard normal distribution.

[0098] The Bayesian-optimized K-nearest neighbor algorithm consists of the following four stages:

[0099] Phase 1: Define the hyperparameter space Ω: the number of neighbors K∈{3,5,7,10,15,20}, and the distance metric M includes Euclidean distance, Manhattan distance, and cosine similarity, etc.; the hyperparameter space Ω is discrete and is composed of the Cartesian product of the set of values ​​of K and the set of values ​​of M.

[0100] Phase Two: Initial Modeling

[0101] Random selection Group of hyperparameter combinations: ,in, To initialize the number of samples, in this embodiment of the invention, the value is 5, which determines the number of hyperparameter combinations that need to be evaluated during the initialization phase. Let i be the i-th hyperparameter combination, which is the object to be evaluated in this iteration; Let be the number of neighbors in the i-th hyperparameter combination. One component of , whose value comes from the hyperparameter space Ω; Let be the distance metric function in the i-th hyperparameter combination. The other component, whose value comes from the hyperparameter space Ω, is optional. {Euclidean distance, Manhattan distance, cosine similarity}.

[0102] For each group Training the KNN model: .

[0103] in, The training dataset consists of patient feature vectors processed by steps S1 and S2 and their corresponding treatment response labels; This is the training function for the KNN model, which is based on the input hyperparameters. Configure the model (e.g., set the number of neighbors to 1). The distance metric is And fit the data .

[0104] The performance of the KNN model was evaluated on the validation set, with the objective function being the combined efficacy (TRG grade, tumor regression rate) and safety indicators (such as the incidence of ≥ grade 3 adverse reactions).

[0105]

[0106] in, : The overall performance score of the i-th model; the higher the score, the better the model performance. Tumor Regression Grade (TRG) prediction accuracy represents the proportion of TRG grades predicted by the model that match the actual values. The false positive rate for serious toxic reactions is (1 - specificity). Specifically, it represents the proportion of patients who did not actually experience grade ≥3 toxic reactions, but were incorrectly predicted by the model to have such reactions. The calculation formula is as follows: Where FP (False Positive) represents the number of false positives and TN (True Negative) represents the number of true negatives. α: The weighting coefficient for efficacy accuracy, a pre-defined constant ranging from [0, 1], used to measure... The importance of α in the total score, for example, α is 0.7; β: security weighting coefficient, is a pre-set constant with a value in the range [0, 1], used to measure security indicators. The importance of α and β in the total score is determined by factors such as β being 0.3 and usually satisfying α + β = 1. The specific values ​​of α and β can be set by clinical experts based on the priority considerations of efficacy and safety, or they can be verified and determined on historical data through hyperparameter optimization methods such as grid search.

[0107] All obtained during the initialization phase Given 10 data pairs (hyperparameter combinations, performance scores), fit an initial Gaussian process (GP) surrogate model:

[0108]

[0109] Here, GP stands for Gaussian process surrogate model, which is used to establish a probabilistic mapping from the hyperparameter space Ω to the performance score. This model will be continuously updated in subsequent iterations. Gaussian process fitting function: a standard machine learning regression method used to learn a probabilistic model based on the input-output pair.

[0110] Input: Training dataset Validation set ;

[0111] Output: Initial Gaussian process model GP.

[0112] Phase 3: Bayesian Iterative Optimization (total) Second-rate)

[0113] According to the acquisition function Select the current potentially optimal combination of hyperparameters ;

[0114] Based on the current potential optimal hyperparameter combination Train the KNN model and evaluate the error on the validation set. , It is the error function;

[0115] Update the Gaussian process model, Add training data.

[0116] Phase Four: Termination and Output

[0117] If the data collection function is collected in 5 consecutive rounds Improvement of less than 1%, or reaching the maximum number of iterations. If the termination condition is met, the optimization is terminated and the optimal hyperparameter combination is output.

