Drug effect trend prediction method based on regression model

Abstract

The invention relates to the technical field of information retrieval, in particular to a drug effect trend prediction method based on a regression model, and the method comprises the following steps: based on a drug effect query request and a drug target input by a user, accessing a data source directory, calling the access cost of each data source, selecting an optimal item, and establishing an optimal data source execution path. On the basis that a user inputs a drug effect query request and a drug target, access cost calling and optimal path selection are performed on a plurality of data sources, the information positioning and retrieval precision is improved, initial sample distribution is obtained and accurate comparison with target distribution is realized by means of statistics of sample data dimensions, and the accuracy of drug effect query is improved. The quantitative index of data deviation is captured, efficient division of the dominant group and the disadvantage group is further achieved, meanwhile, the dominant group is processed through undersampling guided by information entropy, effective oversampling of physiological index interpolation is adopted for the disadvantage group associated with drug interaction, and the data distribution achieves the remarkable equalization effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information retrieval technology, and particularly relates to a drug efficacy trend prediction method based on a regression model. BACKGROUND

[0002] The field of information retrieval technology is a professional technical field related to quickly and accurately locating, extracting and analyzing user required information from a large data set through specific search strategies and matching algorithms.

[0003] The prior art is difficult to provide effective data screening and analysis for the association between specific drug doses and drug efficacy trends, lacks sufficient consideration of data access costs and quality differences, and thus the retrieved data set often has problems such as insufficient representation and excessive data bias, which affects the accuracy and reliability of the retrieval results. In addition, the prior art lacks fine management of sample data in multidimensional distribution, and fails to achieve effective group division and data balancing, which easily leads to unbalanced data distribution and excessive bias, and is difficult to truly and objectively reflect the actual drug efficacy. Therefore, improvement is needed. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art, and to provide a drug efficacy trend prediction method based on a regression model.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme, a drug efficacy trend prediction method based on a regression model, comprising the following steps:

[0006] Based on the user input drug efficacy query request and the drug target, access the data source directory and call the access cost of each data source, select the optimal item, and establish an optimal data source execution path;

[0007] Based on the optimal data source execution path, initiate retrieval to the target data source, obtain a candidate data set and count the number of samples in the age and disease staging dimensions, obtain an initial sample distribution map, based on the initial sample distribution map, compare with the set target distribution, obtain a candidate data bias quantization index;

[0008] Based on the candidate data bias quantization index, identify the advantage group with positive bias value and the disadvantage group with negative bias value, generate a data group division identifier, based on the data group division identifier, perform information entropy weighted undersampling on the advantage group and physiological index interpolation oversampling on the disadvantage group associated with drug interaction, and establish an equalized drug efficacy feature data set;

[0009] Based on the equalization pharmacodynamic characteristic data set, the drug dose characteristics are associated with the pharmacodynamic results to generate a pharmacodynamic dose response scatter plot, and based on the pharmacodynamic dose response scatter plot, a continuous change trajectory is fitted by connecting all data points in the scatter plot to obtain a pharmacodynamic prediction trend curve.

[0010] Preferably, the preferred data source execution path acquisition step is:

[0011] Based on the user input pharmacodynamic query request and the drug target, the drug target parameter value described in the pharmacodynamic query request is parsed word by word, the drug target parameter value recorded in the data source directory is called and matched item by item, the corresponding data accuracy value, data integrity and data update distance from today in the data source record matched successfully are extracted to generate a data source quality normalization sub-item set;

[0012] According to the data source quality normalization sub-item set, the quality comprehensive score of the data source is calculated;

[0013] According to the quality comprehensive score, the data source record with the highest quality comprehensive score is selected as the optimal data source item to form the preferred data source execution path.

[0014] Preferably, the initial sample distribution map acquisition step is:

[0015] Based on the preferred data source execution path, retrieval requests are initiated to the target data source in sequence, the age parameter value and the disease staging parameter value of each sample recorded in the target data source are extracted, the number of samples in each age category and the number of samples in each disease staging category are counted one by one to generate a sample dimension statistical set;

[0016] According to the sample dimension statistical set, the sample distribution balance index is calculated;

[0017] Based on the sample distribution balance index, the number of category samples in each dimension in the sample dimension statistical set is drawn into a column chart form, and the sample distribution balance index of the corresponding dimension is labeled in the graph to generate an initial sample distribution map.

[0018] Preferably, the candidate data bias quantification index acquisition step is:

[0019] Based on the initial sample distribution map, the number of samples in each category in the age dimension and the disease staging dimension and the sample distribution balance index calculated for the corresponding dimension are extracted respectively, the proportion value of the number of samples in each category to the total number of samples in the respective dimension is calculated one by one to generate an initial sample distribution proportion set;

[0020] According to the initial sample distribution proportion set, the absolute difference values of each class proportion value in the initial sample distribution proportion set and the corresponding class proportion value of the set target distribution are calculated respectively, the absolute difference values of each class proportion value are accumulated respectively, and the sample distribution cumulative difference value of each dimension is generated;

[0021] Based on the sample distribution cumulative difference value and in combination with the sample distribution balance index, the sample distribution cumulative difference value of each dimension is divided by the sample distribution balance index of the corresponding dimension, the standardized deviation quantity value of each dimension is calculated respectively, and the standardized deviation quantity values of all dimensions are added and averaged to obtain the candidate data deviation quantification index.

