Method and system for recommending personalized treatment scheme of lung cancer and storage medium

Abstract

The invention relates to the technical field of medical treatment, and discloses a lung cancer personalized treatment scheme recommendation method and system and a storage medium. The method comprises the following steps: acquiring clinical and molecular indexes of a patient, and constructing a simplified feature set; distributing weights for treatment targets according to the simplified feature set, constructing and optimizing a scheme evaluation matrix, and generating a preliminary scheme sorting list; through threshold screening and patient feature matching degree verification, a verified scheme set is obtained; the schemes are classified based on gene mutation and driver gene features, and classified optimization scheme subsets are obtained; constructing an interaction model to analyze interaction influence of toxic and side effects and life quality on curative effects, and dynamically adjusting scheme scores; and if the score is lower than a threshold value, triggering iterative optimization, obtaining a scheme list after iteration, and further determining an optimal treatment scheme according to treatment collaboration and target balance. According to the method, intelligent and closed-loop optimization from multi-source data to personalized treatment decision is realized, and the personalization and accuracy of a treatment scheme are improved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to a method, system, and storage medium for recommending personalized treatment plans for lung cancer. Background Technology

[0002] Lung cancer, one of the leading causes of cancer death worldwide, typically requires treatment tailored to each patient's specific circumstances, such as pathological type, tumor stage, and gene mutations. However, traditional treatment methods often fail to provide precise solutions based on individual patient differences, resulting in limited efficacy, significant side effects, and high economic costs associated with the treatment process.

[0003] In recent years, with the development of medical technology, personalized medicine has gradually become a research hotspot in the field of cancer treatment. Precision medicine, by combining genomics, molecular biomarkers, and clinical data, aims to provide the most suitable treatment plan for each patient. However, current methods for recommending personalized treatment plans still have many problems, such as difficulty in effectively integrating multidimensional data, lack of in-depth analysis of the interaction between different treatment goals, and the inability to achieve the best balance between treatment efficacy and side effects during the recommendation process.

[0004] To address these issues, this invention proposes a personalized treatment plan recommendation method for lung cancer. By comprehensively considering the patient's clinical indicators, molecular data, and gene mutation information, and utilizing advanced optimization algorithms and interactive models, it can accurately recommend the most suitable treatment plan for the patient, thereby improving treatment efficacy, reducing side effects, and providing a scientific basis for clinical decision-making. Summary of the Invention

[0005] This application provides a method, system, and storage medium for recommending personalized treatment plans for lung cancer. The aim is to provide the most suitable treatment plan for each patient by comprehensively considering multi-dimensional data such as clinical indicators, molecular data, gene mutations, and PD-L1 expression, combined with optimization algorithms and interactive models. The core idea of ​​this invention is to optimize treatment plans by integrating multiple treatment objectives (such as treatment efficacy, side effects, cost burden, and quality of life) through dimensionality reduction, weight optimization, classification processing, and interactive model analysis, thereby improving the accuracy and effectiveness of personalized treatment.

[0006] Firstly, this application provides a method for recommending personalized treatment plans for lung cancer, the method comprising: S1. Obtain multidimensional feature data from the patient's clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set; S2. Set multidimensional treatment goals, assign weights to each goal according to the simplified feature set, construct a treatment plan evaluation matrix and optimize it to generate a preliminary treatment plan ranking list; S3. Filter the schemes in the preliminary scheme ranking list by a preset score threshold to obtain a list of candidate optimal schemes, and verify the matching degree of the candidate optimal scheme list to obtain a set of verified schemes. S4. Extract classification feature data from the validated scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes; S5. Based on the subset of classification optimization schemes, construct an interaction model to analyze the interaction between toxic side effects and quality of life on treatment effects. Calculate the interaction impact data using the interaction model, and then adjust the scores of each scheme to obtain the adjusted scheme scores. S6. If the score of the adjusted solution is lower than the preset score threshold, an iterative solution list will be generated through iterative optimization. S7. Select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment plan.

[0007] Secondly, this application provides a personalized treatment plan recommendation system for lung cancer, the system comprising: The feature extraction module is used to obtain multidimensional feature data from patients' clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set. The treatment plan evaluation module is used to set multidimensional treatment goals, assign weights to each goal based on a simplified feature set, construct and optimize the treatment plan evaluation matrix, and generate a preliminary ranking list of treatment plans. The scheme filtering module is used to filter the schemes in the preliminary scheme ranking list by a preset score threshold, obtain a list of candidate optimal schemes, and verify the matching degree of the candidate optimal scheme list to obtain a set of verified schemes. The classification optimization module is used to extract classification feature data from the validated scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes. The interactive analysis module is used to construct an interactive model based on a subset of classification optimization schemes to analyze the interactive impact of toxic side effects and quality of life on treatment effects. The interactive model is used to calculate the interactive impact data, and then the scores of each scheme are adjusted to obtain the adjusted scheme scores. The iterative optimization module is used to generate an iterative list of solutions if the score of the adjusted solution is lower than the preset score threshold. The optimal decision module is used to select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment plan.

[0008] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned method for recommending a personalized treatment plan for lung cancer.

[0009] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. It achieves precise mapping from high-dimensional data to personalized weights, improving the individual adaptability of the treatment plan. By reducing the dimensionality of clinical and molecular indicators, redundant noise is eliminated while retaining key information; and based on this, the weights of multiple treatment goals are dynamically determined by simplifying the feature set. This solves the problems of insufficient personalization and weak targeting of the treatment plan caused by high data dimensionality, feature redundancy, and fixed weight allocation, making the recommended plan closely integrated with the patient's specific pathological characteristics.

[0010] 2. By quantifying the interactive effects, a scientific and dynamic trade-off among multiple objectives is achieved, optimizing the overall balance of the treatment plan. Interactive models (such as Bayesian networks) are constructed to quantitatively analyze the correlation between factors such as side effects and quality of life on treatment efficacy, and the plan score is dynamically adjusted accordingly. This makes plan optimization no longer dependent on simple linear weighting of individual objectives, but rather on a nonlinear trade-off that reflects complex clinical realities. This effectively solves the technical challenge of potentially neglecting safety or quality of life in the pursuit of therapeutic efficacy, generating a more balanced treatment combination under multiple constraints.

[0011] 3. By streamlining the screening process and focusing on key optimization areas, the overall efficiency and stability of the intelligent decision-making system have been improved. Multiple rounds of solution screening and classification using preset thresholds allow subsequent complex interactive analysis and iterative optimization to focus on high-potential subsets of solutions, avoiding blind and time-consuming calculations across the entire solution space. This "gradual focusing and deep optimization" strategy achieves rational and efficient allocation of computing resources while ensuring recommendation quality.

[0012] 4. An adaptive closed-loop optimization system was constructed to ensure the continued excellence and reliability of the recommendation results. By introducing a scoring-based iterative optimization mechanism, the system can automatically adjust core parameters and re-optimize when the output plan does not meet the requirements, forming a complete closed loop of "generation-evaluation-feedback-optimization". This overcomes the limitations of the unidirectional and static nature of traditional recommendation methods, endowing the system with the ability to self-improve, thereby continuously outputting reliable and stable optimal treatment plans and providing strong intelligent support for clinical decision-making. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a method for recommending a personalized treatment plan for lung cancer according to this application; Figure 2 This is a schematic diagram of the Bayesian network structure according to an embodiment of this application; Figure 3 This is a schematic diagram of a high-probability heatmap of the treatment effect in an embodiment of this application. Figure 4 This is a schematic diagram of the 3D surface plot of the interaction effect in an embodiment of this application; Figure 5 This is a schematic diagram of a personalized treatment plan recommendation system for lung cancer according to this application. Detailed Implementation

[0015] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of a method for recommending a personalized treatment plan for lung cancer provided by the present invention. The flowchart specifically includes the following steps: S1. Obtain multidimensional feature data from the patient's clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set.

