A method, system, device, and medium for feature extraction from medical clinical data.

Abstract

This invention discloses a method, system, device, and medium for feature extraction from medical clinical data, relating to the field of data processing technology. It includes: acquiring a medical clinical dataset; replacing the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, removing the cryptic effect of the moss optimization algorithm, and introducing a bimodal propagation search to obtain a bio-learning moss optimization algorithm; using the bio-learning moss optimization algorithm to find the subset of clinical features most relevant to sample classification during the feature extraction process of the medical clinical dataset; and using a K-nearest neighbor classifier to classify the subset of clinical features most relevant to sample classification, thereby obtaining the optimal subset of clinical features for the medical clinical data. This invention overcomes the deficiency of the moss optimization algorithm in its insufficient ability to exploit local optimal solutions, effectively improving the algorithm's convergence accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, device, and medium for feature extraction from medical clinical data. Background Technology

[0002] Lung cancer is the leading cause of cancer-related death worldwide, with an overall five-year survival rate of less than 20%. Based on histopathological characteristics, lung cancer is mainly divided into non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), with NSCLC accounting for approximately 80%–85% of lung cancer cases. NSCLC includes three main subtypes: squamous cell carcinoma, adenocarcinoma, and large cell carcinoma. Current clinical treatment strategies encompass a variety of modalities, including surgical resection, radiotherapy, chemotherapy, molecular targeted therapy, and immunotherapy.

[0003] In recent years, breakthroughs in molecularly targeted drugs, immune checkpoint inhibitors (ICIs), and anti-angiogenic agents have significantly improved the prognosis of NSCLC patients. For patients with advanced NSCLC, the choice of first-line treatment is mainly based on the driver gene mutation status and the expression level of programmed death-ligand 1 (PD-L1).

[0004] In patients without targeted driver gene mutations, ICIs, represented by PD-1 / PD-L1 inhibitors, have become the core treatment approach. To screen for the optimal combination of biomarkers that can accurately predict the efficacy and prognosis of immunochemotherapy in non-small cell lung cancer (NSCLC) patients, the Moss Growth Optimization (MGO) algorithm can be used. The implementation process is as follows: Clinical data obtained from NSCLC patients is used as input, where each feature is considered a dimension in the search space. The algorithm encodes each possible combination of biomarkers (i.e., a subset of features) as a "moss patch," forming an initial population. Subsequently, the quality of each "moss patch" is evaluated using a pre-defined fitness function. This function typically combines the classification accuracy (or AUC value) of the prediction model built using the current feature subset with a penalty term for the number of selected features, achieving a balance between high accuracy and model simplicity. During the iterative optimization process, the moss optimization algorithm simulates the natural growth and reproduction strategies of moss, primarily using mechanisms such as spore dispersal for large-scale global exploration to quickly locate "suitable areas" with high predictive potential within a broad feature combination space. As iterations proceed, the entire population migrates and concentrates towards the region where the current optimal solution is located based on its fitness value, gradually converging. This process continues until a preset termination condition is reached, ultimately outputting the feature subset corresponding to the "moss patch" with the highest fitness, serving as the biological basis for constructing the final prediction model.

[0005] However, existing moss optimization algorithms lack the ability to exploit local solutions in the later stages of the search, resulting in limited convergence accuracy and a tendency for the results to deviate from the global optimum or for the final solution to be inaccurate. Summary of the Invention

[0006] Therefore, it is necessary to provide a method, system, device, and medium for feature extraction of medical clinical data to address the aforementioned technical problems.

[0007] This invention provides a method for feature extraction from medical clinical data, comprising:

[0008] Obtain medical clinical datasets, which include blood test data and clinical datasets;

[0009] By replacing the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, removing the cryptic effect of the moss optimization algorithm, and introducing bimodal propagation search, a bio-learning moss optimization algorithm is obtained. The bio-learning moss optimization algorithm is used to find the subset of clinical features most relevant to sample classification during the feature extraction process of medical clinical datasets.

[0010] The K-nearest neighbor classifier is used to classify the clinical feature subset that is most relevant to the sample classification, thus obtaining the optimal clinical feature subset of medical clinical data.

[0011] Specifically, the bio-learning moss optimization algorithm is used to find the subset of clinical features most relevant to sample classification during feature extraction from medical clinical datasets. This includes:

[0012] Initialize a moss population, with each moss individual representing candidate medical clinical data and possessing an independent feature value, resulting in a feature value set; perform feature selection on the feature value set to obtain the optimal feature subset;

[0013] A biogeographical learning strategy is used to construct an example vector set corresponding to the optimal feature subset. The particles in the example vector set are sorted according to the historical best solution of each moss individual to obtain the sorting value of each particle. Based on the sorting value of each particle, the immigration rate and immigration rate are assigned to the moss individuals corresponding to the optimal feature subset. A two-mode propagation search is performed based on the immigration rate and immigration rate to dynamically update the position of the moss individuals and obtain the global optimal solution, which is the clinical feature subset most relevant to the sample classification.