[0022] Preferably, the data population division identification acquisition step is:

[0023] Based on the candidate data deviation quantification index, the proportion value of the number of samples of each class in the age dimension and the disease staging dimension to the target distribution is extracted respectively, the difference value of each class proportion value and the corresponding class proportion value of the target distribution is calculated one by one, the class with a positive difference value is defined as the dominant class, and the class with a negative difference value is defined as the disadvantaged class, and the initial class deviation identification is generated.

[0024] According to the initial class deviation identification, the class of each sample is judged one by one, if the class of the sample is the dominant class, the sample is marked as the dominant population sample, if the class of the sample is the disadvantaged class, the sample is marked as the disadvantaged population sample, and the samples are marked one by one to obtain the data population division identification.

[0025] Preferably, the acquisition step of the balanced pharmacodynamic feature data set is:

[0026] Based on the data population division identification, the class identification and the corresponding information entropy value of the dominant population sample are extracted in turn, the information entropy threshold is determined after the information entropy values are sorted from large to small, it is judged whether the information entropy value of each dominant population sample is lower than the information entropy threshold, the samples with information entropy values lower than the information entropy threshold are removed, and the information entropy weighted undersampling result set is generated.

[0027] According to the information entropy weighted undersampling result set, the class identification and the corresponding physiological index value of the disadvantaged population sample are extracted one by one, the disadvantaged population samples associated with the drug interaction are screened, and the samples are expanded by interpolation calculation with a fixed step within the physiological index value range to generate the physiological index interpolation oversampling result set.

[0028] Based on the physiological index interpolation oversampling result set, the information entropy weighted undersampling result set and the physiological index interpolation oversampling result set are integrated to generate the balanced pharmacodynamic feature data set.

[0029] Preferably, the obtaining step of the efficacy dose-response scatter plot is:

[0030] Based on the equalized efficacy feature data set, the corresponding drug dose feature value and efficacy result value in each sample record are extracted in sequence, the drug dose feature value and the efficacy result value are combined as a data pair according to the sample number, and a drug dose-efficacy combination data set is generated.

[0031] According to the drug dose-efficacy combination data set, the data pairs of drug dose feature values and efficacy result values are checked one by one, it is verified that the drug dose feature values and the efficacy result values in the data pairs are valid values and are within the predefined value range, and a drug dose-efficacy drawing data file is generated.

[0032] Based on the drug dose-efficacy drawing data file, each data pair in the file is read one by one, the drug dose feature value is plotted as the horizontal axis of the scatter plot, the efficacy result value is plotted as the vertical axis of the scatter plot, and the scatter data points are numbered according to the plotting order, and an efficacy dose-response scatter plot is generated.

[0033] Preferably, the obtaining step of the efficacy prediction trend curve is:

[0034] Based on the efficacy dose-response scatter plot, each scatter data point recorded in the efficacy dose-response scatter plot is read one by one, the drug dose value of each scatter data point is taken as the independent variable, the corresponding efficacy result value is taken as the dependent variable, and all data points are sorted in ascending order according to the drug dose value, forming an efficacy dose-response value set.

[0035] According to the efficacy dose-response value set, a given drug dose value is calculated by a nonlinear regression method.

[0036] A uniformly distributed drug dose value sequence is generated within the drug dose observation range, each dose value in the drug dose value sequence is substituted into the predicted efficacy result value calculation formula one by one to calculate, and all calculated predicted efficacy result values and corresponding drug dose values are sequentially connected into a continuously changing trajectory to generate an efficacy prediction trend curve.

[0037] Compared with the prior art, the advantages and positive effects of the present application are:

[0038] The application improves information positioning and retrieval accuracy by accessing and calling multiple data sources and selecting optimal paths based on user input drug efficacy query request and drug target, obtains initial sample distribution and realizes accurate comparison with target distribution by statistics of sample data dimension, captures quantitative indicators of data deviation, further realizes efficient division of advantage group and disadvantage group, and through information entropy guided undersampling of advantage group and physiological index interpolation oversampling of disadvantage group related to drug interaction, makes data distribution reach significant equalization effect, and improves data accuracy, representativeness and generalization ability; and then adopts correlation mapping to effectively reveal the internal relationship between drug dose and drug efficacy, and supplements continuous change trajectory fitting to improve the continuity and reliability of trend prediction, so that the stability, explainability and practicality of prediction are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The figure is a schematic diagram of the steps of the application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the application more clear, the application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0041] Please refer to Figure 1 The application provides a technical scheme, a drug efficacy trend prediction method based on a regression model, including the following steps:

[0042] Based on the user input drug efficacy query request and drug target, access the data source directory and call the access cost of each data source, select the optimal item, and establish the optimal data source execution path;

[0043] Based on the optimal data source execution path, initiate retrieval to the target data source, obtain the candidate data set and count the sample quantity in the age and disease staging dimension, obtain the initial sample distribution map, compare the initial sample distribution map with the set target distribution, and obtain the candidate data deviation quantitative indicator;

[0044] Based on the candidate data deviation quantitative indicator, identify the advantage group with positive deviation value and the disadvantage group with negative deviation value, generate the data group division identifier, based on the data group division identifier, perform information entropy weighted undersampling on the advantage group and physiological index interpolation oversampling on the disadvantage group related to drug interaction, and establish the equalized drug efficacy characteristic data set;

[0045] Based on the equalization of pharmacodynamic characteristic data set, the drug dose characteristics are associated with the pharmacodynamic results to generate a pharmacodynamic dose response scatter plot, based on the pharmacodynamic dose response scatter plot, by connecting all data points in the scatter plot, a continuous change trajectory is fitted to obtain a pharmacodynamic prediction trend curve.