[0017] In one specific embodiment, the process of performing step S1 may specifically include the following steps: Multidimensional feature data are obtained from the patient's clinical and molecular indicators. The multidimensional feature data includes at least tumor size, lymph node metastasis status, gene mutation type, PD-L1 expression level and driver gene status. Using patient samples as rows and data dimensions as columns, an initial data matrix is ​​constructed. After standardizing the initial data matrix, a multidimensional data matrix is ​​obtained. Principal component analysis was used to reduce the dimensionality of the multidimensional data matrix, and the top k eigenvectors whose cumulative variance contribution rate reached the preset contribution threshold were selected as the principal feature components. The multidimensional data matrix is ​​projected onto the feature space spanned by the main feature components to generate a simplified feature set.

[0018] Specifically, structured and unstructured data are obtained from hospital information systems and molecular testing reports to ensure that the clinical indicators comprehensively reflect the patient's condition, while molecular indicators provide important information related to treatment efficacy. Clinical and molecular indicators reflect the patient's tumor characteristics, immune response, and genetic factors. Clinical indicators are derived from imaging reports and pathology records; for example, the maximum diameter of the tumor is measured using CT images as tumor size data, and the number of lymph nodes involved in the pathology report represents lymph node metastasis status data. Molecular indicators are derived from gene sequencing and immunohistochemistry reports; for example, mutation status of genes such as EGFR, ALK, and ROS1 is detected through next-generation sequencing, and the percentage expression level of PD-L1 protein is detected through immunohistochemistry. Driver gene status is binarized based on the presence of specific gene fusions or amplifications. These data are then integrated into a multidimensional feature data matrix, forming a matrix with multiple dimensions, where each row represents a patient sample and each column represents different feature data.

[0019] After constructing the initial data matrix, standardization is performed to ensure data consistency. The purpose of standardization is to bring data with different features to the same order of magnitude, preventing certain features from dominating subsequent calculations due to differences in scale. For example, tumor size and gene mutation type have significantly different numerical ranges. Without standardization, tumor size might have an excessively large impact on the final result, while neglecting the equally crucial feature of gene mutation. Standardization transforms the data into a distribution with a mean of 0 and a variance of 1 by subtracting the mean from each column and dividing by the standard deviation. This ensures that all features are compared under the same standard, avoiding biases caused by scale and data distribution.

[0020] Due to the high dimensionality and numerous features involved in the data, direct processing may lead to excessive computational complexity and susceptibility to redundant features and noise. Therefore, Principal Component Analysis (PCA) is employed for dimensionality reduction. PCA is a commonly used data dimensionality reduction method that aims to map the original high-dimensional data to a new low-dimensional space through linear transformation, while preserving the variance and information of the data to the greatest extent possible. In PCA, the covariance matrix of the data matrix is ​​first calculated. The covariance matrix reflects the relationships between the various features, i.e., how they change together. Then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvectors represent the new direction, while the eigenvalues ​​represent the magnitude of the variance of the data in that direction. By selecting the first k eigenvectors with the largest eigenvalues, it is ensured that the selected principal components can preserve the variance of the original data to the greatest extent possible, thereby achieving dimensionality reduction. These k eigenvectors constitute the principal eigencomponents, and each eigenvector is a linear combination of all the original dimensions. The choice of k value is usually determined by the cumulative variance contribution rate to ensure that most of the information of the original data is retained. Setting a cumulative variance contribution rate threshold (such as 90% or 95%) can help select the most suitable k value and ensure that the data is still representative after dimensionality reduction.

[0021] The essential operation of dimensionality reduction is to project the original multidimensional data matrix onto a new coordinate system spanned by these k principal feature components. This projection process is achieved through matrix multiplication. The standardized data matrix is ​​multiplied by a transformation matrix composed of the k feature vectors, resulting in a new matrix with the same number of rows as the samples, but with the number of columns reduced to k. This new matrix is ​​the simplified feature set, where each column represents a new synthetic feature. These synthetic features are independent of each other and capture the most prevalent variation patterns in the original data. The simplified feature set contains the most important information from the original data, removing redundancy and noise, thus providing clearer and more concise input data for subsequent treatment recommendations.

[0022] S2. Set multidimensional treatment goals, assign weights to each goal based on a simplified feature set, construct a treatment plan evaluation matrix and optimize it to generate a preliminary treatment plan ranking list.

[0023] In one specific embodiment, the process of performing step S2 may specifically include the following steps: Set multidimensional treatment goals, which should include at least treatment efficacy, side effects, cost burden, and quality of life. Based on the feature components related to each target in the simplified feature set, an initial weight parameter is determined for each treatment target to form an initial weight vector; Obtain a set of candidate treatment options and determine the expected score of each option on multidimensional treatment goals. Construct an evaluation matrix for the options based on the expected scores. Starting with the initial weight vector, a genetic algorithm is used to iteratively optimize the weight vector. During the optimization process, the weight vector being evaluated is weighted and calculated with the scheme evaluation matrix to obtain the comprehensive score of each scheme. The comprehensive score is used as the basis for evaluating the fitness of the weight vector until the optimal weight vector that meets the convergence condition is obtained. The evaluation matrix of each scheme is weighted according to the optimal weight vector to obtain the final comprehensive score of each scheme and generate a preliminary sorted list of schemes.

[0024] Specifically, multiple dimensions of treatment goals need to be optimized. These dimensions directly correspond to core considerations in clinical decision-making, including key factors such as treatment efficacy, side effects, cost burden, and quality of life. Treatment efficacy represents the inhibitory or curative effect of treatment on the tumor; side effects describe adverse reactions caused by treatment; cost burden involves the economic affordability of the treatment plan; and quality of life refers to the patient's quality of life during treatment. These goals are interdependent, ensuring that the treatment plan can effectively treat the tumor while also taking into account the patient's health and economic burden.

[0025] When assigning initial weights to each treatment objective, it is first necessary to analyze the relevant feature components of each objective to determine which features have the greatest impact on that objective. For example, features related to treatment efficacy objectives may include tumor size, gene mutation type, etc., while features related to toxic side effects may involve immune response level, patient tolerability, etc. By analyzing the importance of these features in simplifying the feature set, and combining medical expert experience or historical data, an initial weight parameter can be assigned to each treatment objective. The initial weight parameter reflects the weight proportion of each treatment objective in the entire optimization process, initially reflecting the priority of each objective when recommending treatment plans. Specifically, the determination of the initial weight parameter can be achieved by quantitatively analyzing the variance, correlation, and actual impact on the treatment effect of each treatment objective's relevant features, ensuring that the weight allocation of each objective reflects its clinical importance while also taking into account the balance between different treatment objectives. Preferably, the normalized values ​​of relevant feature components can also be mapped to initial weight values ​​between 0 and 1 using a preset linear or nonlinear function. All weight values ​​constitute an initial weight vector, and the sum of its elements is usually normalized, for example, to 1.

