A method and system for individual adjustment of tumor chemotherapy regimen and evaluation of therapeutic effect

Abstract

The application belongs to the technical field of medical information, and discloses a tumor chemotherapy scheme individualized adjustment and curative effect evaluation method and system. Static basic data, dynamic treatment data, single cell sequencing data and structured data of doctor-patient conversation are comprehensively collected in the data fusion quality control stage. Format adaptation and semantic alignment of cross-modal data are realized through a unified standardized interface. Data quality control is carried out in combination with an isolated forest algorithm, and only high-quality data with a score of 80 or more are included to provide a reliable basis for scheme development. In the multi-module collaborative prediction stage, a medical knowledge graph and a hybrid modeling framework are fused, multi-center data is jointly trained, and drug sensitivity scores are accurately output. When the scheme is generated, the clinical guidelines and patient needs are linked, multi-constraint rule reasoning and multi-objective optimization are used to develop individualized schemes that adapt to the patient's genetic characteristics, pathological types and quality of life demands, and reduce the risk of poor efficacy and serious adverse reactions caused by individual differences.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology, specifically a method and system for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation. Background Technology

[0002] The core goal of chemotherapy for cancer is to inhibit or kill tumor cells with chemical drugs, thereby prolonging patient survival and improving quality of life. However, traditional chemotherapy regimens often adopt a standardized, one-size-fits-all approach, failing to adequately consider individual patient differences, such as genetic characteristics, tumor pathology type, immune status, liver and kidney function, age, and comorbidities. This often leads to problems in clinical treatment, including poor efficacy, severe adverse reactions, and the development of drug resistance. While some existing technologies attempt to adjust the type of chemotherapy drug based on the patient's genetic testing results, the following technical challenges remain: First, the data is limited in scope, failing to integrate clinical diagnosis and treatment data, imaging data, multi-omics data (genomics, transcriptomics, proteomics, etc.), and real-time treatment response data, making it difficult to comprehensively characterize individual patient features. Second, efficacy assessment is delayed, relying heavily on traditional imaging examinations such as CT and MRI or tumor marker detection, which makes it difficult to reflect the effects of chemotherapy drugs on tumor cells in real time, resulting in untimely adjustments to treatment plans.

[0003] Existing technologies lack a closed-loop mechanism for the entire process of data, models, solutions, evaluation, and optimization. Models and knowledge systems cannot be dynamically iterated based on clinical data throughout the entire lifecycle, making it difficult to continuously improve the adaptability of individualized solutions and effectively link drug resistance early warning and solution adjustment, thus hindering the formation of precise control throughout the entire process. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, comprising the following specific steps: Preferably, the data fusion quality control stage integrates patient static basic data, dynamic treatment data, single-cell sequencing data, and structured doctor-patient dialogue data through the coordinated advancement of multi-dimensional panoramic data acquisition, cross-modal data standardized interface design, and full-process data quality control. The static basic data includes gene testing, tumor pathology, and basic clinical data; the dynamic treatment data includes chemotherapy drug dosage, medication time, adverse reactions, real-time hematological indicators, imaging monitoring data, and real-time monitoring data of chemotherapy drug blood concentration; the single-cell sequencing data obtains tumor heterogeneity-related gene expression information through single-cell RNA sequencing technology; and the structured doctor-patient dialogue data is generated through a large language model embedded with medical knowledge, including structured medical record entries such as chief complaint, treatment intention, and quality of life demands. A unified multi-source data standardization interface is designed to achieve format adaptation and semantic alignment of genomic data, imaging data, clinical text data, and single-cell sequencing data. Dedicated standardized processes are designed for different types of data: genomic data is screened for chemotherapy-sensitive or drug-resistant core sites through sequence alignment and variant annotation; clinical text data is extracted with structured information using improved natural language processing technology; imaging data is quantified with features through deep learning image segmentation algorithms; and single-cell data is extracted with tumor heterogeneity features through dimensionality reduction clustering. Sequence alignment was performed using the BWA-MEM algorithm to align the patient's gene sequence with the human reference genome GRCh38. Variance annotation was performed by integrating information from databases such as dbSNP, COSMIC, and ClinVar using the ANNOVAR tool to screen for nonsynonymous mutations, copy number variations, and fusion genes that are directly related to tumor chemotherapy sensitivity. After excluding benign polymorphic sites, core sites were identified.

[0006] The quantitative features include core radiomics features and dynamic change features. The radiomics features include the maximum diameter of the tumor, volume, mean / standard deviation of CT values, edge irregularity, texture features (gray-level co-occurrence matrix entropy, contrast), and enhancement peak time. The dynamic change features include the rate of change in tumor volume before and after chemotherapy, the magnitude of change in CT values, and the type of change in enhancement pattern. All features were included in the subsequent analysis after Z-score standardization.

[0007] This improved natural language processing technology is designed for the characteristics of doctor-patient dialogue texts. It integrates a medical terminology dictionary and a specialized vocabulary for tumor chemotherapy to enhance semantics, optimizes the named entity recognition module to accurately extract core structured items such as chief complaint, treatment intention, and quality of life demands, and corrects ambiguous expressions through contextual semantic association analysis to improve the completeness and accuracy of structured information extraction.

[0008] A data quality assessment and dynamic correction module is constructed. Outlier detection algorithms such as the isolated forest algorithm are used to identify invalid data. Missing values ​​are filled through cross-validation of multi-source data. A data analysis quality scoring system from 0 to 100 is established. Only standardized data with a score ≥ 80 are included in subsequent analysis to form a standardized data matrix with quality labels.

[0009] The scoring system comprises three dimensions: data completeness (40 points), accuracy (30 points), and consistency (30 points). Data completeness is scored based on the missing rate: 40 points for no missing data, 30 points for a missing rate ≤5%, 20 points for a missing rate between 5% and 10%, and 0 points for a missing rate >10%. Accuracy is scored through cross-validation of multi-source data: 30 points for consistent validation, 20 points for 1-2 minor inconsistencies, and 0 points for core data inconsistencies. Consistency is scored based on the semantic alignment results of cross-modal data: 30 points for complete alignment, 15 points for partial alignment, and 0 points for severe misalignment. The scores from these three dimensions are added together to obtain the final quality score.