[0014] Optionally, medical clinical datasets can be obtained through electronic medical records. These datasets specifically include demographic characteristics, disease characteristics, pre-treatment hematological indicators, immune characteristics, and efficacy outcome indicators.

[0015] The efficacy outcome measures include: efficacy response and survival endpoint; efficacy response includes: disease remission group and disease non-remission group; survival endpoint includes: long survival group and short survival group.

[0016] Optionally, feature selection is performed on the feature set to obtain the optimal feature subset, specifically including:

[0017] Each moss individual is binary encoded and the feature set is encoded using a transformation function. Feature selection is performed based on the position of each moss individual's feature value in the feature set to obtain multiple feature subsets.

[0018] For each feature subset, determine the weight of that feature subset in the feature extraction process, obtain the fitness value of that feature subset, and take the feature subset with a fitness value greater than a set threshold as the optimal feature subset.

[0019] Optionally, the binary encoded values ​​of the initial moss population are obtained by encoding the feature set using a transformation function based on the following formula:

[0020] ;

[0021] ;

[0022] in, For the first The first iteration generated The first individual A binary encoded value, For the first The first iteration generated The first individual 1 eigenvalue, Given a random number in the range [0, 1]. To negate the sign, For conversion functions, This represents the initial moss population;

[0023] The weight of the feature subset in the feature extraction process is determined based on the following formula, and the fitness value of the feature subset is obtained:

[0024] ;

[0025] ;

[0026] in, For the first The fitness value of an individual moss. For classification error rate, The length of the feature subset. For the total length of the feature, The weights for the classification error rate, Weights for selecting the number of feature subsets. The number of correctly classified samples. This represents the number of samples that were misclassified.

[0027] Optionally, a biogeographical learning strategy can be used to construct a set of example vectors corresponding to the optimal feature subset:

[0028] = ];

[0029] ;

[0030] Based on the following formula, the particles in the example vector set are sorted according to the historical best solution of each moss individual to obtain the sorting value of each particle:

[0031] = N- 1 , = N- 2 , … = 0 ;

[0032] Based on the following formula, the immigration and emigration rates are assigned to the moss individuals corresponding to the optimal feature subset according to the ranking value of each particle:

[0033] ;

[0034] in, This is the particle's own historical optimal solution. sN For particle subscripts, For the immigration rate, For emigration rate, N The sorting value of the particles. rank For sorting functions, For example vectors, For example vector set, i For indexing.

[0035] Optionally, a dual-mode propagation search is performed based on immigration and emigration rates to dynamically update the location of moss individuals, specifically including:

[0036] Moss individuals with the best historical solution have the lowest migration rate and tend to update their position to themselves.

[0037] The moss individual with the worst historical best solution has the highest migration rate and tends to update its position to other moss individuals.

[0038] The particle with the best historical optimal solution has the highest migration rate and is used as the renewal direction by other moss individuals.

[0039] The particle with the worst historical best solution has the lowest migration rate and is not used as the renewal direction by other moss individuals.

[0040] This invention provides a feature extraction system for medical clinical data, comprising:

[0041] The data acquisition module is used to acquire medical clinical datasets, which include blood test data and clinical datasets.

[0042] The processing module is used to replace the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, remove the cryptic effect of the moss optimization algorithm, and introduce bimodal propagation search to obtain a biolearning moss optimization algorithm; the biolearning moss optimization algorithm is used to find the subset of clinical features most relevant to the sample classification in the feature extraction process of medical clinical datasets;

[0043] The classification module is used to classify the clinical feature subset that is most relevant to the sample classification using a K-nearest neighbor classifier, thereby obtaining the optimal clinical feature subset of medical clinical data.

[0044] The processing module includes a feature selection module and an optimization module.

[0045] The feature selection module is used to initialize the moss population. Each moss individual represents candidate medical clinical data and has an independent feature value, resulting in a feature value set. Feature selection is performed on the feature value set to obtain the optimal feature subset.

[0046] The optimization module is used to construct an example vector set corresponding to the optimal feature subset using a biogeographical learning strategy. The particles in the example vector set are sorted according to the historical best solution of each moss individual to obtain the sorting value of each particle. Based on the sorting value of each particle, the influx rate and emigration rate are assigned to the moss individuals corresponding to the optimal feature subset. Based on the influx rate and emigration rate, a two-mode propagation search is performed to dynamically update the position of the moss individuals and obtain the global optimal solution, which is the clinical feature subset most relevant to the sample classification.

[0047] This invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a feature extraction method for medical clinical data.

[0048] This invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of a feature extraction method for medical clinical data.

[0049] The method, system, device, and medium for feature extraction of medical clinical data provided in this invention have the following advantages compared with the prior art:

[0050] This invention enhances the diversity and flexibility of the moss optimization algorithm in the search process through a biogeographical learning strategy. The use of bimodal propagation search in the later stages of the search allows for a more detailed exploration of the solution space, screening and extracting the clinical features most relevant to the sample classification. This can compensate for the shortcomings of existing moss optimization algorithms in their ability to exploit local optimal solutions and effectively improve the algorithm's convergence accuracy.