[0026] A set of candidate treatment options is defined using clinical guidelines, drug databases, and expert consensus. These options may include single-agent or combination therapies such as chemotherapy, targeted therapy, and immunotherapy. The treatment strategy, mechanism of action, and expected efficacy of each option are evaluated based on clinical trial data, pharmacoeconomic analysis, and other factors. For each treatment option, expected scores are determined based on existing medical literature and clinical data across four treatment objectives: efficacy, toxicity, cost, and quality of life. These scores reflect the treatment option's performance at specific objectives. For example, the expected efficacy score may be based on the objective response rate in different patient subgroups; the toxicity score is based on the incidence of adverse reactions in phase III clinical trials; the cost is calculated based on the total cost of treatment; and the quality of life score is typically assessed in conjunction with patient-reported outcomes. After obtaining the expected scores for each treatment option at each treatment objective, these scores are organized into a matrix. Each row of the matrix corresponds to a treatment option, each column corresponds to a treatment objective, and each element in the matrix represents the expected score of the treatment option at that specific objective. This matrix, called the treatment option evaluation matrix, demonstrates the performance of each treatment option across various dimensions.

[0027] Next, a genetic algorithm is used to iteratively optimize the weight vector. The genetic algorithm is an optimization method that simulates the principles of natural selection and genetics. It finds the optimal solution by repeatedly iterating the weight vector and adjusting the weight combinations. In this invention, the optimization objective of the genetic algorithm is to maximize the overall score of the treatment objective. Therefore, the algorithm continuously adjusts the weight vector and combines these weights with the scheme evaluation matrix to calculate the overall score of each scheme.

[0028] In the initialization phase, based on the initial weight vector, an initial population P_0 containing N individuals (i.e., candidate weight vectors) is generated by applying a random perturbation Δ following a uniform or Gaussian distribution to each dimension. Each individual in the population represents a four-dimensional weight vector, and the sum of all its elements is constrained to 1. Fitness evaluation is the core driving mechanism of the algorithm. For each weight vector ω_t in the population, it is multiplied by the solution evaluation matrix to obtain a column vector Score_t, where the i-th element represents the comprehensive score of the i-th candidate treatment under weight ω_t. The fitness function F is defined as a certain statistic of the score vector Score_t to quantify the global merit of the weight combination. A typical definition is to take the maximum value in Score_t; another definition could be the mean or a specific quantile of Score_t to accommodate different optimization preferences. In each generation of the evolutionary cycle, the algorithm performs selection operations based on the fitness values ​​of individuals. For example, a tournament selection strategy can be used: K individuals are randomly selected from the population, and the individual with the highest fitness is selected and retained in the next generation parent pool. This process is repeated until the parent pool is full. Then, a crossover operation is performed on the individuals in the parent pool, randomly pairing parent vectors and exchanging the weights of some dimensions with probability P_c to generate new child vectors. Next, each weight value of the child vector is mutated with a small probability P_m by adding a tiny random offset. After mutation, the weight vector needs to be normalized to maintain a sum of 1. The resulting new generation population P_t+1 will enter the next round of evaluation and evolution. The iterative process is controlled by preset convergence conditions. Common convergence conditions include: the number of iterations reaches a preset upper limit T_max; or, the absolute value of the change in the optimal fitness F_best over G consecutive generations is less than a very small threshold ε. When any convergence condition is met, the algorithm terminates and outputs the weight vector with the highest fitness in the current population as the optimal weight vector. This vector can most reasonably reflect the relative importance of each treatment target and ensure the optimal performance of the treatment plan in multiple dimensions.

[0029] Based on the obtained optimal weight vector, the treatment plan evaluation matrix is ​​weighted again to obtain the final comprehensive score for each treatment plan, generating a preliminary ranking list. The ranking process considers not only the performance of each treatment plan across various dimensions but also personalizes the ranking based on the patient's specific characteristics and treatment goals. The ranking results ensure that the recommended treatment plans provide optimal treatment outcomes, minimal side effects, acceptable cost burden, and high quality of life based on the comprehensive assessment.

[0030] This invention addresses the issues of data redundancy and noise interference, the trade-off between conflicting treatment goals, and local optima through effective processing of multidimensional feature data, scientific allocation of target weights, dynamic optimization using genetic algorithms, and precise calculation of the comprehensive score of the treatment plan. The closed-loop optimization process ensures continuous improvement and precision of the treatment plan, enhances the personalization and reliability of lung cancer treatment, and significantly improves treatment outcomes and patient quality of life.

[0031] S3. Filter the schemes in the preliminary scheme ranking list by a preset score threshold to obtain a list of candidate optimal schemes, and verify the matching degree of the candidate optimal scheme list to obtain a set of verified schemes.

[0032] In one specific embodiment, the process of performing step S3 may specifically include the following steps: Obtain the comprehensive score data of each scheme in the preliminary scheme ranking list, filter out the schemes with comprehensive scores higher than the preset score threshold, and form a candidate optimal scheme list; Extract patient feature vectors related to individual patient characteristics from the simplified feature set. The patient feature vectors include at least tumor size, lymph node metastasis status, gene mutation type, PD-L1 expression level, and driver gene status. For each candidate optimal solution, the matching degree is verified based on the treatment attribute vector of the solution and the patient feature vector. The treatment attribute vector represents the treatment strategy and intensity of the treatment solution for each feature in the patient feature vector. Generate a set of validated solutions based on the validation results.

[0033] Specifically, each treatment option in the preliminary ranking list has undergone initial optimization and scoring. Its overall score reflects the treatment option's performance across multiple dimensions (such as treatment efficacy, side effects, quality of life, and cost). By processing the overall scores of the options, treatment options with scores above a preset threshold are selected, forming a candidate optimal option list. The selection threshold is typically determined by medical experts or historical data experience to ensure that the selected treatment options meet certain clinical requirements in terms of overall score. Threshold screening removes low-scoring options that do not meet treatment requirements, ensuring that the options used in subsequent evaluations are high-potential treatment options that meet clinical needs.

[0034] For each selected candidate optimal treatment plan, a matching degree verification is required. The purpose of matching degree verification is to ensure that the candidate plan not only performs well in terms of abstract comprehensive scores, but also closely matches the patient's specific and subtle pathophysiological characteristics, thereby improving the adaptability of the treatment plan to the individual patient's condition. In this process, patient feature vectors related to individual patient characteristics are first extracted from the simplified feature set. These feature vectors include, but are not limited to, tumor size, lymph node metastasis status, gene mutation type, PD-L1 expression level and / or driver gene status, which can comprehensively reflect the patient's pathological characteristics, immune status and genetic background. These specific values ​​are arranged in a fixed order to form a patient feature vector, thus standardizing these complex biological data.

[0035] For each treatment option in the candidate optimal treatment list, its corresponding treatment attribute vector needs to be obtained. This vector represents the treatment strategy and intensity of the treatment option targeting patient characteristics and needs to be constructed based on the specific drug composition and treatment mechanism of the option. The treatment attribute vector reflects how the treatment option designs corresponding treatment strategies for specific pathological characteristics, immune responses, and genetic characteristics of the patient. For example, some targeted therapies may specifically target certain gene mutations, while immunotherapy may adjust the treatment intensity based on PD-L1 expression levels. For example, for a "pembrolizumab combined with platinum-based chemotherapy" option, its treatment attribute vector needs to encode the effect of the option on the aforementioned patient characteristics: its treatment intensity for tumors with high PD-L1 expression may be set to a high value, for tumors with specific driver gene mutations it may be set to a moderate or low value (depending on whether the gene affects the efficacy of immunotherapy), and the inhibitory effect on tumor size and lymph node metastasis is assigned an empirical intensity value based on the objective response rate data of this combination therapy in advanced lung cancer. These intensity values ​​may originate from subgroup analysis data of the registered clinical trials of the protocol, scientific literature on drug mechanisms of action, and expert knowledge bases, and are normalized to a uniform scale, such as between 0 and 1, to form a treatment attribute vector with the same dimension and order as the patient's feature vector.