[0010] Preferably, the multi-model collaborative prediction stage is based on the standardized data matrix output by the data fusion quality control stage, and constructs a multi-model collaborative prediction architecture that integrates medical knowledge graph. It adopts a hybrid architecture that combines medical knowledge graph with deep learning and reinforcement learning. The input layer is the core feature output by the data fusion quality control stage, and the hidden layer has a dual-branch structure. Branch one extracts the deep association between high-dimensional omics and clinical features through convolutional neural network, and branch two captures the temporal features of dynamic treatment data through long short-term memory network. The reinforcement learning module is used to dynamically adjust the feature weights of the dual branches to strengthen the contribution of features directly related to the efficacy of chemotherapy. The reinforcement learning module uses the accuracy of chemotherapy efficacy prediction as the core reward signal. It takes the correlation strength between high-dimensional omics features and clinical features, and the matching degree between the temporal change trend of dynamic treatment data and historical efficacy data as state inputs. Through iterative learning, it optimizes the feature weight allocation strategy, assigns higher weights to core features that are positively correlated with efficacy (such as key gene mutations and tumor heterogeneity features), and suppresses redundant features to ensure that the weight adjustment is directly linked to the efficacy prediction target.

[0011] Simultaneously construct a knowledge graph specifically for tumor chemotherapy, which includes knowledge such as drug-gene interactions, the correspondence between tumor types and chemotherapy regimens, adverse reactions and dosage adjustment rules, etc. Transform the semantic association information of the knowledge graph into constraints and embed them into the model training process. The knowledge graph data sources include DrugBank, PubMedCentral clinical literature, TCGA oncology database, NCCN / CSCO official guideline documents, and multicenter clinical case summary data; update triggers include the publication of new high-quality clinical studies, validation of more than 30 new consistent clinical cases, updates to drug instructions or revisions to guidelines, and the updated data must be reviewed and confirmed by more than 3 experts in the field of oncology chemotherapy.

[0012] A federated learning framework is used to jointly train multi-center clinical data. Model parameters are optimized through cross-validation and a dynamic model update mechanism is established. New case data and knowledge graph update information are regularly incorporated. The output layer is the sensitivity score of different chemotherapy drugs and the scoring criteria.

[0013] Federated learning uses the FedAvg algorithm to aggregate model parameters. After local training is completed, each center only uploads the model parameter gradients and does not transmit the original clinical data. Homomorphic encryption technology is used to ensure privacy and security during data transmission. Weight coefficients are set during parameter aggregation, and aggregation weights are dynamically allocated according to the data volume ratio of each center (60% weight ratio) and data quality score (40% weight ratio) to ensure balanced data contribution from multiple centers.

[0014] Preferably, the chemotherapy regimen generation and interpretation stage is based on the drug sensitivity score and scoring criteria output by the multi-modal collaborative prediction stage, combined with clinical guidelines and patient needs, to generate and interpret the chemotherapy regimen; a multi-constraint rule base is constructed by integrating NCCN guidelines, CSCO guidelines, medical knowledge graph derived rules, and patient quality of life demands and preference constraint rules. The rule base includes standard drug combinations and dosage ranges for different tumor types or stages, as well as adjustment rules corresponding to liver and kidney dysfunction, comorbidities, and patient quality of life demands. The rule base is prioritized according to "rigid constraints > flexible constraints". Rigid constraints include contraindications explicitly stated in the NCCN / CSCO guidelines and dosage restrictions for patients with severe hepatic or renal insufficiency. These have the highest priority and cannot be violated. Flexible constraints include rules derived from medical knowledge graphs and rules based on patients' quality of life demands and preferences. These have the next highest priority. When flexible constraints conflict with rigid constraints, the rigid constraints take precedence. When there is a conflict between flexible constraints, the decision is made by weighting clinical expert consensus (70%) and patient satisfaction (30%).

[0015] A strategy combining rule-based reasoning, weighted scoring, and multi-objective optimization is used to generate chemotherapy regimens. First, candidate drug combinations are screened based on tumor type and stage. The top three groups are selected based on the drug sensitivity scores obtained in the aforementioned prediction stage. Then, the drug dosage and treatment cycle are individually adjusted based on the patient's basic clinical data and quality of life requirements. Finally, three candidate regimens are screened through a multi-objective optimization algorithm, and the optimal regimen is determined. The multi-objective optimization algorithm adopts the non-dominated sorting genetic algorithm (NSGA-II), with "maximizing efficacy, minimizing adverse reaction risk, and minimizing patient tolerance cost" as the three core objectives. The priority weight of each objective is set in combination with the consensus of clinical experts, among which the efficacy objective has the highest weight. The weights of adverse reaction risk and patient tolerance cost are dynamically adapted according to the patient's age, underlying disease status and quality of life requirements. The three Pareto optimal set of schemes are selected through population iteration to ensure that the scheme takes into account both effectiveness and safety.

[0016] For each candidate regimen, generate an explanatory report that includes the basis for drug selection, reasons for dosage adjustment, expected efficacy, potential adverse reaction risk score, and avoidance suggestions.

[0017] Preferably, the real-time efficacy evaluation stage constructs a multi-dimensional real-time efficacy evaluation system that integrates data quality feedback for the chemotherapy regimen determined in the regimen generation and interpretation stage; according to the set time intervals of 1 day before chemotherapy, 3 days after chemotherapy, and 7 days after chemotherapy, multi-dimensional dynamic data such as hematological indicators, imaging indicators, patient subjective symptoms, drug blood concentration indicators, and data quality scores are collected to construct a data quality and efficacy evaluation correction mechanism, and the weight of each monitoring indicator is dynamically adjusted according to the data quality score; Hematological indicators include tumor markers (carcinoembryonic antigen CEA, squamous cell carcinoma antigen SCC, carbohydrate antigen CA19-9, etc.), complete blood count (white blood cell count, platelet count), liver and kidney function (alanine aminotransferase ALT, creatinine Cr), and inflammatory factors (C-reactive protein CRP); imaging indicators include maximum tumor diameter, tumor volume, changes in the number of lesions, enhancement degree, and whether new metastases have appeared. Tumor volume is calculated based on imaging data using a three-dimensional reconstruction algorithm.