[0118] The final KNN model is constructed to recommend personalized treatment plans and output the predicted treatment response level and adverse reaction probability. Within the KNN framework, the following settings are defined: Representing all possible TRG grades (0 for complete remission, 4 for no response), the patient sample... Predicted to be TRG level probability The following formula can be used to calculate the predicted probability of treatment response for the current patient by weighted statistical analysis of the efficacy distribution of neighboring samples:

[0119]

[0120] The current patient sample to be predicted is the input feature vector (including clinical data, imaging features, biochemical indicators, etc.). The rating label indicates the actual level of treatment response; For a specific TRG level whose predicted probability needs to be calculated, ; The set of possible values ​​for the level includes five levels from 0 to 4; :sample of The set of nearest neighbor samples represents the set of samples with... The closest in the feature space One training sample; : The actual TRG level of the i-th neighbor sample; Indicator function, if The value is 1 if the value is 1, otherwise it is 0, used to count the number of times the value is 0. Does each neighbor belong to a certain level? ;

[0121] The weight of the i-th neighbor, representing its contribution to the prediction result, is defined as:

[0122]

[0123] Sample x and its i-th neighbor sample The distance between them can be Euclidean distance, Manhattan distance or cosine distance. A small constant to prevent the denominator from being zero is generally taken as 10. -6 .

[0124] This formula outputs the conditional probability of a certain TRG level by weighted statistical analysis of the proportion of samples belonging to a certain TRG level among K neighbors. For example, if a patient's predicted result is "the probability of TRG level 2 is 60% and the probability of level 3 is 30%", the doctor can assess whether intensive treatment is needed to further improve the degree of remission, or to make conservative intervention based on the risk.

[0125] This modeling process enables nonlinear modeling and efficient optimization of complex clinical variables, improving the clinical interpretability and practicality of the model under multi-indicator constraints, and providing intelligent and personalized decision support for adjuvant treatment of gastric antral cancer. Step S3 constructs a Bayesian-optimized KNN model, which automatically searches for the optimal hyperparameter combination to accurately predict the probability distribution of the patient's individualized tumor regression grade (TRG), providing a reliable basis for treatment decisions.

[0126] S4: The probability distribution of tumor regression level is used as the core component of the state space and input into a two-layer reinforcement learning framework based on a penalty mechanism to optimize treatment decisions and output a set of personalized treatment strategy parameters, which includes recommended chemotherapy drug dosage, treatment cycle duration, and surgical timing suggestions.

[0127] To further achieve personalized optimization of neoadjuvant therapy regimens for patients with gastric antral cancer, this invention proposes a two-layer reinforcement learning algorithm based on a penalty mechanism. This algorithm models the optimization of treatment strategies as a two-layer Markov decision problem and introduces value-penalized gradients and Bellman-penalized gradients, transforming the originally difficult-to-solve two-layer problem into a single-layer optimization problem, thereby achieving intelligent and adaptive treatment decision modeling.

[0128] like Figure 3 The diagram shown illustrates the reinforcement learning optimization process of an embodiment of the present invention. Step S4 includes the following operations: using the probability distribution vector obtained in step S3 as a forced input component of the state space in step S4; establishing a two-layer Markov decision framework, wherein the upper layer sets a comprehensive efficacy objective function and optimizes the meta-parameters of the treatment strategy; the lower layer constructs a state space containing the TRG probability distribution to simulate the patient's physiological state transitions; transforming the two-layer optimization into a single-layer problem through a strategy suboptimal penalty mechanism; and iteratively updating the treatment strategy parameters in conjunction with clinical safety constraints to generate a personalized treatment strategy parameter set. The clinical safety constraints include: the maximum safe dose threshold of chemotherapy drugs, the lower limit of the treatment cycle interval, and the surgical implementation time window range.