[0036] The matching degree is calculated by matching the patient's feature vector with the treatment attribute vector of the treatment plan. This matching degree measures the degree of consistency between the treatment plan and the patient's characteristics. The level of matching degree directly affects the personalization and adaptability of the treatment plan, ensuring that the recommended treatment plan can best meet the patient's actual needs in clinical practice. The matching degree is calculated by comparing the similarity between the patient's feature vector and the treatment attribute vector of the treatment plan, typically using methods such as cosine similarity or Euclidean distance to quantify the degree of matching.

[0037] After completing the matching degree verification, treatment plans are screened based on the matching degree results to select the most suitable treatment plan for the patient, forming a verified plan set. Preferably, a matching degree threshold can be set, and treatment plans with a matching degree higher than the threshold are selected to form the verified plan set. The verified plan set fully considers the adaptability between the patient's individual characteristics and the treatment plan, ensuring that the recommended treatment plan not only has good therapeutic effect but also minimizes side effects and improves quality of life.

[0038] Through the above steps, this invention resolves the problems of conflicting treatment goals, insufficient personalization in treatment recommendations, and lack of optimization. The pre-set score threshold screening ensures that only high-potential treatment plans that meet treatment efficacy are selected, avoiding inefficient or unsuitable plans. The matching verification step, by comprehensively considering the patient's characteristics and the suitability of the treatment plan, ensures that the recommended treatment plan can be optimized according to the patient's individual differences, thereby maximizing the personalization and precision of the treatment plan, ultimately improving treatment efficacy and reducing side effects. This process not only ensures the accuracy of the treatment plan but also ensures continuous improvement in treatment efficacy through iterative optimization.

[0039] S4. Extract classification feature data from the validated scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes.

[0040] In one specific embodiment, the process of performing step S4 may specifically include the following steps: Categorical feature data are extracted from the molecular indicators associated with the validated protocol set. The categorical feature data includes at least the gene mutation status and the driver gene status. Based on the classification feature data, the schemes in the validated scheme set are grouped to obtain the classification results; Based on preset screening criteria, a subset of classification optimization schemes is determined according to the classification results. The preset screening criteria include at least the degree of matching with patient characteristics, comprehensive evaluation of treatment goals, and / or clinical efficacy.

[0041] Specifically, the validated protocol set, after matching verification, is further refined to address the issue of high heterogeneity within the protocol set, which hinders subsequent unified and in-depth interaction analysis. The core of this step lies in classifying the protocols based on their inherent molecular biological characteristics, thereby building a more homogeneous analytical foundation for subsequent steps.

[0042] Each treatment regimen in the validated regimen set corresponds to a specific patient context, from which key molecular marker states are retrieved. These markers include, but are not limited to, gene mutation status and driver gene status. Gene mutation status and driver gene status are key factors influencing the effectiveness of lung cancer treatment regimens, especially in targeted therapy and immunotherapy, where mutations or fusions of certain genes can differentiate treatment outcomes. Therefore, based on these molecular indicators, it is possible to better determine whether each treatment regimen is suitable for a specific patient. For example, gene mutation status involves specific information such as the presence of exon 19 deletion or L858R point mutation in the EGFR gene, whether ALK gene rearrangement has occurred, whether ROS1 has fusion, and KRAS mutation; driver gene status further distinguishes which mutations are identified as key events driving tumor growth, possibly involving the status of EGFR, ROS1, BRAF, etc. This information is usually in the form of structured binary or categorical variables. For example, EGFR mutation status may be "19-del positive," "L858R positive," or "wildtype"; ALK rearrangement status may be "positive" or "negative." For the m schemes in the validated scheme set, the molecular markers corresponding to each scheme are extracted in parallel and encoded into a classification feature vector c_j, where j identifies the scheme number. The set of classification feature vectors of all schemes constitutes an m-row, d-column matrix C, where d is the number of molecular feature dimensions considered. Each row c_j of matrix C uniquely represents the molecular subtype background that is pre-defined when adopting the scheme.

[0043] Using the extracted classification feature data, the treatment plans in the validated plan set are grouped. The classification process can employ unsupervised clustering algorithms, such as the k-means algorithm, which aims to group plans with similar classification feature vectors into the same cluster. In practice, a distance metric needs to be predefined; for vectors containing categorical variables, Jaccard distance or Hamming distance can be used to calculate the differences between plans. The number of clusters to be generated is set to K1, and the value of K1 can be pre-defined using the elbow rule or based on business knowledge. During algorithm initialization, K1 feature vectors of the plans are randomly selected as initial cluster centers, and then the following two steps are iteratively executed: each plan is assigned to the cluster containing the nearest cluster center; the "centroid" of all feature vectors of all plans in each cluster is recalculated (for categorical variables, the centroid can be defined as the vector composed of the modes of each feature dimension). Iteration continues until the cluster centers no longer change significantly or the preset maximum number of iterations is reached. At this point, the algorithm converges, and the cluster label to which each plan belongs is output, forming the classification result. This process groups protocols with similar molecular characteristics into one category. For example, all targeted therapies against EGFR-sensitive mutations may cluster into one cluster, while protocols with high PD-L1 expression and no specific driver gene may cluster into another, thus achieving the structuring of protocol sets at the biomarker level.

[0044] Preferably, the grouping process considers not only the single criterion of gene type, but also other possible influencing factors, such as the patient's treatment history, age, and immune status, thus enabling a more comprehensive evaluation of the suitability of each treatment regimen. This grouping method allows for the initial screening of different treatment regimens based on their effectiveness and patient suitability, ensuring that the selected regimens are optimized based on the patient's characteristics and treatment goals.

[0045] The system interprets and selects from the classification results based on preset screening criteria, which include at least the degree of matching with patient characteristics, a comprehensive assessment of treatment goals, and / or clinical efficacy. The system does not re-evaluate individual treatment plans; instead, it evaluates them on a cluster basis based on the classification results. For example, for the criterion of "degree of matching with patient characteristics," the average matching score μ_M obtained by all plans within each cluster in step S3 can be calculated, and the cluster with the highest μ_M can be selected. For "comprehensive assessment of treatment goals," the average final comprehensive score μ_S obtained by all plans within each cluster in step S2 can be calculated, and the cluster with the highest μ_S can be selected. "Clinical efficacy" may refer to external knowledge; for example, based on the latest clinical research evidence, a certain molecular subtype (corresponding to a specific cluster) has been shown to have an excellent response rate to a certain type of treatment plan, and that cluster can be prioritized. These criteria can be used individually or in a weighted combination to generate a priority score for each cluster. One or more target clusters are determined from the classification results according to preset rules (e.g., selecting the cluster with the highest priority score, or selecting all clusters with priority scores exceeding a threshold γ).