[0018] The dynamic weight adjustment follows the principle of "the higher the quality, the higher the weight": when the data quality score is ≥90, the weight of the indicator is 1.1 times the basic weight; when the score is 80-89, the basic weight remains unchanged; when the score is 70-79, it is reduced to 0.8 times the basic weight; when the score is 60-69, it is reduced to 0.5 times the basic weight; when the score is below 60, the indicator does not participate in the calculation of the comprehensive efficacy score, and its weight is allocated to other high-quality indicators proportionally.

[0019] The basic weights of each indicator were determined using the analytic hierarchy process (AHP), and the comprehensive efficacy score was calculated by combining the dynamic weights after data quality correction, and the efficacy levels were classified. By monitoring the changing trends of various indicators through trend analysis algorithms and combining the correlation characteristics of efficacy decline in medical knowledge graphs, we can provide early warnings of the risk of efficacy decline and output early warning reports and preliminary intervention suggestions.

[0020] The sliding window size is set to three consecutive monitoring time points (i.e., 1 day before chemotherapy, 3 days after chemotherapy, and 7 days after chemotherapy as the first complete window). The slope of the change of each indicator within the window is fitted by linear regression. When the slope meets the preset "effect decline association characteristic threshold" in the medical knowledge graph (such as tumor marker concentration slope > 0.2, blood drug concentration slope < -0.15), an effect decline warning is triggered. At the same time, the sliding window is continuously updated to monitor the trend of indicator changes, ensuring the timeliness and accuracy of the warning.

[0021] Preferably, the precise adjustment phase of the treatment plan is based on the efficacy level, early warning information, and multi-center clinical collaborative data output from the real-time efficacy assessment phase, and constructs an intelligent decision tree to adjust the chemotherapy regimen. The intelligent decision tree is constructed using the C4.5 algorithm, with a training set of full-cycle data from more than 10,000 multicenter tumor chemotherapy cases. The data covers different tumor types, stages, gene characteristics, and treatment responses. During training, redundant branches are removed using a pruning algorithm. The efficacy improvement rate and adverse reaction reduction rate after treatment adjustment are used as evaluation indicators to ensure the generalization ability and clinical adaptability of the decision tree.

[0022] The decision tree nodes contain key parameters such as data quality labels, drug resistance gene mutation types, patient tolerance scores, efficacy levels, and adverse reaction severity. The branches correspond to refined adjustment strategies. For example, if the efficacy is significant and the adverse reactions are mild, the current regimen is maintained and the monitoring frequency is optimized. If the efficacy is acceptable but the adverse reactions are severe, the drug dosage is reduced by 20% to 50% or a low-toxicity alternative drug is replaced based on blood drug concentration and liver and kidney function indicators. If the efficacy is poor, the drug combination is changed or the medication cycle is adjusted based on the drug resistance test results and the updated drug sensitivity prediction model. The proposed adjustment plan is uploaded to the federated learning platform. The feasibility and safety of the plan are verified using multi-center historical data, and a verification report is generated. The Monte Carlo simulation method is used to predict the expected efficacy and adverse reaction risk of the new plan, and an adjustment report is generated that includes the reasons for the adjustment, details of the new plan, expected effects, and key monitoring points. The adjusted plan and monitoring data are fed back to the data fusion quality control stage and the multi-mode collaborative prediction stage to update the data collection system and prediction model parameters, while providing basic data for the drug resistance early warning and intervention stage.

[0023] Preferably, the drug resistance early warning and intervention stage combines the efficacy change data from the real-time efficacy evaluation stage with the drug resistance detection information from the precise adjustment stage to construct a drug resistance early warning and intervention system that integrates single-cell data; it integrates dynamic gene data, efficacy monitoring data, tumor heterogeneity drug resistance characteristics obtained from single-cell sequencing, and drug metabolism enzyme activity data to form a multi-dimensional drug resistance feature set, and constructs a deep learning-based drug resistance early warning model. The model input is the drug resistance feature set, and the output is a drug resistance risk score and early warning level, and a report on the cause analysis of the early warning is generated by combining a medical knowledge graph. Drug-metabolizing enzyme activities were detected by liquid chromatography-tandem mass spectrometry (LC-MS / MS), with a focus on detecting the activities of the cytochrome P450 family (CYP3A4, CYP2D6, CYP2C9) and uridine diphosphate glucuronide transferase (UGT1A1). The activity level was quantified by the enzyme reaction rate. The test samples were peripheral blood lymphocytes or serum collected during chemotherapy.

[0024] A tiered intervention strategy is implemented for different warning levels. At low risk levels, the monitoring interval is shortened to 3 days and blood drug concentration monitoring is strengthened. At medium risk levels, the dosage of chemotherapy drugs is adjusted or targeted drugs to reverse drug resistance are used in combination and dynamic monitoring of drug resistance-related genes is supplemented. At high risk levels, the protocol adjustment process is initiated immediately, a new combination of chemotherapy drugs is used, and a comprehensive protocol is developed in combination with immunotherapy. Record data related to drug resistance occurrence, update the drug resistance feature database and knowledge graph, and feed the updated database back to the data fusion quality control stage.

[0025] Preferably, the model knowledge optimization stage performs iterative optimization of the model and knowledge system based on the full-cycle data generated in the aforementioned stages; after the chemotherapy cycle ends, the patient's final efficacy data, adverse reaction occurrence, regimen adjustment records, and patient satisfaction scores are collected to construct an efficacy evaluation dataset. A comparative analysis method is used to compare the prediction results of the multi-model collaborative prediction stage with the actual efficacy, calculate the prediction error and identify model deficiencies, adjust the feature weights and network structure of the prediction model based on the error analysis results, and update the medical knowledge graph to supplement newly discovered knowledge such as the association between drugs and drug resistance genes and factors affecting efficacy. An optimization report is generated, which includes a performance comparison before and after model optimization, updated knowledge graph content, and optimization suggestions for the data collection system. The optimized model and knowledge graph are applied to the data analysis process of new cases, and the reliability is further improved through multi-center clinical validation. Ultimately, this forms a complete closed loop encompassing data collection, quality control, model prediction, treatment plan generation, efficacy monitoring, treatment plan adjustment, and iterative optimization.