[0129] Step S4.1: Optimize Problem Modeling

[0130] The two-layer reinforcement learning framework is constructed as follows: treatment strategy parameters are used as upper-layer optimization variables, and individual adaptive parameters are used as lower-layer optimization variables. The lower-layer Markov decision process includes a state space composed of tumor indicators, an action space composed of treatment operations, and a reward function composed of efficacy indicators. A policy suboptimal penalty term is constructed through value function differences or Bellman error. Specifically, the two-layer structure is represented by the following formula:

[0131]

[0132] θ: Upper-level optimization variable, representing meta-parameters of the treatment strategy (such as dose adjustment, chemotherapy cycle, etc.). Lower-level optimization variables represent personalized strategy parameters (such as patient-specific dose control strategies). The upper-level objective function measures the overall efficacy of the current treatment strategy, such as treatment response rate and side effect risk. :Strategy In Markov Decision Process (MDP) models Below, from the initial state distribution The expected cumulative return (value function) is used to simulate the evolution of a patient's physiological state under a given strategy. : Parametric Markov Decision Process: S: State space, such as tumor size, blood routine, biochemical indicators; A: Action space, representing treatment options (such as whether to add medication or whether to delay treatment); : Parameterized reward function, which measures the contribution of state-action pairs to the treatment effect; : State transition probability; τ: Regularization coefficient; h: State regularization term, which improves policy smoothness.

[0133] Step S4.2: Penalty Restructuring and Single-Layer Transformation

[0134] To simplify the solution, the constrained problem is transformed into an unconstrained form:

[0135]

[0136] : Penalty term (valued penalty gradient or Bellman penalty gradient, see steps S4.4.1 and S4.4.2 below), characterizing the deviation between the current policy and the optimal policy. λ>0: Penalty coefficient, controlling the degree of relaxation of lower-level constraints. : The transformed single-layer optimization objective.

[0137] By refactoring with penalties, the algorithm explicitly expresses the "the strategy should be optimal" constraint as a penalty term in the objective, thereby achieving unconstrained optimization.

[0138] Step S4.3: Gradient Update Mechanism

[0139] The model is trained using projective gradient descent, iteratively updating the treatment strategy and individualized control parameters.

[0140]

[0141] : The current parameters in the k-th iteration. β>0: The learning rate. Projection operation ensures that parameters meet medical constraints (such as upper dose limits and treatment cycle restrictions). The upper-level objective function is about The gradient. Penalty items regarding The gradient.

[0142] Step S4.4: Penalty term gradient calculation method

[0143] Step S4.4.1: Value-penalized gradient

[0144] Value penalty quantifies the suboptimality of a policy by calculating the difference between the value function of the current policy and the optimal policy.

[0145]

[0146] in, : Consists of parameter θ and hyperparameter The defined Markov Decision Process (MDP) model characterizes environmental dynamics. ρ: Initial state distribution, used to define the initial state probability distribution for value function calculation. Based on the current policy with policy parameter φ, determine the action selection rules for each state. :by The parameter is any strategy used as the traversal object when searching for the optimal strategy. :Strategy In Markov Decision Process (MDP) models Below, from the initial state distribution The expected cumulative return (value function) is used to simulate the evolution of a patient's physiological state under a given strategy. :Model All possible strategies The corresponding maximum expected cumulative return.

[0147] The gradient expression for the value penalty is:

[0148]

[0149] : In parameters The optimal strategy under the given conditions, that is, the strategy that makes the value function... Maximization strategy. : Lower-level optimal strategy The value function with respect to The gradient. The gradient of the current policy value function with respect to parameter θ reflects the rate at which parameter θ affects the performance of the current policy. It is calculated as follows:

[0150]

[0151] : No. The state at any given time is generated by the state transition process. : No. Always according to strategy The action chosen. : Under parameter θ, from state Execute action The subsequent state transition probability function. The reward function, defined by parameter θ, represents the state... Execute action The instant reward received. Discount factor: Used to calculate the present value weight of future rewards, with a value range of [value missing]. . This indicates that the expected value of the entire expression within the brackets [] is calculated based on the given current state. and actions Next, the next state It is determined by the state transition probability function Randomly generated.

[0152] Step S4.4.2: Bellman penalized gradient

[0153] Bellman penalty is based on Bellman error to construct a penalty term, and is suitable for scenarios where the policy is not unique.

[0154]

[0155] The minimum value corresponds to the target value of the optimal strategy at the next lower level. The regularized Bellman objective function consists of the inner product of the policy and value functions and a regularization term. :state Next strategy and The inner product, For state The distribution of actions below; Regularization term The penalty coefficient is... Ensure strategy smoothness.