[0046] Based on the identified target cluster labels, all protocols belonging to these clusters are extracted from the validated protocol set. This subset constitutes the classification optimization protocol subset. Protocols within this subset not only perform well in initial comprehensive scores and individual matching, but more importantly, they exhibit high homogeneity in key molecular subtyping. This homogeneity is crucial for the interaction model constructed in step S5, as the correlation patterns between toxic side effects, quality of life, and treatment efficacy can vary depending on the molecular subtype. Limiting the analysis to a homogeneous subset of protocols avoids confounding effects from different biological mechanisms, allowing the interaction model to more clearly and stably capture the relationships between factors within a specific patient subgroup, thereby improving the reliability and specificity of subsequent score adjustments. Through this step, the system further refines and focuses upon the "validated high-potential protocol pool" based on core biological logic, providing technical support for addressing the lack of consideration for inherent biological heterogeneity in solution evaluation and the potential inaccuracies in interaction analysis due to data confounding.

[0047] S5. Based on the subset of classification optimization schemes, construct an interaction model to analyze the interaction between toxic side effects and quality of life on treatment efficacy. Calculate the interaction impact data using the interaction model, and then adjust the scores of each scheme to obtain the adjusted scheme scores.

[0048] In one specific embodiment, the process of performing step S5 may specifically include the following steps: Quantitative score data for each classification optimization scheme in terms of toxic side effects and quality of life were extracted to obtain interaction impact data. Construct an interactive model to analyze the interaction between toxic side effects and quality of life on treatment efficacy; Input the interaction impact data into the interaction model to obtain the interaction impact results; Based on the results of the interaction effects, the scores of each scheme in the subset of classification optimization schemes are adjusted to obtain the adjusted scheme scores.

[0049] Specifically, the purpose of this step is to precisely adjust the overall score of the treatment plan by considering the impact of the treatment plan's performance in terms of toxic side effects and quality of life on the treatment outcome, thereby ultimately providing a more personalized and precise treatment plan.

[0050] The subset of optimized treatment options, obtained through preliminary screening, includes various treatment plans suitable for patients. For each option, toxicity and quality of life are two important indicators for evaluating treatment efficacy. First, for each optimized treatment plan, corresponding quantitative scores are extracted for both toxicity and quality of life. These scores are typically derived from clinical trial data, patient-reported quality of life assessment scales, and drug side effect databases. Toxicity scores are usually quantified by the incidence and severity of adverse reactions (e.g., the incidence of grade 3 or higher adverse reactions), while quality of life scores are assessed using patient quality of life scores and health status scales. Toxicity and quality of life scores are paired to form interactive data pairs, providing an interactive data set for subsequent interactive analysis.

[0051] Interaction models are pre-built models designed to analyze the interactive effects of side effects and quality of life on treatment outcomes. The core of interaction models lies in establishing the conditional dependencies among treatment outcomes, side effects, and quality of life, and determining the interaction patterns among them. Interaction models are typically built using methods such as multiple regression analysis, machine learning algorithms (e.g., decision trees, support vector machines), or Bayesian networks. These models can simulate the complex relationships between side effects, quality of life, and treatment outcomes, and quantify their interactive effects. The data used to build interaction models is extracted from historical patient clinical data containing real treatment outcomes.

[0052] Preferably, an interaction model is established using Bayesian networks. This structure can effectively represent the interdependencies between different treatment factors, especially when facing complex and uncertain biological and clinical data. Bayesian networks provide a flexible framework for representing the conditional probabilities between variables. This model enables quantitative analysis of the impact of toxic side effects and quality of life on treatment efficacy, addressing the lack of modeling and analysis of the complex relationships inherent in treatment regimens in traditional methods. Through conditional probability tables, the interaction model can quantify the changes in treatment efficacy under different levels of toxic side effects and quality of life. The conditional probability table records the impact of quality of life on treatment efficacy at a given level of toxic side effects. For example, when toxic side effects are severe, a decline in quality of life may significantly weaken the treatment effect, thus affecting the patient's treatment response. By estimating historical clinical data or simulated data, the marginal probability of treatment efficacy for each treatment regimen under different conditions can be accurately calculated, providing a precise basis for subsequent treatment regimen evaluation.

[0053] For example, an interaction model is constructed using a Bayesian network, the structure of which is designed for the current analytical objective. The model contains three explicit nodes: node X representing the toxicity score, node Y representing the quality of life score, and node Z representing the treatment outcome score. Nodes X and Y serve as parent nodes, and node Z as a child node, with directed edges pointing from X and Y to Z. This structure encodes the prior knowledge that "toxicity and quality of life jointly influence treatment outcome." The core of the model is the conditional probability table for node Z. This table defines the probability P(Z=high|X=x, Y=y) that the treatment outcome score z takes a high value (e.g., representing objective remission or disease control) given a specific set of toxicity scores x and quality of life scores y. The parameters of the conditional probability table need to be learned based on historical data. In one embodiment, it can be trained using a separate historical clinical database containing numerous observational records of toxicity, quality of life, and treatment outcomes after patients received different treatment regimens. Each historical record contains an actual regimen used and its actual results in the X, Y, and Z dimensions. Statistical learning algorithms can be used to estimate the probability distribution P(Z|X,Y). For example, it might be found that when the side effect score x is at a low to medium level and the quality of life score y is at a high level, the probability of P(Z=high) increases significantly; while when x is very high, the probability of P(Z=high) is low regardless of y. This model links discrete score data with the probability of treatment success.

[0054] Interaction data is input into the constructed interaction model for forward inference calculations, yielding the interaction results for each treatment plan. These results reflect the true impact of the treatment plan on the overall treatment effect after considering the interaction between side effects and quality of life. After the interaction results are generated, each treatment plan is scored and adjusted using these results. By applying a weighted adjustment coefficient to the marginal probabilities of the treatment effect, the score of the treatment plan is adjusted upwards or downwards based on the strength of the interaction. Specifically, when the interaction model calculates a high probability of a treatment effect, the score of the treatment plan is increased accordingly; if the interaction results show that the treatment effect is negatively affected by side effects or quality of life, the score is decreased. This dynamic score adjustment ensures that the treatment plan can be optimized according to changes in actual treatment effects, solving the problem of the inability to dynamically adjust target weights in traditional treatment plan evaluations.

[0055] This process resolves multi-objective conflicts and complex dependencies between treatment goals, especially in scenarios where side effects and quality of life significantly impact treatment outcomes. By introducing an interactive model, the effectiveness of treatment plans can be evaluated more comprehensively, avoiding the oversight of intrinsic connections and interactions between treatment goals by simple weighted summation methods. This allows for treatment recommendations that better align with patients' individual needs. In clinical settings such as non-small cell lung cancer treatment, this model optimization significantly improves the precision of treatment plans, reduces side effects, and enhances patients' quality of life, providing strong technical support for clinical decision-making.

[0056] Figure 2 , Figure 3 and Figure 4 The interaction model effect verification diagrams provided in this application embodiment use three diagrams to illustrate the core principles and quantification effects of the interaction model in a three-dimensional way: Figure 2 The diagram shows a Bayesian network structure, illustrating a model constructed to analyze the relationship between side effects, quality of life, and treatment effectiveness. In the diagram, the side effect score node and the quality of life score node serve as parent nodes, both pointing to the treatment effectiveness score child nodes. This directed acyclic graph structure intuitively encodes the prior medical knowledge that "side effects and quality of life jointly influence treatment effectiveness," providing a framework for subsequent probabilistic reasoning.