[0051] In addition, this invention utilizes the K-nearest neighbor classifier to classify the clinical feature subset most relevant to sample classification, further avoiding deviations from the global optimum, thereby improving the accuracy of the optimal clinical feature subset and providing a more efficient and accurate method for the analysis and processing of medical clinical data. Attached Figure Description

[0052] Figure 1 A flowchart of a feature extraction method for medical clinical data provided in one embodiment;

[0053] Figure 2A graph showing the function results of BLMGO with different parameters for a feature extraction method for medical clinical data provided in one embodiment;

[0054] Figure 3 The image shows the feature selection count results of bBLMGO, a feature extraction method for medical clinical data provided in one embodiment, on a dataset of 241 patients (binary classification) with laboratory data from patients with PFS less than 12 months and greater than 12 months.

[0055] Figure 4 The image shows the result of feature selection counts for bBLMGO, a feature extraction method for medical clinical data provided in one embodiment, on a dataset of 289 patients (binary classification) with clinical efficacy data (PD, SD, PR, CR).

[0056] Figure 5 The flowchart shows a bio-learning moss optimization algorithm for a feature extraction method for medical clinical data provided in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] Immune checkpoint inhibitors (ICIs), such as anti-programmed death-1 (PD-1) / programmed death-ligand 1 (PD-L1) or anti-cytotoxic T-lymphocyte-associated antigen 4 (CTLA-4) drugs, can induce durable responses in subgroups of patients with advanced cancer. ICIs combined with platinum-based chemotherapy regimens have demonstrated clear survival benefits in multiple randomized controlled trials (RCTs), providing a new first-line treatment strategy for metastatic NSCLC. Notably, this survival benefit was observed in patients with different PD-L1 expression levels, and the overall safety profile was manageable.

[0059] However, it should be noted that although ICI combined with chemotherapy significantly improves the objective response rate and median survival, this treatment only provides long-term benefit to a small percentage of patients. Treatment resistance and disease progression remain major clinical challenges. The vast majority of patients do not achieve remission or experience disease progression during treatment. Furthermore, immune checkpoint inhibitor therapy can lead to immune-related adverse events.

[0060] The significant heterogeneity of immunotherapy responses underscores the importance of developing predictive biomarkers, which can help screen potential beneficiaries and achieve precision medicine. Current biomarker systems can be divided into four main categories: (1) intrinsic tumor markers, including PD-L1 expression, tumor mutational burden (TMB), and specific driver gene variants; (2) tumor microenvironment markers, covering the composition of tumor-associated immune cells (TAICs) and T cell receptor (TCR) diversity; (3) liquid biopsy markers, such as circulating tumor DNA (ctDNA) and peripheral blood immune cell subsets; and (4) host-related factors, involving human leukocyte antigen class I (HLA-I) polymorphism, clinical features, and metabolomics characteristics.

[0061] Although the FDA has approved PD-L1 expression levels and TMB as predictive biomarkers for ICI efficacy, their clinical application still has significant limitations and cannot fully reflect the complex interaction between tumors and the immune system. PD-L1 lacks standardized staining interpretation criteria, and its expression differs greatly between primary and metastatic lesions; TMB requires sufficient tumor tissue and resources to sequence DNA, and more importantly, these indicators alone cannot accurately predict durable remission and survival.

[0062] Against this backdrop, peripheral blood biomarkers have garnered significant attention due to their accessibility and dynamic monitoring advantages. Several indicators and ratios of the complete blood count (CBC) have been suggested as biomarkers for predicting ICI efficacy and patient outcomes. Among blood biomarkers, a low neutrophil-lymphocyte ratio (NLR < 5), low systemic immune inflammation index (SII), and decreased C-reactive protein (CRP) are associated with better prognosis, while high lactate dehydrogenase (LDH) and elevated baseline CRP predict poor outcomes. For example, in a study of patients with advanced NSCLC, Diem et al. and Ren et al. reported that high pre-treatment neutrophil-lymphocyte ratio (NLR) and platelet-lymphocyte ratio (PLR) were prognostic biomarkers significantly associated with poor survival and lower remission rates in patients receiving nivolumab monotherapy.

[0063] The current limitations in the predictive efficacy of single biomarkers highlight the urgent need to develop multidimensional prognostic models. Multidimensional integrated models can improve the accuracy of clinical decision-making, enabling physicians to identify high-response patients and those at risk of rapid progression in immunotherapy (ICI). Artificial intelligence (AI) technology offers a breakthrough path to this goal. AI / ML models can accurately predict immunotherapy response, progression-free survival, and overall survival in NSCLC patients, tailoring treatment decisions to each patient's unique molecular, clinical, and imaging characteristics.

[0064] The existing defects include, but are not limited to:

[0065] (1) Limited convergence speed: In high-dimensional complex or multi-peak optimization problems, although the MGO population has strong expansion ability in the global scope, its convergence speed is slow, resulting in low efficiency of the algorithm in solving practical engineering problems.

[0066] (2) Limited convergence accuracy: Existing MGOs have insufficient local development capabilities for the optimal solution in the later stages of the search, which can easily lead to the results deviating from the global optimum or the accuracy of the final solution being low.