[0026] This invention also provides a system for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, based on the above method, comprising: Data acquisition and preprocessing module: integrates multi-source data acquisition units, is equipped with standardized processing units to achieve cross-modal data format adaptation and semantic alignment, and is equipped with a quality control unit to filter outliers, fill missing values ​​and filter by quality scores, and outputs a standardized data matrix with quality level labels; Multi-modal collaborative prediction module: Constructs a hybrid modeling unit that integrates medical knowledge graph, integrates a dual-branch network of convolutional neural network and long short-term memory network and reinforcement learning unit, realizes joint training of multi-center data through federated learning unit, and outputs drug sensitivity score and correlation evidence; The protocol generation and interpretation module has a built-in multi-constraint rule library. It generates three candidate chemotherapy protocols through rule reasoning and multi-objective optimization units, and simultaneously generates an interpretation report containing the selection basis, expected efficacy and risk avoidance suggestions. Real-time efficacy assessment module: Collects dynamic diagnosis and treatment data at preset intervals, outputs efficacy level through dynamic weight adjustment and efficacy scoring unit, and provides early warning 3 to 5 days in advance of efficacy decline by trend analysis unit, and outputs early warning report; The protocol adjustment and drug resistance early warning module: Based on the efficacy evaluation results, an intelligent decision tree unit is constructed to complete the fine-tuning of the protocol and validated by multi-center data; a multi-dimensional drug resistance feature is integrated to construct an early warning unit, outputting the drug resistance risk level and graded intervention strategy, and updating the drug resistance feature database; Model knowledge optimization module: Collect full-cycle diagnosis and treatment data to construct efficacy evaluation dataset, and achieve iterative optimization through model parameter adjustment and knowledge graph update unit. After multi-center clinical validation, the results are fed back to the aforementioned modules.

[0027] The beneficial effects of this invention are as follows: 1. This invention comprehensively collects static basic data, dynamic treatment data, single-cell sequencing data, and structured data from doctor-patient dialogues during the data fusion and quality control stage. A unified standardized interface is used to achieve format adaptation and semantic alignment of cross-modal data. Data quality control is implemented using algorithms such as the Isolation Forest algorithm, incorporating only high-quality data with scores ≥80 points to provide a reliable foundation for treatment plan development. In the multi-modal collaborative prediction stage, medical knowledge graphs and hybrid modeling architectures are integrated, utilizing multi-center data for joint training to accurately output drug sensitivity scores. During treatment plan generation, clinical guidelines and patient needs are considered, and through multi-constraint rule reasoning and multi-objective optimization, individualized treatment plans are developed to suit the patient's genetic characteristics, pathological type, and quality-of-life requirements, reducing the risk of poor efficacy and serious adverse reactions due to individual differences.

[0028] 2. This invention constructs a multi-dimensional real-time efficacy evaluation system, collecting dynamic data from multiple sources at preset intervals of 1 day before chemotherapy, 3 days after chemotherapy, and 7 days after chemotherapy. The system dynamically adjusts indicator weights based on data quality to accurately classify efficacy levels. Through trend analysis algorithms linked with a medical knowledge graph, it provides early warning of the risk of declining efficacy 3-5 days in advance, allowing time for regimen adjustments. During the regimen adjustment phase, relying on an intelligent decision tree, it formulates refined strategies based on key parameters such as efficacy level and adverse reaction severity. For example, if adverse reactions are severe, the drug dosage may be reduced by 20%-50% or a less toxic alternative drug may be used; if efficacy is poor, the drug combination may be changed. Multi-center data validation and effect prediction ensure the feasibility and safety of the adjusted regimen.

[0029] 3. After the chemotherapy cycle of this invention is completed, an evaluation dataset is constructed by integrating the patient's final efficacy data, adverse reaction information, and protocol adjustment records. The predicted results are compared with the actual efficacy, the model parameters are adjusted, the medical knowledge graph is updated, and newly discovered drug-gene associations and efficacy influencing factors are added. The optimized model and knowledge system are fed back to each pre-processing stage, and multi-center data joint training and dynamic model updates are achieved by relying on the federated learning framework, continuously improving the prediction accuracy and adaptability of subsequent case protocols. Through continuous iterative optimization and clinical validation, the technology of personalized chemotherapy for tumors is steadily upgraded. Attached Figure Description

[0030] Fig. 1 This is an overall flowchart of the method of the present invention; Fig. 2 This is a flowchart of the data fusion quality control stage of the present invention; Fig. 3 This is a flowchart illustrating the generation and adjustment process of the solution in this invention. Detailed Implementation

[0031] 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.

[0032] like Figs. 1 to 3 As shown, this embodiment of the invention provides a method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, including the following specific steps: The data fusion quality control stage integrates patient static basic data, dynamic treatment data, single-cell sequencing data, and structured doctor-patient dialogue data through the coordinated advancement of multi-dimensional panoramic data acquisition, cross-modal data standardization interface design, and full-process data quality control. Static basic data includes gene testing, tumor pathology, and basic clinical data; dynamic treatment data includes chemotherapy drug dosage, medication time, adverse reactions, real-time hematological indicators, imaging monitoring data, and real-time monitoring data of chemotherapy drug blood concentration; single-cell sequencing data obtains tumor heterogeneity-related gene expression information through single-cell RNA sequencing technology; and structured doctor-patient dialogue data is generated through a large language model embedded with medical knowledge, including structured medical record items such as chief complaint, treatment intention, and quality of life demands. A unified multi-source data standardization interface is designed to achieve format adaptation and semantic alignment of genomic data, imaging data, clinical text data, and single-cell sequencing data. Dedicated standardized processes are designed for different types of data: genomic data is screened for chemotherapy-sensitive or drug-resistant core sites through sequence alignment and variant annotation; clinical text data is extracted with structured information using improved natural language processing technology; imaging data is quantified with features through deep learning image segmentation algorithms; and single-cell data is extracted with tumor heterogeneity features through dimensionality reduction clustering. The UMAP algorithm was used to reduce the dimensionality of high-dimensional single-cell data, preserving the local and global structural features of the data. The K-means algorithm was used for clustering, with the K value determined based on prior knowledge of tumor subtypes and the elbow rule (usually set to 3-8). Clustering results with a silhouette coefficient ≥0.7 were identified as effective cell subpopulations, and differentially expressed genes of each subpopulation were extracted as core features of tumor heterogeneity.