[0156] : The negative optimal Q value, used to construct the linear term of the Bellman objective function. Strategy class, consisting of parameters Defined set of strategies It is a convex set. :Strategy exist The Q-value function under the state, i.e. Execute action Expected discount return after:

[0157]

[0158] : The reward function defined by parameter θ gives the immediate reward for performing action a in state s. : State transition probability, under parameter θ, the probability of transitioning from state s to the next state by performing action a. The probability of. Strategy π in the model Next state The value function. γ: Discount factor, used to calculate the present value of future rewards, with a range of values... .

[0159] The expression for the Bellman penalty gradient is:

[0160]

[0161] The gradient of the Bellman penalty function with respect to the parameter θ guides the direction of parameter optimization. The gradient of the Q-value function with respect to the parameter θ consists of the expected values ​​of the immediate reward gradient and the future value gradient. Based on state distribution ρ and current policy The expected distribution of actions. Based on state distribution ρ and optimal policy The expected distribution of actions.

[0162] Step S4 innovatively integrates the TRG probability distribution into the state space and employs a two-layer reinforcement learning algorithm based on a penalty mechanism to transform the complex medical decision-making problem into a solvable optimization problem, outputting personalized treatment strategy parameters that balance efficacy and safety. Through this framework, the algorithm can automatically optimize adjuvant treatment strategies for gastric antral cancer tailored to individual differences. The upper-layer optimization reflects the overall goal setting driven by medical knowledge (such as maximizing efficacy and minimizing side effects), while the lower-layer strategy dynamically matches the individual's disease evolution process through reinforcement learning, constructing an intelligent, adaptive, and personalized cancer treatment system.

[0163] S5: Perform real-time scanning and verification of the drug dosage parameters in the set of personalized treatment strategy parameters, and output the finally approved and safe personalized treatment plan.

[0164] In step S5, the drug dosage parameters in the personalized treatment strategy parameter set generated in step S4 are scanned and verified in real time. If any dosage parameter is detected to exceed its corresponding maximum safe dosage threshold, a dynamic intervention mechanism is triggered, and the finally approved and safe personalized treatment plan is output. Specifically, this includes the following steps:

[0165] S5.1 Real-time detection: The strategy parameters generated in step S4 are scanned in real time to detect whether they exceed hard constraints such as the maximum safe dose threshold;

[0166] S5.2 Dynamic Intervention: Once a dose breakthrough event is detected, the dose decay mechanism is immediately triggered, and the adjusted dose is calculated according to the following formula: Adjusted dose = min(current recommended dose) Safety threshold) Where λ is the attenuation coefficient, which is determined by referring to Table 1 below based on the patient's creatinine clearance rate (CrCl). Table 1 shows the correspondence between creatinine clearance rate and attenuation coefficient λ; t is the cumulative monitoring time after the dose overrun event occurs, in days, starting from when the system detects the dose overrun, and increasing daily.

[0167] Table 1. Relationship between creatinine clearance rate and attenuation coefficient λ

[0168]

[0169] S5.3 Output: Outputs the final approved, safe, and individualized treatment plan.

[0170] Step S5 is a critical redundancy design to ensure the safe clinical implementation of the treatment plan. Acting as a real-time watchdog, it ensures that the strategy generated in step S4 undergoes a final safety verification before clinical deployment, preventing dangerous doses due to model uncertainties or boundary conditions. A dose attenuation mechanism performs a final hard constraint verification on the optimized strategy, ensuring that any output treatment plan absolutely meets clinical safety standards, which is a crucial guarantee for the reliable implementation of the system.