[0057] Figure 3A heatmap of high probability of treatment efficacy is generated, which visualizes the conditional probability distribution learned from historical clinical data. The horizontal axis represents the toxicity score (1 being the mildest and 10 being the most severe), and the vertical axis represents the quality of life score (1 being the worst and 10 being the best). The color intensity of each coordinate point in the graph represents the probability value P (Z=high|X=x, Y=y) of achieving high efficacy under that specific combination of toxicity and quality of life level. As shown in the diagram: In region ① (low side effects, high quality of life), the probability of achieving high efficacy is highest (light-colored region). For example, when the side effect score is 2 and the quality of life score is 8, the probability of high efficacy can reach above 0.85. In region ② (high side effects, low quality of life), the probability of achieving high efficacy is lowest (dark-colored region). For example, when the side effect score is 8 and the quality of life score is 2, the probability of high efficacy drops below 0.15. In region ③ (high side effects but also high quality of life), the probability of high efficacy remains at a low level (approximately 0.3–0.4). This quantifies the complex interactive phenomenon that the severity of side effects may weaken or even offset the positive contribution of high quality of life to efficacy.

[0058] Figure 4 The interaction effect 3D surface plot further demonstrates the effect in three dimensions. Figure 3 The three-dimensional shape of the probability distribution more clearly reveals the nonlinear trend and gradient change of the probability of treatment effect with the two input factors. The peak-valley distribution on the surface intuitively shows that after the toxic side effects exceed a certain threshold, their negative impact on the efficacy increases at an accelerated rate, while the improvement of quality of life has a "diminishing marginal effect" on the efficacy.

[0059] Through this interactive model, the system in this application surpasses traditional linear weighted evaluation methods, accurately capturing and quantifying the complex nonlinear trade-offs between efficacy, safety, and quality of life in clinical practice. During the protocol scoring adjustment phase, the final score of each candidate protocol is dynamically adjusted based on the conditional probabilities calculated by this model: for... Figure 3 The scores for regimens corresponding to the combination of toxicity / quality of life in the lighter-colored areas will be significantly improved; for regimens in the darker-colored areas, the scores will be correspondingly reduced. This dynamic adjustment mechanism based on probabilistic reasoning ensures that the recommended regimens not only pursue theoretically high efficacy but also take into account the level of side effects that patients can tolerate and maintain their quality of life in actual clinical practice, thereby generating more personalized treatment plans that meet real-world clinical needs.

[0060] S6. If the score of the adjusted solution is lower than the preset score threshold, an iterative solution list will be generated through iterative optimization.

[0061] In one specific embodiment, the process of performing step S6 may specifically include the following steps: Obtain the adjusted solution score and compare it with the preset score threshold. If the highest solution score is lower than the preset score threshold, trigger the iterative optimization process. Using the weight parameters of the current multidimensional treatment goals and / or candidate treatment plans as optimization variables, a genetic algorithm is used for re-optimization, adjusting the weight parameters and plan combinations; Based on the optimization results, updated weight parameters and / or treatment plans are generated, and the comprehensive score of all relevant candidate treatment plans is recalculated to generate an iterative list of plans.

[0062] Specifically, the adjusted score for each candidate treatment plan is obtained, and then the score of each treatment plan is compared with a preset scoring threshold to determine whether optimization is needed. This threshold is set based on the clinically acceptable minimum efficacy standard or the score distribution of historical successful cases, for example, set at 80 points. If the highest score is lower than the set threshold, it means that the current treatment plan has not yet achieved the expected therapeutic effect or the patient's needs, and the optimization process is initiated. Based on this, a genetic algorithm is used to optimize the treatment plan. As an optimization method that simulates natural selection and genetic mechanisms, the genetic algorithm can effectively handle multi-objective optimization problems.

[0063] The process of re-optimizing treatment plans using genetic algorithms, adjusting weight parameters and treatment combinations, is based on current multidimensional treatment goals and candidate treatments. Specifically, treatment goals include therapeutic efficacy, side effects, quality of life, and cost burden, while treatment combinations consist of the drug components, efficacy, and adaptability to different patients among various treatment options. Through optimization using genetic algorithms, the weight parameters of these goals and the combination of treatment plans can be dynamically adjusted to achieve optimal treatment outcomes.

[0064] After triggering an iteration, the system uses the current round's state as the starting point for optimization, rather than starting from scratch. During optimization, the weight parameters of multidimensional treatment goals and candidate treatment plans are input into the genetic algorithm as optimization variables. The weight parameters of the treatment goals are obtained by clinical experts or through data training in the preliminary analysis, reflecting the importance of different treatment goals in the comprehensive evaluation. During the genetic algorithm optimization process, the initial set of weight parameters and treatment plans are treated as individuals in the population. The genetic algorithm performs selection, crossover, and mutation operations on these individuals through multiple generations of iteration to generate the optimal solution. During population initialization, initial individuals are randomly generated from the weight parameters and plans of the current treatment plans. Each individual represents a specific combination of treatment plans and weight parameters. For example, assuming the initial weight parameters are treatment efficacy 0.4, toxicity 0.3, quality of life 0.2, and cost burden 0.1, the initial treatment plan may include a combination of a chemotherapy drug and a targeted therapy drug. By randomly perturbing these weight and plan combinations, multiple initial individuals are generated, forming the initial population. Next, the genetic algorithm selects individuals with high fitness from the population through a selection operation. These individuals represent the current optimal combination of treatment options. Fitness is evaluated based on the comprehensive score of each treatment option, which is calculated by weighting the weight parameters with the effectiveness data of the treatment options. For example, options with high treatment effectiveness scores may be prioritized, while options with greater toxicity may be eliminated. The selection operation ensures that only the most promising individuals can enter the next generation. During the crossover operation, selected individuals exchange some of their weight parameters or combinations of treatment options. The crossover operation mimics gene exchange in nature, aiming to combine the advantages of different individuals to generate new solutions. It allows the exchange of partial weight components or subsets of the treatment option set between different individuals. For example, crossing an individual's highly effective treatment combination with another individual's low-toxicity treatment combination may result in a new, comprehensively optimized solution. This process helps discover potential highly effective treatment options. The mutation operation simulates mutation phenomena in nature by randomly changing the weights or treatment options of some individuals. It randomly fine-tunes a weight value with a preset probability or randomly replaces a treatment option in the set of options. This step increases the diversity of the optimization process and prevents the algorithm from getting trapped in local optima. Mutation operations typically fine-tune the weights of individuals, changing the drug combination of the treatment plan or adjusting the weight ratio of treatment goals. Through mutation, the genetic algorithm can explore a wider solution space and discover treatment plans more suitable for the patient. After several generations of iterative optimization, the genetic algorithm converges to an optimal solution based on the fitness evaluation results—that is, a combination of treatment plans that maximizes the therapeutic effect while also considering goals such as toxicity, side effects, and quality of life. In this process, each generation of individuals is updated based on their fitness until a treatment plan that meets the patient's individual needs is found.The optimized treatment plan will generate a new weight parameter and treatment plan, reflecting the optimal performance of the treatment plan across multiple target dimensions.

[0065] Through iterative genetic algorithms, not only can the therapeutic advantages of a treatment plan be ensured, but the impact of side effects and quality of life can also be taken into account. By handling multi-objective conflict problems, especially the trade-off between efficacy and toxic side effects, the personalization and precision of treatment plans are greatly improved. Genetic algorithms ensure dynamic adjustment of treatment plans, avoiding the limitations of fixed weights that cannot meet the diverse needs of patients.