[0033] A data quality assessment and dynamic correction module is constructed. Outlier detection algorithms such as the isolated forest algorithm are used to identify invalid data. Missing values ​​are filled through cross-validation of multi-source data. A data analysis quality scoring system from 0 to 100 is established. Only standardized data with a score ≥ 80 are included in subsequent analysis to form a standardized data matrix with quality labels.

[0034] The multi-model collaborative prediction stage is based on the standardized data matrix output from the data fusion quality control stage. It constructs a multi-model collaborative prediction architecture that integrates medical knowledge graphs. It adopts a hybrid architecture that combines medical knowledge graphs with deep learning and reinforcement learning. The input layer is the core features output from the data fusion quality control stage, including key gene mutations, tumor heterogeneity features, liver and kidney function indicators, etc. The hidden layer has a dual-branch structure. Branch 1 extracts the deep association between high-dimensional omics and clinical features through convolutional neural networks (CNN). Branch 2 captures the temporal features of dynamic treatment data through long short-term memory networks (LSTM). The reinforcement learning module is used to dynamically adjust the feature weights of the dual branches to strengthen the contribution of features directly related to the efficacy of chemotherapy. Simultaneously construct a knowledge graph specifically for tumor chemotherapy, which includes knowledge such as drug-gene interactions, the correspondence between tumor types and chemotherapy regimens, adverse reactions and dosage adjustment rules, etc. Transform the semantic association information of the knowledge graph into constraints and embed them into the model training process. A federated learning framework is used to jointly train multi-center clinical data to avoid data privacy leaks. Model parameters are optimized through cross-validation and a dynamic model update mechanism is established. New case data and knowledge graph update information are regularly incorporated to improve the model’s adaptability to different tumor subtypes. The output layer is the sensitivity score (0 to 10 points) of different chemotherapy drugs and the scoring basis (related knowledge graph nodes).

[0035] Cross-validation adopts a 5-fold cross-validation method, which divides the multi-center joint training dataset into training and validation sets in an 8:2 ratio. During each fold of training, the model architecture is fixed, and only the feature weights and regularization parameters are optimized. The model dynamic update cycle is set to every 50 new complete diagnosis and treatment data or every 3 months. After the update, the performance is verified through a 20% independent test set to ensure the stability of the model.

[0036] The chemotherapy regimen generation and interpretation stage is based on the drug sensitivity score and scoring criteria output by the multi-modal collaborative prediction stage. It combines clinical guidelines and patient needs to generate and interpret the chemotherapy regimen. The system integrates NCCN guidelines, CSCO guidelines, medical knowledge graph derived rules, and patient quality of life demands and preference constraint rules to construct a multi-constraint rule base. The rule base includes standard drug combinations and dosage ranges for different tumor types or stages, as well as adjustment rules corresponding to liver and kidney dysfunction, comorbidities, and patient quality of life demands. A strategy combining rule-based reasoning, weighted scoring, and multi-objective optimization is employed to generate chemotherapy regimens. First, candidate drug combinations are screened based on tumor type and stage. The top three groups are then selected based on the drug sensitivity scores obtained in the aforementioned prediction stage. Next, drug dosage and treatment cycle are individually adjusted based on the patient's basic clinical data and quality of life requirements. Finally, a multi-objective optimization algorithm is used to screen three candidate regimens and determine the optimal regimen. The objective function is calculated as follows: maximize efficacy minus λ multiplied by minimize adverse reactions minus μ multiplied by the patient's tolerance cost.

[0037] The weighted scoring is distributed as follows: drug sensitivity score accounts for 60%, clinical guideline matching degree accounts for 25%, and patient quality of life requirement fit accounts for 15%. Among them, the drug sensitivity score is linearly mapped to 0-10 points as the weight percentage, the clinical guideline matching degree is scored according to the degree of fit between the protocol and the NCCN / CSCO guideline recommended protocol, and the patient quality of life requirement fit is quantified by the percentage of requirement satisfaction.

[0038] For each candidate regimen, generate an explanatory report that includes the basis for drug selection (sensitivity score, knowledge graph association evidence), reasons for dosage adjustment, expected efficacy (tumor control rate, progression-free survival prediction), potential adverse reaction risk score, and avoidance suggestions.

[0039] The real-time efficacy assessment stage, for the chemotherapy regimen determined in the regimen generation and interpretation stage, constructs a multi-dimensional real-time efficacy assessment system that integrates data quality feedback to achieve dynamic monitoring and early warning of efficacy during chemotherapy. At set time intervals of 1 day before chemotherapy, 3 days after chemotherapy, and 7 days after chemotherapy, multi-dimensional dynamic data such as hematological indicators, imaging indicators, patient subjective symptoms (collected using standardized scales), drug blood concentration indicators, and data quality scores are collected. A data quality and efficacy assessment correction mechanism is constructed, and the weight of each monitoring indicator is dynamically adjusted according to the data quality score. The higher the quality score, the greater the weight of the corresponding indicator. The standardized scale uses a combination of the "EORTCQLQ-C30 Cancer Patient Quality of Life Core Scale + QLQ-CHEMO24 Chemotherapy-Specific Symptom Scale". It focuses on collecting subjective symptoms directly related to the efficacy and tolerability of chemotherapy, such as fatigue, nausea and vomiting, pain, loss of appetite, constipation, and diarrhea. Each symptom is scored from 0 to 100, and the average score of each item is used as the comprehensive subjective symptom score.

[0040] The basic weights of each indicator were determined using the analytic hierarchy process (AHP): tumor volume change rate 0.35, tumor marker concentration change 0.3, improvement in patient subjective symptoms 0.2, adverse reaction severity 0.1, and blood drug concentration target achievement rate 0.05. A comprehensive efficacy score was calculated using dynamic weights after data quality correction, and efficacy levels were assigned: 80-100 points for "significant efficacy," 60-79 points for "decent efficacy," and below 60 points for "poor efficacy." By monitoring the changing trends of various indicators through trend analysis algorithms (such as linear regression combined with sliding window), and combining the efficacy decline correlation characteristics in the medical knowledge graph, such as the continuous increase of tumor markers and abnormal fluctuations in blood drug concentration, the risk of efficacy decline can be warned 3 to 5 days in advance, and warning reports and preliminary intervention suggestions can be output.