[0067] (3) The robustness of escaping local optima needs to be improved: Although a dormancy mechanism is set, individuals may still fall into new local optima in more complex or dynamically changing search spaces, making it difficult to guarantee the attainment of the global optimum.

[0068] This invention provides a method for feature extraction from medical clinical data, such as... Figure 1 As shown. The method includes:

[0069] Obtain medical clinical datasets, which include blood test data and clinical datasets.

[0070] By replacing the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, removing the cryptic effect of the moss optimization algorithm, and introducing bimodal propagation search, a biolearning moss optimization algorithm is obtained. This biolearning moss optimization algorithm is then used to find the subset of clinical features most relevant to sample classification during feature extraction from medical clinical datasets.

[0071] The K-nearest neighbor classifier is used to classify the clinical feature subset that is most relevant to the sample classification, thus obtaining the optimal clinical feature subset of medical clinical data.

[0072] Specifically, the bio-learning moss optimization algorithm is used to find the subset of clinical features most relevant to sample classification during the feature extraction process of medical clinical datasets. This includes: initializing a moss population, where each moss individual represents candidate medical clinical data and has an independent feature value, resulting in a feature value set; and performing feature selection on the feature value set to obtain the optimal feature subset.

[0073] A biogeographical learning strategy is employed to construct an example vector set corresponding to the optimal feature subset. The particles in the example vector set are then sorted according to the historical best solutions of individual moss individuals, yielding a ranking value for each particle. Immigration and emigration rates are assigned to the moss individuals corresponding to the optimal feature subset based on these ranking values. A bimodal propagation search is then performed based on the immigration and emigration rates to dynamically update the positions of the moss individuals, ultimately yielding the global optimum, which is the clinical feature subset most relevant to the sample classification.

[0074] According to another preferred embodiment of the present invention, for clinical data of non-small cell lung cancer patients, the present invention employs a moss optimization method improved by a biological learning strategy for feature extraction. These feature extraction methods can capture key features from clinical data, providing important features for subsequent analysis and classification. When processing routine blood tests and clinical data of non-small cell lung cancer patients, challenges such as limited sample size, high feature dimensionality, and correlations between data are often encountered. To overcome these problems, a 10-fold cross-validation method is used to verify the reliability and stability of their analytical results. Figure 5 As shown, the dataset is divided into 10 equal parts, with one part used as the test set and the other nine parts as the training set. This process is repeated 10 times, each time selecting a different training and test set. The final average of all results is used as the final performance evaluation metric. This approach ensures consistent and consistent training and testing, reducing errors caused by improper data partitioning and improving the accuracy and reliability of model evaluation.

[0075] Step S1: Extract the training sample set from the above clinical dataset of non-small cell lung cancer patients, and initialize the training sample population. ,in, The number of individuals in the training sample. The number of dimensions for each training sample. Indicates the first The population obtained in the next iteration Indicates in In the nth iteration The j-th feature value of an individual This indicates the current iteration number, with a value range of [0, 1000].

[0076] Step S2: Design a classifier implementation method and optimization technique based on k-Nearest Neighbor (KNN) according to the training sample set, and perform classification.

[0077] The basic idea behind KNN is a simple, parameter-free, and efficient learning method that has shown excellent performance in problems such as function classification and approximation, achieving high completion rates and classification accuracy. Recent research has also demonstrated superior training speed and classification accuracy; therefore, KNN was used as a classifier for experimental evaluation. KNN is an instance-based learning model that predicts the class of a new sample by majority voting on the k nearest neighbor classes of the test sample. In classification decisions, the class of the new sample is determined based on the minimum distance (based on a similarity metric) between it and all samples in the training set. The most widely used similarity metric is Euclidean distance. For two D-dimensional points... and The Euclidean distance is calculated as follows:

[0078] ;

[0079] in, D-dimensional points and The distance. Representing D-dimensional points and The coordinate values.

[0080] Specifically, it includes:

[0081] Step S1: Collect blood test data and clinical datasets from patients with non-small cell lung cancer, and extract them through electronic medical records.

[0082] Step S2: Divide the population into training and testing sets, initialize the population, encode each moss individual in the current population into binary form, and encode the features using a transformation function.

[0083] Step S3: Calculate the fitness value of the current population using the moss optimization algorithm.

[0084] Step S4: Improve the moss reproduction mechanism, remove the cryptogenic (dormant) effect, introduce a dual-mode propagation search, and update the location of individual populations.

[0085] Step S5: Test the relevant parameters of the biological learning strategy, select the optimal biological learning moss optimization algorithm, and dynamically adjust the individual moss.

[0086] Step S6: Optimize the moss algorithm using the optimal biological learning strategy and construct a prediction model with the KNN classifier.

[0087] Step S7: Update the loop to the maximum number of iterations and output the optimal feature subset of routine blood data and clinical data.