[0171] This invention discloses an AI-driven method for optimizing neoadjuvant treatment for gastric antral cancer. It comprehensively utilizes key technologies such as iterative ReliefF feature selection, Bayesian-optimized KNN prediction models, and penalized two-layer reinforcement learning to construct a complete closed-loop system encompassing data preprocessing, feature engineering, response prediction, decision optimization, and safety monitoring. This invention achieves true personalization of treatment plans, predicting responses and customizing strategies based on patients' unique clinical characteristics. By integrating efficacy and safety in multi-objective optimization, it significantly improves the accuracy and safety of treatment. Finally, this invention possesses good clinical interpretability and practicality, effectively assisting physicians in decision-making and potentially improving treatment effectiveness and patient quality of life.

[0172] Figure 4 This is a schematic diagram of the architecture of an artificial intelligence-driven neoadjuvant therapy optimization system for gastric antral cancer according to an embodiment of the present invention; as shown below. Figure 4 As shown, an AI-driven neoadjuvant therapy optimization system for gastric antral cancer includes:

[0173] The data acquisition and preprocessing module 210 is used to acquire and preprocess clinical data of patients with gastric antral cancer. The clinical data includes basic information, diagnostic data, treatment data, and follow-up data.

[0174] The feature selection and construction module 220 is used to screen core features and construct composite features using the iterative ReliefF algorithm. The composite features include cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammatory indicators, and lesion complexity features.

[0175] The treatment response prediction module 230 constructs a prediction model based on the Bayesian-optimized K-nearest neighbor algorithm and uses the composite features as input to the prediction model to predict the treatment response, outputting a patient-specific tumor regression level probability distribution; wherein, the construction of the prediction model includes: optimizing the KNN hyperparameters through a Gaussian process model and an expectation improvement strategy;

[0176] The treatment decision optimization module 240 is used to take the probability distribution of the tumor regression level as the core component of the state space and input it into a two-layer reinforcement learning framework based on the penalty mechanism to optimize the treatment decision and output a set of personalized treatment strategy parameters, which includes recommended chemotherapy drug dosage, treatment cycle duration and surgical timing suggestions.

[0177] The safety verification and adjustment module 250 is used to scan and verify the drug dosage parameters in the personalized treatment strategy parameter set in real time, and output the finally approved and safe personalized treatment plan.

[0178] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0179] The acquisition, storage, and application of user personal information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0180] 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.

[0181] 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.

[0182] 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. An artificial intelligence driven optimization method for neoadjuvant therapy regimen of gastric antrum cancer, characterized in that, The method comprises the following steps: S1: obtaining and preprocessing clinical data of a gastric antrum cancer patient, the clinical data comprising basic information, diagnosis data, treatment data and follow-up data; The preprocessing comprises: using multiple imputation method to process missing values, correcting abnormal values in combination with clinical reference ranges, encoding classification data, and standardizing numerical data; S2: screening core features and constructing composite features using an iterative ReliefF algorithm, the composite features comprising cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammation indicators and lesion complexity features; The iterative ReliefF algorithm introduces a dynamic weight updating mechanism, iteratively corrects feature weights based on Manhattan distance, and adjusts feature weights in cycles to screen discriminative features related to efficacy or adverse reactions by comparing feature distance differences between same-class effective samples and different-class ineffective samples; wherein the specific execution process of the iterative ReliefF algorithm introducing a dynamic weight updating mechanism and iteratively correcting feature weights based on Manhattan distance comprises: calculating the feature distance between the current sample and the same-class effective sample as the positive correlation weight increment; calculating the feature distance between the current sample and the different-class effective sample as the negative correlation weight decrement; and obtaining the final feature weight ranking through multiple rounds of iteration convergence; S3: constructing a prediction model based on a K-nearest neighbor algorithm optimized by Bayes, and taking the composite features as inputs of the prediction model to predict treatment response and output individualized tumor regression grade probability distribution of the patient; The construction process of the prediction model constructed by the K-nearest neighbor algorithm optimized by Bayes comprises: optimizing KNN hyperparameters through a Gaussian process model and an expected improvement strategy, specifically comprising: Defining the hyperparameter search space, including the value range of the number of neighbors K and the distance measurement method; In the initialization stage, N0 groups of hyperparameter combinations are randomly selected to train an initial KNN model, and a Gaussian process model is fitted based on the validation set efficacy-safety comprehensive score; In the iterative optimization stage, the expected improvement function is used to iteratively select hyperparameters, including: selecting the current potential optimal hyperparameter combination based on the expected improvement function value, training the KNN model based on the current potential optimal hyperparameter combination and evaluating the error on the validation set, updating the Gaussian process model, and updating the Gaussian process model until the termination condition is met; In the output stage, the historical optimal parameters are selected to construct the final prediction model, and the tumor regression grade probability distribution of the patient is calculated, the efficacy distribution of the neighbor samples is weighted and counted as the current patient treatment response probability prediction value; S4: taking the tumor regression grade probability distribution as a core component of the state space, inputting it into a double-layer reinforcement learning framework based on a punishment mechanism to optimize treatment decisions, and outputting a set of individualized treatment strategy parameters, including recommended chemotherapy drug doses, treatment cycle lengths and surgery timing suggestions, comprising: Taking the probability distribution vector as a mandatory input component of the state space; A double-layer Markov decision framework is established, wherein an upper layer sets a comprehensive therapeutic effect target function to optimize therapeutic strategy element parameters, and a lower layer constructs a state space including a TRG probability distribution to simulate patient physiological state transitions; The double-layer optimization is converted into a single-layer problem through a strategy suboptimality penalty mechanism; Therapeutic strategy parameters are iteratively updated in combination with clinical safety constraint conditions to generate a personalized therapeutic strategy parameter set; The double-layer reinforcement learning framework is constructed in the following manner: Therapeutic strategy parameters are used as upper-layer optimization variables, and individual adaptability parameters are used as lower-layer optimization variables; The lower-layer Markov decision process includes a state space composed of tumor indicators, an action space composed of therapeutic operations, and a reward function composed of therapeutic effect indicators; A strategy suboptimality penalty term is constructed through a value function difference or a Bellman error; S5: Real-time scanning and verification are performed on drug dosage parameters in the personalized therapeutic strategy parameter set, and a final approved safe-to-execute personalized treatment plan is output.