[0041] The precise adjustment phase of the treatment plan is based on the efficacy level, early warning information, and multi-center clinical collaborative data output from the real-time efficacy assessment phase. An intelligent decision tree is constructed to adjust the chemotherapy regimen. The decision tree nodes include key parameters such as data quality labels, drug resistance gene mutation types, patient tolerance scores, efficacy levels, and adverse reaction severity. The branches correspond to refined adjustment strategies. That is, if the efficacy is significant and the adverse reactions are mild (level 1 to 2), the current regimen is maintained and the monitoring frequency is optimized. If the efficacy is acceptable but the adverse reactions are severe (level 3 to 4), the drug dosage is reduced by 20% to 50% or a low-toxicity alternative drug is replaced based on blood drug concentration and liver and kidney function indicators. If the efficacy is poor, the drug combination is changed or the medication cycle is adjusted by combining the drug resistance test results and the updated drug sensitivity prediction model. The patient tolerance score is based on a 100-point scale, with the severity of adverse reactions accounting for 40 points (30-40 points for grade 1-2 adverse reactions, and 0-29 points for grade 3-4 adverse reactions), the degree of improvement in subjective symptoms accounting for 30 points (20-30 points for symptom relief, 10-19 points for no change, and 0-9 points for worsening), and liver and kidney function indicators accounting for 30 points (25-30 points for normal indicators, 15-24 points for mild abnormalities, and 0-14 points for moderate or above abnormalities). The sum of the scores of each item is the final tolerance score, with 80 points or above indicating "good tolerance", 60-79 points indicating "moderate tolerance", and below 60 points indicating "poor tolerance".

[0042] The proposed adjustment plan is uploaded to the federated learning platform. The feasibility and safety of the plan are verified using multi-center historical data, and a verification report is generated. The Monte Carlo simulation method is used to predict the expected efficacy and adverse reaction risk of the new plan, and an adjustment report is generated that includes the reasons for the adjustment, details of the new plan, expected effects, and key monitoring points. The Monte Carlo simulation iterations were set to 1000. The simulation variables included drug dose fluctuations (±20%), liver and kidney function index variations (referenced to the clinical normal range ±15%), tumor proliferation rate changes (based on the 0.5-1.5 times range of multicenter data), and drug-metabolizing enzyme activity fluctuations (±25%). The probability distribution of each variable was determined by fitting data from more than 5000 multicenter historical cases. The mean of the simulation results at the 95% confidence interval was used as the expected efficacy and risk assessment value.

[0043] The adjusted plan and subsequent monitoring data are fed back to the data fusion quality control stage and the multi-model collaborative prediction stage to update the data collection system and prediction model parameters, strengthen the linkage between data, models and plans, and provide basic data for the adjusted plan in the drug resistance early warning and intervention stage.

[0044] The drug resistance early warning and intervention stage combines efficacy change data from the real-time efficacy assessment stage with drug resistance detection information from the precise adjustment stage to construct a drug resistance early warning and intervention system that integrates single-cell data. It integrates dynamic gene data, efficacy monitoring data, tumor heterogeneity drug resistance characteristics obtained from single-cell sequencing, and drug metabolism enzyme activity data to form a multi-dimensional drug resistance feature set. A deep learning-based drug resistance early warning model is then constructed. The model input is the drug resistance feature set, including new drug resistance gene mutations, mutation frequency changes, tumor volume growth rate, and blood drug concentration changes. The output is a drug resistance risk score (0 to 10 points) and an early warning level: low risk is 0 to 3 points, medium risk is 4 to 7 points, and high risk is 8 to 10 points. A report analyzing the causes of the early warning is also generated using a medical knowledge graph. The drug resistance early warning model adopts a fusion architecture of "CNN-LSTM-attention mechanism". The CNN layer is used to extract spatial correlation features in the multi-dimensional drug resistance feature set (such as the correlation between gene mutation and drug-metabolizing enzyme activity). The LSTM layer is used to capture the temporal change pattern of dynamic features (such as the time series changes of mutation frequency and tumor volume). The attention mechanism layer focuses on strengthening the contribution of core features directly related to the occurrence of drug resistance (such as the frequency of drug resistance gene mutation and fluctuations in blood drug concentration). The model outputs a drug resistance risk score and the corresponding early warning level judgment criteria.

[0045] A tiered intervention strategy is implemented for different warning levels. At low risk levels, the monitoring interval is shortened to 3 days and blood drug concentration monitoring is strengthened. At medium risk levels, the dosage of chemotherapy drugs is adjusted or targeted drugs to reverse drug resistance (such as cetuximab) are used in combination and dynamic monitoring of drug resistance-related genes is supplemented. At high risk levels, the protocol adjustment process is initiated immediately, a new combination of chemotherapy drugs is used, and a comprehensive protocol is developed in combination with immunotherapy. Record data related to drug resistance occurrence, update the drug resistance feature database and knowledge graph, and feed the updated database back to the data fusion quality control stage.

[0046] The conditions for updating the drug resistance feature database are: ≥20 newly added drug-resistant cases with common drug resistance features, newly discovered drug resistance gene mutation sites confirmed by PCR, and statistical verification of the association between drug-metabolizing enzyme activity and drug resistance (P<0.05). After the update, it must be subject to a dual mechanism of "algorithm model verification + clinical expert review". The validity of the newly added features must be confirmed by two oncology experts and one bioinformatics expert to ensure the clinical suitability of the updated feature database.