[0088] The implementation is as follows:

[0089] Step S1: Collect data on the efficacy of immunochemotherapy in non-small cell lung cancer (NSCLC) patients. Initial screening of 1799 NSCLC patients admitted up to July 2024 was conducted using a hospital's electronic medical record system. After strict inclusion and exclusion criteria, 289 patients with locally advanced or metastatic NSCLC who received first-line PD-L1 immune checkpoint inhibitors (ICIs) combined with platinum-based chemotherapy were ultimately included. Eligibility criteria included age ≥18 years, histological / cytological diagnosis, completion of baseline hematological testing within 30 days prior to treatment, and at least one RECIST 1.1 standard efficacy assessment. Exclusion criteria included prior targeted therapy or radiotherapy, exclusion of cases receiving other anti-tumor regimens during treatment (e.g., surgery, radiotherapy), missing key clinical data (e.g., staging, efficacy records), or loss to follow-up where complete follow-up data could not be obtained.

[0090] Extracted from electronic medical records: ① Demographic characteristics (sex, age, BMI, smoking history and comorbidities); ② Disease characteristics (TNM stage, pathological type, PD-L1 TPS%); ③ Hematological parameters 30 days before treatment (including complete blood count, liver and kidney function, inflammatory markers, coagulation function and tumor markers); ④ Immune characteristics (lymphocyte subsets, immunoglobulins and complement); ⑤ Efficacy outcome indicators: The study established dual outcome indicators: efficacy response was divided into disease remission group (complete remission [CR] + partial remission [PR]) and disease non-remission group (stable disease [SD] + disease progression [PD]) based on RECIST 1.1 criteria; the survival endpoint was divided into long survival group (12 months) and short survival group based on progression-free survival (PFS, from treatment initiation to progression / death).

[0091] For continuously variable data that conform to a normal distribution, the Kolmogorov-Smirnov test was used for data analysis, expressed as mean ± standard deviation (X ± S), and the u test was used for comparisons between groups. For continuously variable data that are not normally distributed, nonparametric tests were used, described as median and interquartile range (IQR), and count data were expressed as percentages (%), with chi-square tests or Fisher's exact test used for comparisons between groups. Statistical analysis results are shown in Tables 1 and 2. Let the initial population of the training samples be... ,in, , The number of individuals in the training sample. It is the dimension of the training samples. In the first The population obtained in the next iteration Indicates in In the nth iteration The first individual j 1 eigenvalue, This indicates the current iteration number, with a value range of [0, 1000].

[0092] Table 1. Baseline characteristics of locally advanced / metastatic NSCLC patients treated with first-line immunotherapy combined with chemotherapy: comparison between the disease response group (CR+PR) and the non-response group (SD+PD).

[0093]

[0094] Table 2. Baseline characteristics of locally advanced / metastatic NSCLC patients treated with first-line immunotherapy combined with chemotherapy: comparison of PFS <12 months and PFS >12 months.

[0095]

[0096] Figure 3 The image shows the number of feature selections for bBLMGO, a feature extraction method for medical clinical data provided in one embodiment, on a dataset of 241 patients (binary classification) with laboratory data from patients with pre-existing conditions (PFS) less than 12 months and greater than 12 months.

[0097] Figure 4 The image shows the result of feature selection counts for bBLMGO, a feature extraction method for medical clinical data provided in one embodiment, on a dataset of 289 patients (binary classification) with clinical efficacy data for PD, SD, PR, and CR.

[0098] Step S2: Divide the dataset into training and testing sets, and initialize the population. Each feature value of each individual in the model is simulated as a binary coded value using formulas (1) and (2).

[0099] In step S2, the population Each feature value of each individual in the model is simulated as a binary encoded value using formulas (1) and (2):

[0100] (1)

[0101] (2)

[0102] in, It is the first The first iteration generated The first individual A binary encoded value, For the first The first iteration generated The first individual 1 eigenvalue, This represents a random number in the range [0, 1]. To negate the sign, For conversion functions, moss population .

[0103] Step S3: Calculate the moss population using formulas (3) and (4). Fitness value:

[0104] (3)

[0105] (4)

[0106] in, For the first The fitness value of an individual moss. For classification error rate, Indicates the length of the feature subset. It is the total length of the feature. It is the weight of the classification error rate. Weights for selecting the number of feature subsets. It is the number of correctly classified samples. It represents the number of misclassified samples.

[0107] The following explanations are provided for steps S2 and S3:

[0108] (1) The role and status of binary encoded values.

[0109] The binary encoded values ​​serve to transform the natural feature space (i.e., continuous feature values) into a discrete feature selection representation. Through the encoding in step S2, the position of each individual (i.e., the moss individual) in the original feature space is expressed using a binary string of 0s and 1s, where each 0 or 1 corresponds to whether a certain feature is selected. This encoding method can conveniently represent the feature subset selection problem and lays the foundation for subsequent evolutionary operations (such as crossover, mutation, etc.) and fitness evaluation.

[0110] In short, binary encoded values ​​are the basis of feature selection schemes (i.e., individual representations), which give each individual a distinct subset of features.

[0111] (2) The role and status of fitness value.

[0112] The fitness value is an evaluation of each individual given a subset of features. The fitness function typically considers factors such as the classification accuracy and the number of features in the selected feature subset, reflecting the quality of that subset and serving as a key indicator for population evolution. In step S3, the fitness value, combined with the classification error rate and the feature subset length, serves as the basis for the evolutionary algorithm's selection and update operations.