[0046] The preferred data source execution path acquisition step is:

[0047] Based on the user input pharmacodynamic query request and the drug target, the drug target parameter value described in the pharmacodynamic query request is parsed word by word, the drug target parameter value recorded in the data source directory is called and matched item by item, the corresponding data accuracy value, data integrity and data update distance from today in the matched data source record are extracted to generate a data source quality normalization sub-item set;

[0048] According to the data source quality normalization sub-item set, the quality comprehensive score of the data source is calculated, and the calculation formula is:

[0049]

[0050] Among them, S m is the quality comprehensive score of the mth data source, A′ m is the normalized data accuracy of the mth data source, I′ m is the normalized data integrity of the mth data source, U′ m is the normalized data update distance from today of the mth data source, w a is the weight index of data accuracy, w i is the weight index of data integrity, and λ is a constant parameter of time decay coefficient for adjusting the amplitude of data quality decay with time;

[0051] According to the quality comprehensive score, the data source record with the highest quality comprehensive score is selected as the optimal data source item to form the preferred data source execution path.

[0052] Specifically, based on the user input of the pharmacodynamic query request and the drug target, first, the natural language parsing technology based on the pre-defined rules and keyword matching is adopted to perform the structured processing on the request text, such as “evaluate the inhibitory effect of drug Gefitinib on EGFRL858R mutant”, by identifying the core entities such as “drug” and “target”, the “Gefitinib” is extracted as the drug parameter value and the “EGFRL858R” is extracted as the target parameter value, then, the internally maintained data source directory is accessed, which is stored in the form of a table and contains the data source identifier, the supported drug target list, the data accuracy evaluation value, the data integrity evaluation value and the last update timestamp and the like fields, the system performs the string matching or synonym matching between the extracted target parameter value “EGFRL858R” and the “supported drug target list” of each data source in the directory, for all the matching successful data sources, the system automatically extracts the “data accuracy evaluation value”, “data integrity evaluation value” and “last update timestamp” from the corresponding record row, and calculates the time difference between the current date and the “last update timestamp” to obtain the data update distance from today, for example, if the data source A matches successfully, the accuracy is 0.98, the integrity is 0.95, and the last update is 30 days ago, then the three values are extracted as the original quality sub-items of the data source, and the original quality sub-items of all the matching successful data sources are summarized to generate the data source quality normalized sub-item set.

[0053] Formula: The formula has the advantages that: the form of the Cobb-Douglas production function and the exponential decay model are combined to evaluate the comprehensive quality of the data source. By combining the data accuracy (A' m ) and integrity (I' m ) in the form of product and weighted exponential, it is ensured that the extreme disadvantage of any one dimension (for example, the accuracy is zero) will lead to the total score close to zero, avoiding the risk of selecting the poor data source due to the higher value of other indicators. The time effectiveness index decay term can dynamically punish the data source that is not updated in time, which is crucial for the field of knowledge rapid iteration such as pharmacodynamic research. The design makes the scoring system not only comprehensive, but also reflects the internal constraint relationship of the data in different dimensions and the value change over time.

[0054] A' m is the normalized data accuracy of the mth data source, which reflects the consistency of the pharmacodynamic data in the data source with the recognized benchmark (such as verified clinical trial results or gold standard database). The original accuracy A mThe acquisition method is to randomly extract a certain number (for example, 1000) of efficacy data records from the data source regularly, compare them with the corresponding records in the gold standard database, and calculate the proportion of the number of matching records to the total number of extracted records, that is, In order to eliminate the influence of the dimensions of evaluation indicators from different data sources, the maximum and minimum normalization method is used, and the calculation formula is: Among them A max and A min are the highest and lowest raw accuracy values ​​among all candidate data sources. For example, there are three candidate data sources: DrugBank, ChEMBL, and PubChem, whose raw accuracy is 0.98, 0.95, and 0.92 respectively. max =0.98, A min =0.92, for DrugBank, its normalized accuracy In this example, a data source is selected, and its normalized data accuracy A′ m is 0.9.

[0055] I′ m The data integrity of the mth data source after normalization, which measures the completeness of key field information related to drug efficacy analysis in the data source. m The calculation method is to count the non-null value ratios of key fields (such as drug dosage, dosing frequency, patient age, disease stage, and efficacy endpoint) for the records related to the query request. The calculation formula is: Where N is the total number of samples, K is the total number of key fields, f(v jk ) is a judgment function, which is 1 if the kth field value of the jth sample is not empty, otherwise it is 0. The maximum and minimum normalization method is also used: For example, a data source has 5000 key fields in 1000 related records, of which 4600 fields are non-empty. m =4600 / 5000=0.92. After normalizing the integrity of all candidate data sources, the normalized data integrity of the data source selected in this example is I′ m is 0.8.