[0047] The model knowledge optimization stage is based on the full-cycle data generated in the previous stage to perform iterative optimization of the model and knowledge system. After the chemotherapy cycle ends, the final efficacy data (progression-free survival, overall survival, etc.), adverse reaction occurrence, regimen adjustment records, and patient satisfaction scores are collected to construct an efficacy evaluation dataset. The prediction results of the multi-model collaborative prediction stage are compared with the actual efficacy using a comparative analysis method. The prediction error is calculated and the model deficiencies are identified. Based on the error analysis results, the feature weights and network structure of the prediction model are adjusted. At the same time, the medical knowledge graph is updated to supplement newly discovered knowledge such as the association between drugs and drug resistance genes and factors affecting efficacy. Patient satisfaction was assessed using the "Healthcare Service Satisfaction Scale (HCSS) Tumor Chemotherapy Specific Version," which includes four dimensions: regimen suitability, adverse reaction management, communication convenience, and perceived efficacy, with a total of 20 items. Each item is scored from 1 to 5, and the total score is converted to 0-100 points. ≥80 points indicates "very satisfied," 60-79 points indicates "basically satisfied," and <60 points indicates "unsatisfied."

[0048] The comparative analysis method uses "the matching degree between the drug sensitivity prediction score and the actual efficacy level, and the deviation rate between the expected progression-free survival and the actual progression-free survival" as the core indicators. When the matching degree is lower than 80% or the deviation rate is higher than 20%, it is determined that the model has room for optimization. Parameter adjustment prioritizes correcting the weight allocation of high-dimensional omics features and clinical features, increases the weight of tumor subtype-related features with large prediction deviations, and updates the corresponding drug-gene association rules in the medical knowledge graph to ensure that the model optimization is targeted.

[0049] An optimization report is generated, which includes a performance comparison before and after model optimization, updated knowledge graph content, and optimization suggestions for the data collection system. The optimized model and knowledge graph are applied to the data analysis process of new cases, and the reliability is further improved through multi-center clinical validation. Ultimately, this forms a complete closed loop encompassing data collection, quality control, model prediction, treatment plan generation, efficacy monitoring, treatment plan adjustment, and iterative optimization.

[0050] Multicenter clinical validation must cover at least 10 tertiary oncology hospitals in different regions, with each center providing data on no fewer than 100 new cases. The cases must cover common tumor types (lung cancer, gastric cancer, colorectal cancer, etc.) and different stages and genetic characteristics. After successful validation, the system must meet the performance indicators of prediction accuracy ≥85% and adverse reaction risk prediction deviation ≤10% before it can be officially applied.

[0051] This invention also provides a system for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, based on the above method, including: Data acquisition and preprocessing module: integrates multi-source data acquisition units (gene detection, imaging, clinical diagnosis and treatment, single-cell sequencing, doctor-patient dialogue acquisition equipment), equipped with standardized processing units to achieve cross-modal data format adaptation and semantic alignment, and equipped with quality control units to filter through outlier detection, missing value imputation and quality scoring, and outputs a standardized data matrix with quality level labels; Multi-modal collaborative prediction module: Constructs a hybrid modeling unit that integrates medical knowledge graph, integrates a dual-branch network of convolutional neural network and long short-term memory network and reinforcement learning unit, realizes joint training of multi-center data through federated learning unit, and outputs drug sensitivity score and correlation evidence; The protocol generation and interpretation module has a built-in multi-constraint rule base, which includes clinical guidelines, knowledge graph rules, and patient preferences. It generates three candidate chemotherapy protocols through rule reasoning and multi-objective optimization units, and simultaneously generates an interpretation report containing the selection basis, expected efficacy, and risk avoidance suggestions. Real-time efficacy assessment module: Collects dynamic diagnosis and treatment data at preset intervals, outputs efficacy level through dynamic weight adjustment and efficacy scoring unit, and provides early warning 3 to 5 days in advance of efficacy decline by trend analysis unit, and outputs early warning report; The protocol adjustment and drug resistance early warning module: Based on the efficacy evaluation results, an intelligent decision tree unit is constructed to complete the fine-tuning of the protocol and validated by multi-center data; a multi-dimensional drug resistance feature is integrated to construct an early warning unit, outputting the drug resistance risk level and graded intervention strategy, and updating the drug resistance feature database; Model knowledge optimization module: Collect full-cycle diagnosis and treatment data to construct efficacy evaluation dataset, and achieve iterative optimization through model parameter adjustment and knowledge graph update unit. After multi-center clinical validation, the results are fed back to the aforementioned modules.

[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, 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 process, method, article, or apparatus.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, characterized in that, The specific steps include the following: Data fusion and quality control stage: Through the collaborative operation of multi-dimensional full-cycle patient data panoramic collection, cross-modal data standardization processing and full-process data quality control, the static basic data and dynamic diagnosis and treatment data of patients are integrated, and after format adaptation, semantic alignment and quality screening and correction, a standardized data matrix with quality level labels is formed. Multi-model collaborative prediction stage: Based on the standardized data matrix, a multi-model collaborative prediction architecture integrating medical professional knowledge graph is constructed. A hybrid modeling approach combining knowledge graph with deep learning and reinforcement learning is adopted. After joint training and optimization with multi-center clinical data, drug sensitivity scores and correlation evidence are output. The protocol generation and interpretation phase: Based on the drug sensitivity prediction results, a multi-dimensional constraint rule base is constructed by integrating clinical treatment guidelines and patient needs. Through rule reasoning and multi-objective optimization, candidate chemotherapy protocols suitable for patients are generated, and an interpretation report containing selection criteria, expected efficacy and risk avoidance suggestions is generated simultaneously. Real-time efficacy assessment phase: A multi-dimensional real-time efficacy assessment system is constructed for the determined chemotherapy regimen. Dynamic diagnosis and treatment data are collected at preset intervals, and efficacy scores are calculated and classified after dynamic weighting. The risk of efficacy decline is warned 3 to 5 days in advance. Precise adjustment phase of the treatment plan: Based on the efficacy assessment results and multi-center data, an intelligent decision tree is constructed to design adjustment strategies. After multi-center data verification and effect prediction, an adjustment report is output. Drug resistance early warning and intervention phase: Integrate multi-dimensional drug resistance-related features to construct an early warning model, output drug resistance risk scores and levels, implement graded intervention strategies, record relevant data, and update the knowledge base; Model knowledge optimization stage: Based on the full-cycle diagnosis and treatment data, the model and knowledge system are iteratively optimized. Parameters are adjusted and the knowledge graph is updated through efficacy comparison analysis. A complete closed loop is formed through clinical validation.

2. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 1, characterized in that, In the data fusion quality control phase, the integrated multi-dimensional data includes patient static basic data, dynamic treatment data, single-cell sequencing data, and structured data of doctor-patient dialogue; The static basic data includes gene testing, tumor pathology and basic clinical data, while the dynamic treatment data includes chemotherapy drug information, adverse reactions, hematological indicators, imaging monitoring data and drug blood concentration monitoring data. A unified multi-source data standardization interface is used to achieve format adaptation and semantic alignment of cross-modal data. Dedicated standardization processes are designed for different types of data to extract core features. A data quality assessment and dynamic correction module is constructed, and outlier detection and multi-source data cross-validation are used to fill missing values. A quality scoring system from 0 to 100 is established, and only standardized data with a score ≥ 80 is included in subsequent analysis.

3. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 2, characterized in that, In the multi-model collaborative prediction stage, the input layer of the multi-model collaborative prediction architecture is the core feature output by the data fusion quality control stage. The core feature includes high-dimensional omics features, clinical features and dynamic treatment-related features. The hidden layer is set with a dual-branch structure to extract the deep correlation between high-dimensional omics and clinical features and the temporal features of dynamic treatment data respectively. The reinforcement learning module dynamically adjusts the weights of the dual-branch features. The medical knowledge graph contains core knowledge, including drug-gene interactions and the correspondence between tumor types and chemotherapy regimens. Its semantic association information is transformed into constraint embedding model training. A federated learning framework is used to achieve joint training of multi-center clinical data, and the model parameters are optimized through cross-validation and a dynamic update mechanism is established.

4. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 3, characterized in that, During the scheme generation and interpretation phase, the multi-dimensional constraint rule base integrates NCCN guidelines, CSCO guidelines, rules derived from medical knowledge graphs, and patient quality of life demands and preference constraint rules, including standard drug combinations, dosage ranges, and adjustment rules for special circumstances for different tumor types or stages. The strategy for generating the treatment plan combines rule-based reasoning, weighted scoring, and multi-objective optimization. First, candidate drug combinations are screened and ranked based on drug sensitivity scores. Then, individual adjustments are made based on patient data. Finally, three candidate treatment plans are selected and the optimal plan is determined. The explanatory report explains the basis for drug selection, the reasons for dosage adjustment, the expected efficacy, and the potential adverse reaction risks and avoidance suggestions for each candidate treatment plan.

5. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 4, characterized in that, In the real-time efficacy assessment phase, the preset collection intervals include 1 day before chemotherapy, 3 days after chemotherapy, and 7 days after chemotherapy. The collected data includes hematological indicators, imaging indicators, patient subjective symptoms, drug blood concentration indicators, and data quality scores. The basic weights of each indicator were determined by the analytic hierarchy process, and the weights were dynamically adjusted in conjunction with the data quality score to calculate the comprehensive efficacy score and grade. By monitoring the changing trends of indicators through trend analysis and combining the correlation characteristics of declining efficacy in the medical knowledge graph, early warning reports and preliminary intervention suggestions are generated.

6. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 5, characterized in that, In the precise adjustment phase of the proposed solution, the nodes of the intelligent decision tree include key parameters such as data quality labels, drug resistance gene mutation types, patient tolerance scores, efficacy levels, and adverse reaction severity, with branches corresponding to refined adjustment strategies; the patient tolerance scores are calculated based on the severity of adverse reactions, subjective symptoms, and liver and kidney function indicators collected during the real-time efficacy assessment phase. The proposed adjustment plan was validated using historical data from multiple centers on the federated learning platform. Simulation methods were used to predict the expected efficacy and adverse reaction risks, and an adjustment report was generated. The adjusted plan and monitoring data were fed back to the data fusion quality control stage and the multi-model collaborative prediction stage to update the data collection system and prediction model parameters.

7. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 6, characterized in that, In the drug resistance early warning and intervention phase, a multi-dimensional drug resistance feature set integrates dynamic gene data, efficacy monitoring data, tumor heterogeneity drug resistance characteristics obtained from single-cell sequencing, and drug metabolism enzyme activity data. The drug resistance early warning model outputs a drug resistance risk score, early warning level, and early warning cause analysis report; the graded intervention strategy implements intervention measures such as shortening the monitoring interval, adjusting drug dosage or combination, and changing the initiation plan according to the risk level; and records drug resistance-related data and updates the drug resistance feature database and knowledge graph.

8. The method for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation according to claim 7, characterized in that, In the model knowledge optimization phase, patient final efficacy data, adverse reactions, protocol adjustment records, and satisfaction scores collected after the end of chemotherapy cycles are collected to construct an efficacy assessment dataset. The predicted results are compared and analyzed with the actual efficacy to adjust model parameters and network structure and update the medical knowledge graph. An optimization report is generated, which includes model performance comparisons, updated knowledge graph content, and suggestions for optimizing the data collection system. The optimized model and knowledge graph are then applied to the analysis of new cases after multi-center clinical validation.

9. A system for individualized adjustment of tumor chemotherapy regimens and efficacy evaluation, based on the method of claim 8, characterized in that, include: Data acquisition and preprocessing module: integrates multi-source data acquisition unit, standardized processing unit and quality control unit, and outputs standardized data matrix with quality level labels; Multimodal collaborative prediction module: includes hybrid modeling unit that integrates medical knowledge graph, dual-branch network unit, reinforcement learning unit and federated learning unit, outputting drug sensitivity score and correlation evidence; The scheme generation and interpretation module has a built-in multi-constraint rule library and generates three candidate chemotherapy schemes and corresponding interpretation reports through rule reasoning and multi-objective optimization units. Real-time efficacy assessment module: includes dynamic data acquisition unit, weight adjustment and efficacy scoring unit and trend analysis unit, outputting efficacy level and early warning report; The protocol adjustment and drug resistance early warning module includes an intelligent decision tree unit, a multi-center data validation unit, a drug resistance early warning unit, and a graded intervention unit, and outputs adjustment reports, drug resistance risk levels, and intervention strategies. The model knowledge optimization module includes a efficacy assessment dataset construction unit, a model parameter adjustment unit, a knowledge graph update unit, and a clinical validation unit, which feeds back the optimization results to the aforementioned modules.

Images