2. The artificial intelligence-driven optimization method of a new adjuvant therapy regimen for gastric antrum cancer according to claim 1, characterized in that, The strategy suboptimality penalty mechanism includes: A value penalty mode: The cumulative reward difference between the current strategy and the theoretically optimal strategy is calculated; A Bellman penalty mode: A strategy uniqueness constraint condition is constructed through a regularization term. 3.The method of claim 2, wherein the method is characterized by, The clinical safety constraint conditions include a maximum safe dose threshold of chemotherapeutic drugs, a lower limit of treatment cycle interval time, and a surgical implementation time window range.

4. An artificial intelligence driven optimization system for neoadjuvant therapy regimen of gastric antrum cancer, for implementing the artificial intelligence driven optimization method of neoadjuvant therapy regimen of gastric antrum cancer according to any one of claims 1-3, characterized in that, It includes: A data acquisition and preprocessing module for obtaining and preprocessing clinical data of a gastric antrum cancer patient, the clinical data including basic information, diagnosis data, treatment data, and follow-up data; A feature selection and construction module for screening core features and constructing composite features using an iterative ReliefF algorithm, the composite features including cumulative drug exposure features, postoperative recovery risk features, comprehensive inflammation indicators, and lesion complexity features; A treatment response prediction module for constructing a prediction model based on a K-nearest neighbor algorithm of Bayesian optimization and inputting the composite features into the prediction model to predict treatment response and output a patient-specific tumor regression grade probability distribution; wherein the construction of the prediction model includes optimizing KNN hyperparameters through a Gaussian process model and an expected improvement strategy; A treatment decision optimization module for inputting the tumor regression grade probability distribution as a core component of a state space into a double-layer reinforcement learning framework based on a penalty mechanism to optimize treatment decisions and output a personalized therapeutic strategy parameter set including recommended chemotherapeutic drug dosages, treatment cycle lengths, and surgical timing suggestions; A safety verification and adjustment module for performing real-time scanning and verification on drug dosage parameters in the personalized therapeutic strategy parameter set and outputting a final approved safe-to-execute personalized treatment plan.

Images