[0056] U′ m The number of days since the data update of the mth data source after normalization is used to quantify the timeliness of the data. m This value is obtained by obtaining the "Last Update Date" in the data source metadata and calculating the difference in days between it and the current date. The smaller the value, the newer the data. To unify the evaluation, the maximum and minimum normalization process is also used: Among them U maxand U min are the maximum and minimum update interval days of all candidate data sources respectively. For example, if the update interval days of candidate data sources are 15 days, 60 days and 180 days respectively, then U min = 15, U max = 180. For the data source with an update interval of 60 days, its normalized value In this example, a data source is selected, and its normalized data update interval U' m is 0.2.

[0057] w a is the weight index of data accuracy, which is not fixed but dynamically adjusted according to the nature of the user query request. Its setting basis is to analyze the risk level and application scenario of the query request, and to determine it by using the analytic hierarchy process (AHP). For example, "accuracy" and "completeness" are compared with each other, and if the query is used for "early drug discovery", completeness may be more important; if it is used for "clinical decision support", accuracy is crucial. A judgment matrix is constructed by domain experts (such as pharmacologists, clinicians), and the maximum eigenvalue and the corresponding eigenvector are calculated, and the normalized eigenvector component is the weight. In this example, for the regular efficacy trend prediction, the importance of accuracy is slightly higher than that of completeness, and the value of w a is calculated to be 0.6.

[0058] w i is the weight index of data completeness, which is set in the same way as w a , determined by the analytic hierarchy process (AHP). In the weight system, the sum of the weights is usually set to 1, i.e. w a +w i = 1, to reflect the relative importance of the two in the evaluation system. Based on the setting of w a , the weight index w i here is w i = 1-0.6 = 0.4.

[0059] λ is the time decay coefficient, which is used to adjust the magnitude of the decay of data quality over time. Its setting refers to the knowledge half-life in a specific research field. For example, in the field of tumor drug research, new target and mechanism research results are published frequently, and the knowledge half-life is short, so a larger λ value should be set to increase the punishment for old data. The specific method of setting λ is as follows: first, determine a normalized value U' crit (0.8, for example) corresponding to an acceptable maximum update time lag, and set the timeliness factor to a lower value (0.1, for example) at this time, and then solve λ inversely. The calculation formula is For a comprehensive platform, a compromise value can be taken. In this example, λ is set to 1.4.

[0060] The calculation process is as follows:

[0061] Now, for the mth data source, substitute the specific values of the parameters obtained into the quality comprehensive score S m .

[0062] The known parameters are as follows:

[0063] The normalized data accuracy A′ m = 0.9;

[0064] The normalized data integrity I′ m = 0.8;

[0065] The normalized data update distance from today U′ m = 0.2;

[0066] The data accuracy weight index w a = 0.6;

[0067] The data integrity weight index w i = 0.4;

[0068] The timeliness decay coefficient λ = 1.4;

[0069] The calculation process is as follows:

[0070] S m = (0.9) 0.6 · (0.8) 0.4 · e ；1.4·0.2 ;

[0071] S m = 0.9392 · 0.9123 · e ；0.28 ;

[0072] S m = 0.9392 · 0.9123 · 0.7558;

[0073] S m = 0.6473;

[0074] The result shows that the quality comprehensive score of the mth data source is 0.6473. This value is a standardized measure between 0 and 1, which comprehensively reflects the performance of the data source in the three dimensions of accuracy, integrity, and timeliness. The system will calculate this score for all candidate data sources, and the higher the score, the better the comprehensive quality of the data source. For example, if another data source scores 0.81, its quality is better than the current data source. If the score is lower than the preset minimum quality threshold (for example, 0.5), the data source will be directly excluded.

[0075] According to the quality comprehensive score, the system will all candidate data sources and their corresponding quality comprehensive score S m Organized into a list, and in descending order of score value, for example, get the ranking list is [{data source B: 0.92}, {data source A: 0.85}, {data source C: 0.71}], on this basis, the system will rank first data source, that is, data source B, selected as the optimal master data source, at the same time, in order to ensure the absolute quality of the selected data source, the system will also apply a preset minimum quality access threshold, for example 0.70, the threshold is determined by analyzing the historical score distribution of the recognized authority database in the field (such as KEGG, DrugBank) on the standard test set, taking its 10% quantile as the benchmark, since the score of the optimal data source B is 0.92, which is higher than the threshold 0.70, its selection is finally confirmed, then, the system not only selects a single optimal item, but also constructs a priority access sequence containing multiple high-quality data sources, arranges all data sources with scores higher than 0.70 in order of score height, forming an ordered candidate list, for example [data source B, data source A, data source C], this ordered list constitutes the preferred data source execution path.

[0076] The initial sample distribution map is obtained by the following steps:

[0077] Based on the preferred data source execution path, the system initiates a retrieval request to the target data source one by one, extracts the age parameter value and disease stage parameter value of each sample recorded in the target data source, and counts the number of samples in each age category and the number of samples in each disease stage category, generating a sample dimension statistical set;

[0078] According to the sample dimension statistical set, the sample distribution balance index is calculated, and the calculation formula is:

[0079]

[0080] Among them, G c is the sample distribution balance index of the cth dimension, Y cb is the sample number corresponding to the bth category in the cth dimension, N c is the total number of samples in the cth dimension, and M is the total number of categories in the cth dimension.

[0081] Based on the sample distribution balance index, the number of category samples in each dimension in the sample dimension statistical set is drawn into a column chart, and the sample distribution balance index of the corresponding dimension is marked in the graph, generating an initial sample distribution map.