[0113] In short, the fitness value is a performance evaluation of binary encoding (feature selection), and it is the goal and driving force for algorithm evolution and optimization.

[0114] (3) The logical connection between the two

[0115] A. Each moss individual is first mapped to a binary code value by its feature value through formula (1) and formula (2), thus obtaining an individual representation (i.e. a feature subset) that represents the feature selection scheme.

[0116] B. After obtaining the binary code of each individual, the performance of each individual in the classification task is calculated based on these selected feature subsets according to formulas (3) and (4) to obtain the fitness value.

[0117] C. Binary encoded values ​​(feature selection representation) are an indispensable representation for searching the optimal feature subset using the moss optimization algorithm; without them, effective feature subset search and combination are impossible. Fitness values ​​are the core of evaluating and driving population iterative evolution; without fitness values, feature selection results cannot be judged and optimized. Therefore, binary encoded values ​​are a prerequisite for fitness value calculation.

[0118] Step S4: Improve the moss reproduction mechanism, remove the cryptogenic (dormant) mechanism, introduce a dual-mode propagation search, and update the location of individual populations.

[0119] Step S5: Test the relevant parameters of the biological learning strategy, select the optimal biological learning moss optimization algorithm, and dynamically adjust the individual moss. The core idea of ​​the Biogeography-Based Learning Strategy (BLS) is to generate example vector indexes through biogeography-based optimization (BBO) algorithm. = According to the moss individual's own historical optimal solution ( The mass of all particles is sorted, and taking the minimization problem as an example, the steps for generating the example in BLS are described as follows:

[0120] ;

[0121] in, s 1 is the best The subscript of the moss individual, s 2 is the second best The subscript of the moss individual, sN It has the worst The subscript of the particle.

[0122] Then, the particle sorting values ​​are assigned as follows:

[0123] = N- 1 , = N- 2 , … = 0 (5)

[0124] According to formula 5, s 1 Best The highest ranking value of moss individuals, Having the worst The particles that receive the lowest ranking value.

[0125] Next, we assign immigration and immigration rates to all particles. Taking the linear migration model as an example, the immigration and immigration rates of all particles can be calculated using the following formula (6):

[0126] (6)

[0127] According to formula 6, Possess the best The solution will have the lowest migration rate. and the highest emigration rate ; And possessing the worst solution It will have the highest immigration rate and the lowest emigration rate .

[0128] Finally, biologically based particles i As can be seen from the example generation method:

[0129] (1) Possess the best moss individuals Due to its immigration rate The lowest, therefore they tend to learn in their own direction;

[0130] (2) Possessing the worst moss individuals Due to its immigration rate The highest, therefore more inclined to other particles study;

[0131] (3) Possess the best particles Its emigration rate is the highest, therefore its It is easier for other particles to learn from it;

[0132] (4) Possessing the worst The particles with the lowest migration rate, therefore their It is not easily learned by other particles.

[0133] First, the performance of BLMGO was compared across six different transfer models. These six versions of BLMGO are denoted as BLMGOi. i From 1 to 6, corresponding to the... i The migration model was analyzed. Overall, BLMGO2 performed best among the six BLMGO versions, achieving optimal results on 12 out of 30 CEC 2017 test functions, as shown in Tables 3 and 4. The function curves are as follows: Figure 2 As shown.

[0134] Figure 2 In the diagram, the blue dashed line represents BLMGO, the yellow solid line represents BLMGO1, the red solid line represents BLMGO2, the green dashed line represents BLMGO3, the blue solid line represents BLMGO4, the purple dashed line represents BLMGO5, and the pink dashed line represents BLMGO6.

[0135] Table 3. Test results of BLMGO and MGO with different parameters on the CEC2017 function.

[0136]

[0137] Table 4 Ranking Results

[0138]

[0139] Step S6: Optimize the moss algorithm using the optimal biological learning strategy and construct a prediction model with the KNN classifier; design a classifier implementation method and optimization technique based on the training sample set using k-Nearest Neighbor (KNN) and perform classification.

[0140] For example, a KNN classifier is designed and classified based on a sample set. The distance between the sample to be classified and all samples in the training set is calculated, and the k closest samples are selected. The class of the sample to be classified is then determined by voting based on the classes of these k samples. The population is then... Each feature value of each individual in the population is simulated into a binary code value through formulas (1) and (2). The fitness value of the current population individuals is obtained through formulas (3) and (4), and sorted according to the fitness value from smallest to largest.

[0141] The basic idea behind KNN is a simple, parameter-free, and efficient learning method that has shown excellent performance in problems such as function classification and approximation, achieving high completion rates and classification accuracy. In recent research, KNN as a classifier has also demonstrated superior training speed and classification accuracy. Feature selection is a process of obtaining optimal classification accuracy using the fewest subsets, i.e., improving classification accuracy while reducing the number of features. This problem is described as a combinatorial optimization problem. To satisfy each objective, a fitness function is defined using a linear combination of error rate and the number of features. Candidate solutions are comprehensively evaluated to accurately predict the efficacy of immunochemotherapy in non-small cell lung cancer patients.