[0066] After the optimization results are generated, the updated weight vector is applied to re-execute the complete scoring calculation process on the updated solutions (or the complete candidate solution library). This involves performing a weighted comprehensive score and applying an interactive model for adjustment, resulting in a completely new set of solution scores. Based on these new scores, all involved solutions are sorted in descending order to generate an iterative solution list. This list reflects the new understanding of the optimal treatment plan after one round of feedback optimization.

[0067] Through this iterative optimization process, this invention can effectively improve the clinical applicability of treatment plans, especially in cases of multidimensional treatment goals and complex patient characteristics. The introduction of genetic algorithms solves the problems of the inability to dynamically adjust weights and plan combinations, as well as the unidirectional static nature of the plan recommendation process and the lack of feedback mechanisms. This allows treatment plans to achieve a reasonable balance and optimization across various dimensions, ultimately providing patients with personalized and efficient treatment options.

[0068] S7. Select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment plan.

[0069] In one specific embodiment, the process of performing step S7 may specifically include the following steps: Obtain the comprehensive score data of each solution in the iterated solution list, and filter out the set of high-scoring solutions with comprehensive scores higher than the preset recommendation threshold; For the high-scoring solutions, a final ranking evaluation is conducted, taking into account the treatment synergy of the solutions, the suitability for individual patient characteristics, and the balance of each treatment goal, to generate a final priority ranking. Based on the final priority ranking, the highest-ranked solution is output as the optimal treatment plan.

[0070] Specifically, the comprehensive score data for each treatment regimen is obtained from the iterated list of regimens. These scores reflect the overall performance of each regimen across multiple treatment objectives (such as therapeutic effect, side effects, and quality of life). Regimens are then screened based on a pre-set recommended threshold to ensure that only high-scoring regimens are selected. The recommended threshold is typically based on clinical efficacy standards, individual patient needs, and data from historical success cases, such as 85 points. This screening process effectively eliminates treatment regimens that fail to meet efficacy targets, ensuring that the regimens used in subsequent analyses are those with high therapeutic potential.

[0071] For the set of high-scoring treatment options, further ranking and evaluation will combine the therapeutic synergy of the treatment options, the suitability to individual patient characteristics, and the balance of various treatment goals. In other words, from the set of high-scoring options, a final ranking and evaluation of the treatment options will be conducted, taking into account the interactions between different treatment options, the multi-dimensional balance between the patient's individual needs and treatment goals. The purpose of this process is to personalize the ranking of treatment options according to the patient's specific needs, ensuring that the recommended treatment options are not only excellent in efficacy but also minimize side effects and improve the patient's quality of life. This process will consider factors such as the patient's physiological characteristics, pathological characteristics, and treatment history, combining this information with the therapeutic synergy of the treatment options to ensure that the selected options are most suitable for the patient.

[0072] Therapeutic synergy refers to whether the combined use of different treatment modalities can produce a synergistic effect, thereby improving efficacy or reducing side effects. For example, when certain chemotherapy drugs are used in combination with targeted therapies, they may enhance each other's efficacy and reduce the side effects when used alone. This synergy analysis requires considering the mechanisms of action of the drugs used in each treatment regimen and their interactions. For example, the therapeutic synergy score is obtained by querying a pre-built drug synergy knowledge base, which assigns a synergy level to each pair of common anti-tumor drugs. For regimen sn, the synergy level scores between all pairs of drugs included in the regimen are calculated, averaged, and normalized to obtain the therapeutic synergy score.

[0073] Patient-specific fit refers to whether a treatment plan meets the specific needs of a patient, such as their gene mutation status, immune response level, and tolerability. These characteristics are crucial for determining the optimal treatment plan. For example, some immunotherapies may be more effective for lung cancer patients with high PD-L1 expression, while targeted therapy may be more effective for patients with EGFR mutations. Assessing patient-specific fit requires combining detailed pathological data, gene mutation data, and other biomarker data to ensure that the selected treatment plan can be personalized to the patient's specific pathological background. Specifically, the matching metric between patient characteristics and treatment plan is quantified during calculation, typically using similarity calculation methods such as Euclidean distance or cosine similarity to assess the degree of fit between patient characteristics and treatment plan, ensuring that the treatment plan best meets the patient's individual needs.

[0074] The balance of treatment goals involves a comprehensive evaluation of treatment options across multiple dimensions, ensuring a reasonable balance between objectives such as treatment effectiveness, side effects, quality of life, and cost. For example, some treatment options may score highly in terms of efficacy but have excessive side effects, leading to a significant decline in the patient's quality of life. In such cases, the treatment options will be comprehensively ranked based on their overall balance, ensuring that the final recommended treatment provides effective treatment while minimizing side effects and improving the patient's quality of life. For instance, based on the specific score distribution of each option across the four dimensions of efficacy, side effects, quality of life, and cost, a balance index is calculated. This includes assessing whether any dimension scores too low to be a significant weakness, or evaluating the Gini coefficient of the four scores to measure the degree of balance.

[0075] The three evaluation results are weighted and summed according to preset weights to obtain the final score of each treatment plan, and then sorted in descending order to generate the final priority ranking list.

[0076] Preferably, the final ranking can also rely on a decision function that integrates the evaluation results of the three dimensions mentioned above. This function calculates a final priority value for each high-scoring scheme. In this case, the calculation of the final priority value is not a simple weighting, but may employ a rule-based scoring card or a lightweight discriminant model. For example, for a scheme with extremely high treatment synergy and perfect patient micro-fit, even if its overall score is slightly lower, it may obtain a higher final priority value through the decision function, thus improving its ranking; conversely, for a scheme with clear micro-contraindications or a severe imbalance in treatment goals, even if its overall score is the highest, its final priority value may be lowered. The construction of the decision function is based on the logic of clinical decision-making pathways and expert experience, and its parameters are configurable. By calculating the final priority value of each scheme in the set of high-scoring schemes and arranging them in descending order, a final priority ranking list is generated. This process combines the numerical scores calculated by the machine with the complexity and quantitative factors that must be considered in clinical decision-making.

[0077] The final priority ranking determines the priority of different treatment options, and the highest-ranked treatment option is selected as the optimal treatment option. This optimal option will provide the best treatment strategy based on the patient's specific condition and treatment goals. In this way, the system can dynamically and personally adjust the treatment plan to ensure that each patient receives the most suitable treatment.

[0078] This optimization ranking method, based on a balance of comprehensive scores and multidimensional treatment goals, effectively addresses the challenge of recommending personalized treatment plans. Especially when facing complex situations involving multi-objective optimization and diverse patient characteristics, it ensures that recommended treatment plans meet clinical efficacy standards while minimizing side effects and improving patients' quality of life. Through this process, treatment selection becomes more precise, meeting individualized patient needs and thus effectively improving treatment success rates and patients' quality of life.

[0079] The above describes a method for recommending a personalized treatment plan for lung cancer in the embodiments of this application. The following describes a system for recommending a personalized treatment plan for lung cancer in the embodiments of this application. Please refer to [link / reference]. Figure 5 This application provides a schematic diagram of a personalized treatment plan recommendation system for lung cancer, which includes: The feature extraction module 10 is used to obtain multidimensional feature data from the patient's clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set.

[0080] The protocol evaluation module 20 is used to set multidimensional treatment goals, assign weights to each goal according to a simplified feature set, construct a protocol evaluation matrix and optimize it, and generate a preliminary protocol ranking list.

[0081] The scheme filtering module 30 is used to filter the schemes in the preliminary scheme ranking list by a preset score threshold, obtain a list of candidate optimal schemes, and verify the matching degree of the candidate optimal scheme list to obtain a set of verified schemes.