[0082] Specifically, based on the preferred data source execution path obtained in the previous step, the system constructs and sends a standardized data retrieval request one by one according to the priority order of the data sources in the path. For example, for the first data source in the path, the system generates an HTTPS GET request that meets the API specification of the data source. The request specifies that the entity to be retrieved is the pharmacodynamic research data matching the user's query request, and limits the returned fields to include "age" (age) and "disease stage" (disease_stage).

[0083] (disease_stage), In the returned data (usually in JSON or XML format), the system parses each sample record and extracts the original values of the age and disease stage parameters. For the age parameter, the system classifies the values into four categories according to the pre-set clinical stratification standard: 'adolescent' (0-17 years old), 'young adult' (18-44 years old),'middle-aged' (45-64 years old), and 'elderly' (65 years old and above). For the disease stage parameter, the system directly uses the original classification labels, such as'stage I','stage II','stage III', and'stage IV'. The system maintains two independent counters (e.g., hash table structure) internally, one for the age dimension and the other for the disease stage dimension. The keys are the category names and the values are the sample counts. Each time a sample is processed, the corresponding age category and disease stage category counters are incremented by one. After traversing all the samples returned by all data sources, the data in these two counters form the sample dimension statistical set.

[0084] Formula: The benefits of the formula are as follows: The formula uses Gini impurity, which can quantify the data distribution balance of a categorical variable (such as age stratification and disease stage). Compared with simple variance or standard deviation, the formula's result is normalized to between 0 and 1, providing a universal evaluation index that is independent of the specific sample size and category number. When the sample size of each category is completely equal, the index reaches the maximum value (tends to 1), indicating the most balanced distribution. When all samples are concentrated in a certain category, the index reaches the minimum value 0, indicating the most unbalanced distribution. This quantification method provides clear and objective decision-making basis for subsequent data balancing strategies (such as oversampling and undersampling), allowing the system to automatically determine which dimensions have the most serious bias and prioritize adjustment.

[0085] Y cbis the number of samples corresponding to the b-th category in the c-th dimension. This parameter is obtained directly from the "sample dimension statistical set" generated in the previous step. The system will traverse the set and extract the statistical count values ​​of each category b under the specified dimension c (for example, c = 1 represents the age dimension, c = 2 represents the disease stage dimension). For example, when analyzing the age dimension, four categories (b = 1 to 4) are pre-defined, corresponding to 'adolescents', 'youth', 'middle-aged', and 'elderly' respectively. By querying the "sample dimension statistical set", the number of samples in each age group is obtained as follows: Y 1,1 =50, Y 1,2 =350, Y 1,3 =450, Y 1,4 =150.

[0086] N c is the total number of samples in the cth dimension. This parameter is also calculated based on the "sample dimension statistics set". It is the sum of the number of samples of all categories under dimension c. The calculation formula is: This value reflects the total amount of data used for analysis under the current dimension. For example, continuing with the example of age dimension, the total number of samples is N1 = Y 1,1 +Y 1,2 +Y 1,3 +Y 1,4 =50+350+450+150=1000 people.

[0087] M is the total number of categories in the cth dimension. This parameter defines how many mutually exclusive categories the samples are divided into under a specific dimension. This value is determined based on the classification criteria set in the data preprocessing stage and can be directly read from the structure of the "sample dimension statistical set". For example, in the age dimension, four categories are set: 'adolescent', 'youth', 'middle-aged', and 'elderly', so M = 4. In the disease stage dimension, if it is divided into 'stage I', 'stage II', 'stage III', and 'stage IV' according to the TNM staging system, then M for this dimension is also 4.

[0088] Calculation process:

[0089] Taking the age dimension (c=1) as an example, according to the above parameter acquisition process, the following values ​​are obtained:

[0090] Total number of categories M = 4;

[0091] Number of samples in each category: Y 1,1 =50, Y 1,2 =350, Y 1,3 =450, Y 1,4 =150;

[0092] The total number of dimension samples N1 = 1000;

[0093] Substitute these values into the sample distribution balance index formula:

[0094]

[0095]

[0096] G1 = 1 - [(0.05) 2 +(0.35) 2 +(0.45) 2 +(0.15) 2 ];

[0097] G1 = 1 - [0.0025 + 0.1225 + 0.2025 + 0.0225];

[0098] G1 = 1 - 0.35;

[0099] G1 = 0.65;

[0100] The results show that the sample distribution balance index of the age dimension is 0.65. This value quantitatively describes the degree of imbalance in the distribution of samples in different age groups in the current candidate data set. Generally, an evaluation standard can be set: an index greater than 0.8 indicates that the distribution is very balanced and does not need special processing; an index between 0.5 and 0.8 indicates that there is moderate imbalance and data enhancement adjustment is needed; an index less than 0.5 indicates that it is severely imbalanced and strong balancing measures must be taken.