[0142] (7)

[0143] In the above formula, E The classification accuracy of the KNN classifier. R The length of the selected feature subset. D This represents the total number of features in the dataset. α and β These are the weights for classification error rate and feature reduction, respectively. Since accuracy is clearly more important, [the weights are...]. α Set it to 0.95. β Set it to 0.05.

[0144] The BLMGO2-KNN model is used to select features from routine blood test data and clinical data, identifying the optimal solution, i.e., the optimal feature subset. Feature selection can be viewed as a multi-objective optimization problem, with conflicting objectives: minimizing the number of feature subsets and maximizing classification accuracy. Therefore, an objective function is set to normalize these two objectives into a single function.

[0145] Step S7: Update the loop to the maximum number of iterations and output the optimal solution for predicting non-small cell lung cancer patients, i.e., the optimal subset of clinical features, to improve the accuracy of predicting the efficacy of immunochemotherapy for non-small cell lung cancer patients.

[0146] The embodiments of the present invention have the following beneficial effects:

[0147] Based on the characteristics of clinical data from non-small cell lung cancer (NSCLC) patients, a biological learning strategy is incorporated into the moss optimization algorithm. By screening and extracting the most relevant features, the algorithm improves the predictive performance of immunochemotherapy efficacy in NSCLC patients. This helps medical professionals more accurately identify individuals in the NSCLC patient population who may benefit from long-term progression-free survival, thereby optimizing treatment plans and resource allocation.

[0148] Table 5. Precise Prediction Method for Immunochemotherapy Efficacy in Non-Small Cell Lung Cancer Based on the Invention

[0149]

[0150] Table 5 shows the results of the 10-fold cross-validation experiment for the bBLMGO-KNN model. The ACC value reached 89.647%, the sensitivity was 85.455%, the specificity was 92.692%, the precision was 90.795%, the MCC value was 79.334%, and the F-measurement was 87.685%.

[0151] Table 6 shows the systemic effects of the present invention in the method for accurately predicting the efficacy of immunochemotherapy in non-small cell lung cancer.

[0152]

[0153] Table 6 shows the results of the 10-fold cross-validation experiment for the bBLMGO-KNN model. The ACC value reached 89.647%, the sensitivity was 85.455%, the specificity was 92.692%, the precision was 90.795%, the MCC value was 79.334%, and the F-measurement was 87.685%.

[0154] Based on the same inventive concept, embodiments of the present invention also provide a feature extraction system for medical clinical data, the system comprising:

[0155] The data acquisition module is used to acquire medical clinical datasets, which include blood test data and clinical datasets.

[0156] The processing module replaces the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, removes the cryptic effect of the moss optimization algorithm, and introduces a bimodal propagation search to obtain a bio-learning moss optimization algorithm. This bio-learning moss optimization algorithm is then used to find the subset of clinical features most relevant to sample classification during feature extraction from medical clinical datasets.

[0157] The classification module is used to classify the clinical feature subset that is most relevant to the sample classification using a K-nearest neighbor classifier, thereby obtaining the optimal clinical feature subset of medical clinical data.

[0158] The processing module includes a feature selection module and an optimization module.

[0159] The feature selection module initializes the moss population, where each moss individual represents candidate medical clinical data and has independent feature values, resulting in a feature value set. Feature selection is then performed on this feature value set to obtain the optimal feature subset.

[0160] The optimization module employs a biogeographical learning strategy to construct an example vector set corresponding to the optimal feature subset. It sorts the particles in the example vector set according to the historical best solutions of each moss individual, obtaining a ranking value for each particle. Based on the ranking values, it assigns immigration and emigration rates to the moss individuals corresponding to the optimal feature subset. A two-mode propagation search is then performed based on the immigration and emigration rates to dynamically update the positions of the moss individuals, yielding the global optimal solution. The global optimal solution is the clinical feature subset most relevant to the sample classification.

[0161] Based on the same inventive concept, embodiments of the present invention also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a feature extraction method for medical clinical data.

[0162] Based on the same inventive concept, embodiments of the present invention also provide a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a method for feature extraction of medical clinical data.

[0163] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for feature extraction from medical clinical data, characterized in that, include: Acquire a medical clinical dataset, which includes: blood test data and clinical datasets; By replacing the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, removing the cryptic effect of the moss optimization algorithm, and introducing bimodal propagation search, a bio-learning moss optimization algorithm is obtained. The bio-learning moss optimization algorithm is used to find the subset of clinical features most relevant to sample classification during the feature extraction process of medical clinical datasets. The K-nearest neighbor classifier is used to classify the clinical feature subset that is most relevant to the sample classification, thus obtaining the optimal clinical feature subset of medical clinical data. Specifically, the biological learning moss optimization algorithm is used to find the subset of clinical features most relevant to sample classification during the feature extraction process of medical clinical datasets, including: Initialize a moss population, with each moss individual representing candidate medical clinical data and possessing an independent feature value, resulting in a feature value set; perform feature selection on the feature value set to obtain the optimal feature subset; A biogeographical learning strategy is used to construct an example vector set corresponding to the optimal feature subset. The particles in the example vector set are sorted according to the historical best solution of each moss individual to obtain the sorting value of each particle. Based on the sorting value of each particle, the immigration rate and immigration rate are assigned to the moss individuals corresponding to the optimal feature subset. A two-mode propagation search is performed based on the immigration rate and immigration rate to dynamically update the position of the moss individuals and obtain the global optimal solution, which is the clinical feature subset most relevant to the sample classification.