[0082] The classification optimization module 40 is used to extract classification feature data from the validated scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes.

[0083] The interactive analysis module 50 is used to construct an interactive model based on the subset of classification optimization schemes to analyze the interactive impact of toxic side effects and quality of life on treatment effects. The interactive model is used to calculate the interactive impact data, and then the scores of each scheme are adjusted to obtain the adjusted scheme scores.

[0084] The iterative optimization module 60 is used to generate an iterative list of solutions if the score of the adjusted solution is lower than the preset score threshold.

[0085] The optimal decision module 70 is used to select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment plan.

[0086] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the method for recommending a personalized treatment plan for lung cancer.

[0087] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for recommending personalized treatment plans for lung cancer, characterized in that, The method includes: S1. Obtain multidimensional feature data from the patient's clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set; S2. Set multidimensional treatment goals, assign weights to each goal according to the simplified feature set, construct a scheme evaluation matrix and optimize it to generate a preliminary scheme ranking list; S3. Filter the schemes in the preliminary scheme ranking list by a preset score threshold to obtain a candidate optimal scheme list, and verify the matching degree of the candidate optimal scheme list to obtain a verified scheme set. S4. Extract classification feature data from the verified scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes; S5. Based on the subset of the classification optimization schemes, construct an interaction model for analyzing the interaction between toxic side effects and quality of life on treatment effects. Calculate the interaction impact data using the interaction model, and then adjust the scores of each scheme to obtain the adjusted scheme scores. S6. If the adjusted solution score is lower than the preset score threshold, an iterative solution list is generated through iterative optimization. S7. Select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment solution.

2. The method according to claim 1, characterized in that, S1 includes: The multidimensional feature data are obtained from the patient's clinical and molecular indicators, and the multidimensional feature data includes at least tumor size, lymph node metastasis status, gene mutation type, PD-L1 expression level and driver gene status. Using patient samples as rows and data dimensions as columns, an initial data matrix is ​​constructed. After standardizing the initial data matrix, the multidimensional data matrix is ​​obtained. Principal component analysis is used to reduce the dimensionality of the multidimensional data matrix, and the top k eigenvectors whose cumulative variance contribution rate reaches a preset contribution threshold are selected as the principal feature components. The multidimensional data matrix is ​​projected onto the feature space spanned by the main feature components to generate the simplified feature set.

3. The method according to claim 1, characterized in that, S2 include: Set multidimensional treatment goals, which include at least treatment efficacy, side effects, cost burden, and quality of life; Based on the feature components related to each target in the simplified feature set, an initial weight parameter is determined for each treatment target to form an initial weight vector; Obtain a set of candidate treatment options and determine the expected score of each option on the multidimensional treatment goal, and construct an evaluation matrix based on the expected score; Starting with the initial weight vector, a genetic algorithm is used to iteratively optimize the weight vector. During the optimization process, the weight vector being evaluated is weighted and calculated with the scheme evaluation matrix to obtain the comprehensive score of each scheme. The comprehensive score is used as the basis for evaluating the fitness of the weight vector until the optimal weight vector that meets the convergence condition is obtained. The evaluation matrix of the proposed schemes is weighted according to the optimal weight vector to obtain the final comprehensive score of each scheme and generate a preliminary sorted list of schemes.

4. The method according to claim 1, characterized in that, S3 include: Obtain the comprehensive score data of each scheme in the preliminary scheme ranking list, filter out the schemes with comprehensive scores higher than the preset score threshold, and form a candidate optimal scheme list; From the simplified feature set, extract patient feature vectors related to individual patient characteristics. The patient feature vectors include at least tumor size, lymph node metastasis status, gene mutation type, PD-L1 expression level, and driver gene status. For each candidate optimal solution, the matching degree is verified based on the treatment attribute vector of the solution and the patient feature vector. The treatment attribute vector represents the treatment strategy and intensity of the treatment solution for each feature in the patient feature vector. Generate a set of validated solutions based on the validation results.

5. The method according to claim 1, characterized in that, S4 includes: Classification feature data is extracted from the molecular indicators associated with the validated scheme set, and the classification feature data includes at least gene mutation status and driver gene status; The schemes in the verified scheme set are grouped according to the classification feature data to obtain the classification results; Based on preset screening criteria, a subset of classification optimization schemes is determined according to the classification results. The preset screening criteria include at least the degree of matching with patient characteristics, comprehensive evaluation of treatment goals, and / or clinical efficacy.

6. The method according to claim 1, characterized in that, S5 include: Quantitative score data for each classification optimization scheme in terms of toxic side effects and quality of life were extracted to obtain interaction impact data. Construct an interactive model to analyze the interaction between toxic side effects and quality of life on treatment efficacy; Input the interaction impact data into the interaction model to obtain the interaction impact results; Based on the interaction effect results, the scores of each scheme in the subset of classification optimization schemes are adjusted to obtain the adjusted scheme scores.

7. The method according to claim 1, characterized in that, S6 include: The adjusted solution score is obtained and compared with a preset score threshold. If the highest solution score is lower than the preset score threshold, an iterative optimization process is triggered. Using the weight parameters of the current multidimensional treatment goals and / or candidate treatment plans as optimization variables, a genetic algorithm is used for re-optimization to adjust the weight parameters and plan combinations. Based on the optimization results, updated weight parameters and / or treatment plans are generated, and the comprehensive score of all relevant candidate treatment plans is recalculated to generate the iterative list of plans.

8. The method according to claim 1, characterized in that, S7 includes: Obtain the comprehensive score data of each scheme in the iterated scheme list, and filter out the set of high-scoring schemes with comprehensive scores higher than the preset recommendation threshold; For the solutions in the set of high-scoring solutions, a final ranking evaluation is performed based on the treatment synergy of the solutions, the suitability for individual patient characteristics, and the balance of each treatment goal, and a final priority ranking is generated. Based on the final priority ranking, the highest-ranked solution is output as the optimal treatment plan.

9. A personalized treatment plan recommendation system for lung cancer, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The feature extraction module is used to obtain multidimensional feature data from patients' clinical and molecular indicators, construct a multidimensional data matrix, and then perform dimensionality reduction processing to generate a simplified feature set. The treatment plan evaluation module is used to set multidimensional treatment goals, assign weights to each goal according to the simplified feature set, construct and optimize the treatment plan evaluation matrix, and generate a preliminary treatment plan ranking list. The scheme filtering module is used to filter the schemes in the preliminary scheme ranking list by a preset score threshold, obtain a list of candidate optimal schemes, and verify the matching degree of the list of candidate optimal schemes to obtain a set of verified schemes. The classification optimization module is used to extract classification feature data from the verified scheme set, perform classification processing based on the classification feature data, and then generate a subset of classification optimization schemes. The interactive analysis module is used to construct an interactive model for analyzing the interactive impact of toxic side effects and quality of life on treatment effects based on the subset of the classification optimization schemes. The interactive model is used to calculate the interactive impact data, and then the scores of each scheme are adjusted to obtain the adjusted scheme scores. The iterative optimization module is used to generate an iterative list of solutions if the adjusted solution score is lower than a preset score threshold. The optimal decision module is used to select high-scoring solutions from the iterated solution list, sort the high-scoring solutions, and obtain the optimal treatment plan.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement a method for recommending a personalized treatment plan for lung cancer as described in any one of claims 1 to 8.

Images