[0101] Based on the sample distribution balance index of each dimension calculated in the previous step and the sample dimension statistical set, the system calls a graphics drawing library (such as the Matplotlib library in Python) to automatically generate an independent bar chart for each analysis dimension (age and disease stage). When drawing, for the age dimension, the system uses the age categories in the "sample dimension statistical set" ('teenagers', 'young','middle-aged', 'old') as the horizontal axis (X-axis) scale labels of the bar chart, and the number of samples corresponding to each category (such as 50, 350, 450, 150) as the values of the vertical axis (Y-axis), and accordingly draws bar bodies of different heights. Similarly, for the disease stage dimension, the same operation is performed, using 'I stage', 'II stage', etc. as the horizontal axis labels, and drawing the corresponding sample number bar chart. After generating each chart, the system displays the sample distribution balance index of this dimension (for example, the G cAs a text element, the sample distribution balance index (0.65) is added to a fixed position of the chart, for example, the blank area below the chart title or the upper right corner, formatted as "Sample Distribution Balance Index: 0.65", finally, the system integrates all the bar charts generated for different dimensions with labeled balance indexes to form a multi-chart report or a collection containing all chart objects, which is the initial sample distribution atlas.

[0102] The obtaining step of the candidate data bias quantification index is:

[0103] Based on the initial sample distribution atlas, the sample quantity values and the sample distribution balance indexes calculated for each category in the age dimension and the disease staging dimension are extracted, the proportion values of the sample quantity of each category to the total sample quantity in the respective dimension are calculated one by one, and an initial sample distribution proportion set is generated;

[0104] According to the initial sample distribution proportion set, the category proportions in the corresponding dimensions of the set target distribution are called, the absolute difference values between each category proportion value in the initial sample distribution proportion set and the corresponding category proportion value of the set target distribution are calculated, the absolute difference values of each category proportion value are accumulated respectively, and the sample distribution cumulative difference values of each dimension are generated;

[0105] Based on the sample distribution cumulative difference values and in combination with the sample distribution balance indexes, the sample distribution cumulative difference values of each dimension are divided by the sample distribution balance indexes of the corresponding dimensions, the standardized bias quantity values of each dimension are calculated respectively, all the standardized bias quantity values of all dimensions are added and averaged, and the candidate data bias quantification index is obtained.

[0106] Specifically, based on the initial sample distribution atlas, the system first programmatically parses the data structure associated behind the atlas, that is, the sample dimension statistical set. For the age dimension, the system traverses each of the preset categories, including 'adolescent', 'young','middle-aged', and 'old', and extracts the sample quantity corresponding to each category, for example, 50, 350, 450, and 150 respectively. At the same time, the total number of samples in this dimension is obtained, that is, 1000. Subsequently, a division operation is performed one by one, and the sample quantity of each category is divided by the total sample quantity to calculate the actual proportion of this category in the current data set. For example, the proportion of 'adolescent' is 50 divided by 1000, which is 0.05, the proportion of 'young' is 350 divided by 1000, which is 0.35, the proportion of'middle-aged' is 450 divided by 1000, which is 0.45, and the proportion of 'old' is 150 divided by 1000, which is 0.15. The same process is performed for the disease staging dimension, and the sample quantities of each stage (such as'stage I','stage II','stage III', and'stage IV') are extracted and their proportions are calculated. The proportion values calculated for each category in all dimensions are organized into a data structure with dimensions as keys and key-value pairs of categories and proportions as values, generating an initial sample distribution proportion set.

[0107] According to the initial sample distribution proportion set generated in the previous step, the system calls a set target distribution from a pre-configured knowledge base. The target distribution is not arbitrarily set, but is based on authoritative epidemiological data or large clinical research guidelines on the incidence rate distribution data of the disease in different age and staging populations, and is formed after standardization processing. For example, the set target age distribution is 'adolescent' proportion 0.10, 'young' proportion 0.30,'middle-aged' proportion 0.40, and 'old' proportion 0.20. Next, the system extracts the proportion of each category in the age dimension in the initial sample distribution proportion set (such as 'adolescent' 0.05) and the corresponding proportion in the target distribution (0.10), calculates the absolute value of the difference between the two, that is, 0.05. Repeat this calculation for all categories such as 'young','middle-aged', and 'old' to obtain a series of absolute difference values. Then, the system sums the absolute difference values of all categories in this dimension, for example, 0.05 plus 0.05 (young), 0.05 (middle-aged), and 0.05 (old), to obtain the cumulative difference value of the age dimension as 0.20. This process is also applied to the disease staging dimension, and finally a total difference value is calculated for each dimension to generate the sample distribution cumulative difference value for each dimension.

[0108] Based on the sample distribution cumulative difference value of each dimension, and combined with the previously calculated sample distribution balance index, the system first extracts the two values corresponding to the age dimension, i.e. the sample distribution cumulative difference value 0.20 and the sample distribution balance index 0.65, and calculates the standard deviation value of the age dimension by dividing the former by the latter, which is about 0.308. The purpose of this calculation is to correct the original cumulative difference using the balance index. If the internal difference of a dimension is very large (i.e. the balance index is low), the weight of its cumulative difference value will be amplified, and vice versa. Subsequently, the system also performs the same calculation for the disease staging dimension. For example, if its cumulative difference value is 0.30 and the balance index is 0.685, then its standard deviation value is about 0.438. After obtaining the standard deviation values of all dimensions (in this case, age and disease staging), the system sums these values (0.308 and 0.438) and then divides by the total number of dimensions (i.e. 2) to calculate the average value, which is the final single evaluation value 0.373. This value is the candidate data deviation quantification indicator.