2. The method for feature extraction of medical clinical data as described in claim 1, characterized in that, The medical clinical dataset is obtained through electronic medical records. The medical clinical dataset specifically includes: demographic characteristics, disease characteristics, pre-treatment hematological indicators, immune characteristics, and efficacy outcome indicators. The efficacy outcome indicators include: efficacy response and survival endpoint; the efficacy response includes: disease remission group and disease non-remission group; the survival endpoint includes: long survival group and short survival group.

3. The method for feature extraction of medical clinical data as described in claim 1, characterized in that, The step of selecting features from the feature set to obtain the optimal feature subset specifically includes: Each moss individual is binary encoded and the feature set is encoded using a transformation function. Feature selection is performed based on the position of each moss individual's feature value in the feature set to obtain multiple feature subsets. For each feature subset, determine the weight of that feature subset in the feature extraction process, obtain the fitness value of that feature subset, and take the feature subset with a fitness value greater than a set threshold as the optimal feature subset.

4. The method for feature extraction of medical clinical data as described in claim 3, characterized in that, The binary encoded values ​​of the initial moss population are obtained by encoding the feature set using a transformation function based on the following formula: ； ； in, For the first The first iteration generated The first individual A binary encoded value, For the first The first iteration generated The first individual 1 eigenvalue, Given a random number in the range [0, 1]. To negate the sign, For conversion functions, This represents the initial moss population; The weight of the feature subset in the feature extraction process is determined based on the following formula, and the fitness value of the feature subset is obtained: ； ； in, For the first The fitness value of an individual moss. For classification error rate, The length of the feature subset. For the total length of the feature, The weights for the classification error rate, Weights for selecting the number of feature subsets. The number of correctly classified samples. This represents the number of samples that were misclassified.

5. The method for feature extraction of medical clinical data as described in claim 1, characterized in that, Based on the following formula, a biogeographical learning strategy is used to construct a set of example vectors corresponding to the optimal feature subset: = ]； ； Based on the following formula, the particles in the example vector set are sorted according to the historical best solution of each moss individual to obtain the sorting value of each particle: = N- 1 , = N- 2 , … = 0 ； Based on the following formula, the immigration and emigration rates are assigned to the moss individuals corresponding to the optimal feature subset according to the ranking value of each particle: ； in, This is the particle's own historical optimal solution. sN For particle subscripts, For the immigration rate, For emigration rate, N The sorting value of the particles. rank For sorting functions, For example vectors, For example vector set, i For indexing.

6. The method for feature extraction of medical clinical data as described in claim 5, characterized in that, The method of performing a dual-mode propagation search based on immigration and emigration rates to dynamically update the location of moss individuals specifically includes: Moss individuals with the best historical solution have the lowest migration rate and tend to update their position to themselves. The moss individual with the worst historical best solution has the highest migration rate and tends to update its position to other moss individuals. The particle with the best historical optimal solution has the highest migration rate and is used as the renewal direction by other moss individuals. The particle with the worst historical best solution has the lowest migration rate and is not used as the renewal direction by other moss individuals.

7. A feature extraction system for medical clinical data based on the feature extraction method for medical clinical data according to any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire medical clinical datasets, which include blood test data and clinical datasets. The processing module is used to replace the moss reproduction mechanism of the moss optimization algorithm with a biogeographical learning strategy, remove the cryptic effect of the moss optimization algorithm, and introduce bimodal propagation search to obtain a bio-learning moss optimization algorithm; the bio-learning moss optimization algorithm is used to find the subset of clinical features most relevant to the sample classification during the feature extraction process of the medical clinical dataset; The classification module is used to classify the clinical feature subset that is most relevant to the sample classification using a K-nearest neighbor classifier, thereby obtaining the optimal clinical feature subset of medical clinical data. The processing module includes a feature selection module and an optimization module. The feature selection module is used to initialize the moss population, where each moss individual represents candidate medical clinical data and has an independent feature value, resulting in a feature value set; feature selection is performed on the feature value set to obtain the optimal feature subset; The optimization module is used to construct an example vector set corresponding to the optimal feature subset using a biogeographical learning strategy, sort the particles in the example vector set according to the historical best solution of each moss individual, and obtain the sorting value of each particle; assign the immigration rate and immigration rate to the moss individuals corresponding to the optimal feature subset according to the sorting value of each particle; perform a two-mode propagation search based on the immigration rate and immigration rate to dynamically update the position of the moss individuals and obtain the global optimal solution, which is the clinical feature subset most relevant to the sample classification.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the feature extraction method for medical clinical data according to any one of claims 1-6.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the feature extraction method for medical clinical data according to any one of claims 1-6.

Images