[0109] The acquisition step of the data population division identifier is:

[0110] Based on the candidate data deviation quantification indicator, the proportion of the number of samples in each category in the age dimension and the disease staging dimension to the target distribution is extracted, and the difference between the proportion of each category and the proportion of the corresponding category in the target distribution is calculated. The category with a positive difference value is defined as the dominant category, and the category with a negative difference value is defined as the disadvantaged category, to generate an initial category deviation identifier.

[0111] According to the initial category deviation identifier, the category of each sample is judged one by one. If the sample belongs to the dominant category, it is marked as a dominant population sample, and if the sample belongs to the disadvantaged category, it is marked as a disadvantaged population sample. Each sample is marked to obtain the data population division identifier.

[0112] Specifically, based on the calculation result of the candidate data bias quantification index, the system continues to utilize the existing initial sample distribution proportion set and the set target distribution data. For the age dimension, the system first extracts the proportion of the 'young' category in the initial sample 0.35 and the proportion in the target distribution 0.30, calculates the difference value as 0.35 minus 0.30 to obtain +0.05. Since the difference value is positive, the system determines the 'young' category as the "dominant category". Then, the system extracts the proportion of the 'old' category in the initial sample 0.15 and the proportion in the target distribution 0.20, calculates the difference value as 0.15 minus 0.20 to obtain -0.05. Since the difference value is negative, the system determines the 'old' category as the "disadvantaged category". This process will be applied to all categories in the age dimension and the disease staging dimension one by one, and finally a mapping table is formed, where the key is the combination of the dimension and the category (such as 'age-young'), and the value is the bias state ('dominant' or 'disadvantaged'). This mapping table is the initial category bias identifier.

[0113] According to the initial category bias identifier, the system starts to traverse each sample record in the original candidate data set. For the first sample, the system reads its age parameter value as '48' and disease staging parameter value as 'II stage'. First, the system maps the age '48' to the'middle-aged' category according to the preset rule (45-64 years old as middle-aged), and then queries the state corresponding to the 'age-middle-aged' entry in the initial category bias identifier. If the state is 'dominant', a new field, for example named 'group identifier', is added to this sample record and assigned the value 'dominant group sample'. Then, the system reads the disease staging 'II stage' and also queries the state of 'disease staging-II stage' in the initial category bias identifier. If it is 'disadvantaged', the 'group identifier' field of the sample is marked as 'disadvantaged group sample'. The system stipulates that as long as the sample belongs to the disadvantaged category in any dimension, its final identifier is 'disadvantaged group sample'. Only when it belongs to the dominant category in all dimensions, it is marked as 'dominant group sample'. The system sequentially executes this judgment and marking process for all thousands of samples in the data set, and finally assigns each sample a clear group affiliation, thereby obtaining the complete data group division identifier.

[0114] The acquisition steps of the balanced pharmacodynamic characteristic data set are:

[0115] Based on the data group division identifier, the category identifier and the corresponding information entropy value of the dominant group sample are extracted in turn. The information entropy threshold is determined by sorting the information entropy values from large to small. It is determined whether the information entropy value of each dominant group sample is lower than the information entropy threshold. Samples with information entropy values lower than the information entropy threshold are removed to generate an information entropy weighted undersampling result set.

[0116] According to the information entropy weighted undersampling result set, the class identifier and the corresponding physiological index value of the disadvantaged group sample are extracted piece by piece, the disadvantaged group sample associated with the drug interaction is screened, and the interpolation calculation is performed in the physiological index value range with a fixed step to expand the sample, and the physiological index interpolation oversampling result set is generated.

[0117] Based on the physiological index interpolation oversampling result set, the information entropy weighted undersampling result set and the physiological index interpolation oversampling result set are integrated to generate the balanced pharmacodynamic characteristic data set.

[0118] Specifically, based on the data group division identifier, the system first screens all records marked as 'advantageous group samples', and calculates the information entropy value of each such sample. The information entropy value is calculated based on the probability distribution of the fine subcategory to which the sample belongs (for example, in the'middle-aged' group, it can be further divided into smaller age groups such as '45-49 years old', '50-54 years old', etc.), and is used to quantify the rarity or information amount of the sample. The higher the entropy value of the sample, the more rare the subcategory it belongs to, and the higher the information value. After calculating the information entropy of all advantageous group samples, the system collects these entropy values to form a numerical list, and sorts the list from large to small. Next, the information entropy threshold is determined by calculating the 75th percentile of the entropy value list, i.e. selecting the entropy value at the top 25% position after sorting as the threshold. For example, if there are 100 entropy values, the 25th entropy value after sorting is 0.85, then this value is the information entropy threshold. Subsequently, the system iterates through all the advantageous group samples again, and compares the information entropy value of each sample with the threshold 0.85. If the information entropy value of the sample is lower than 0.85, the sample is removed from the data set. All samples with information entropy not lower than 0.85 are retained to form the information entropy weighted undersampling result set.

[0119] According to the information entropy weighted undersampling result set, the system then turns its attention to the records marked as 'disadvantaged group samples' by the data population division identifier, the system reads these disadvantaged samples one by one, extracts their class identifier (such as 'elderly','stage IV') and all physiological indicator values in the dataset, including, for example,'serum albumin concentration', 'glomerular filtration rate', 'liver function ALT value', etc., then the system accesses a pre-built drug knowledge base that stores known interaction relationships between the target drug and various physiological indicators, for example, 'Gefitinib is mainly metabolized by liver CYP3A4 enzyme, so liver function indicators ALT, AST are highly related', the system filters out those disadvantaged group samples that have clear pharmacokinetic or pharmacodynamic associations with the current research drug through keyword matching and relationship querying, for the filtered samples, for example, an 'elderly' sample with a low liver function indicator ALT value, within the reasonable medical value range of its physiological indicators, for example, the normal range of ALT is 7-56 U / L, the system generates multiple new samples with different but reasonable ALT values by linear interpolation with a fixed step size, for example, 5 U / L, these new samples inherit all the characteristics of the original sample except for the interpolated physiological indicator, by performing this operation on all related disadvantaged group samples, the sample size is expanded, generating a physiological indicator interpolation oversampling result set.

[0120] Based on the physiological indicator interpolation oversampling result set, the system performs the final data set merging operation, first, the system collects the advantaged group samples retained after information entropy weighted undersampling (i.e. the information entropy weighted undersampling result set) and the newly generated disadvantaged group samples (i.e. the physiological indicator interpolation oversampling result set), at the same time, in order to ensure the integrity of the dataset, those original disadvantaged group samples whose physiological indicators are not identified as having a direct association with the drug are also included in their entirety, the system combines these three parts of data (undersampled advantaged group, newly generated disadvantaged group through oversampling, original disadvantaged group not oversampled) into a unified data table structure, and performs a randomized reordering on the merged dataset to disrupt the original order of the samples, finally, the system performs a final check on the merged dataset to check whether the data types and value ranges of all feature fields are consistent and valid, the final dataset obtained is the balanced pharmacodynamic feature dataset.

[0121] The steps for obtaining the pharmacodynamic dose response scatter plot are:

[0122] Based on the balanced pharmacodynamic feature dataset, the corresponding drug dose feature values and pharmacodynamic result values in each sample record are extracted in turn, the drug dose feature values and pharmacodynamic result values are combined into data pairs according to the sample number one by one to generate a drug dose pharmacodynamic combination data set;

[0123] According to the drug dose efficacy combination data set, the drug dose characteristic value and the efficacy result value of each data pair are checked one by one to verify that both the drug dose characteristic value and the efficacy result value in the data pair are valid values and are within the predefined value range, and a drug dose efficacy drawing data file is generated;

[0124] Based on the drug dose efficacy drawing data file, each data pair in the file is read one by one, the drug dose characteristic value is plotted as the horizontal axis of the scatter plot, the efficacy result value is plotted as the vertical axis of the scatter plot, and the scatter data points are numbered in the order of plotting, and an efficacy dose response scatter plot is generated.

[0125] Specifically, based on the balanced efficacy characteristic data set, the system starts an iteration process, scans each sample record in the data set line by line, for each record, the system locates the "drug dose" feature column and the "efficacy result" feature column, and reads their values respectively, for example, extracts the drug dose as 50mg and the efficacy result as tumor shrinkage rate 20% from the first record, then the system takes the two values as an ordered data pair (50, 20), and associates the unique sample ID (for example, ID-001) of the record, then adds the tuple containing the sample ID, drug dose and efficacy result to a temporary in-memory list, this process continues until all sample records in the balanced efficacy characteristic data set are traversed, finally, the temporary in-memory list containing all sample dose and efficacy pairing information constitutes the drug dose efficacy combination data set.

[0126] According to the drug dose efficacy combination data set, the system initializes an empty list for storing valid data, i.e. the drug dose efficacy drawing data, then the system starts to traverse each data pair in the drug dose efficacy combination data set, for each data pair, the system first checks whether the drug dose characteristic value and the efficacy result value are valid non-empty, non-numeric value types, then compares the two values with the predefined reasonable value range, which is set based on pharmacological common sense and specific drug clinical trial data, for example, the valid range of drug dose is set to greater than 0mg to 1000mg, and the range of efficacy result (such as tumor shrinkage rate) is set to -50% (indicating growth) to 100% (indicating complete remission), if a data pair, for example (50, 20), both of its values pass the non-empty check, and 50 is within the range (0, 1000], and 20 is within the range [-50, 100], then the data pair is determined to be valid and is added to the drug dose efficacy drawing data list, if any of the checks fails, for example, the dose is -10mg or the efficacy result is null, the data pair will be discarded, after the traversal is completed, the final list is the drug dose efficacy drawing data file.

[0127] Based on the drug dose efficacy plot data file, the system calls the graphical component to create a two-dimensional coordinate system chart, and sets the horizontal axis (X-axis) title as "Drug dose (unit: mg)", and the vertical axis (Y-axis) title as "Efficacy result (unit: %)", then the system reads each data pair in the drug dose efficacy plot data file one by one, for example, the first data pair is (50, 20), the system locates the position of the horizontal coordinate 50 and the vertical coordinate 20 on the chart, and draws a solid dot as a scatter point at this position, at the same time, the system associates a unique number with this point, for example, "point 1", which can be displayed in the form of a prompt box when the user hovers the mouse over the point, but will not be directly drawn on the chart, the system continues to process the next data pair, for example, (100, 45), and draws the second scatter point at the corresponding position, marked as "point 2", this process continues until all data pairs in the drug dose efficacy plot data file are converted into scatter points on the chart, and the final completed chart is the efficacy dose response scatter plot.

[0128] The acquisition steps of the efficacy prediction trend curve are:

[0129] Based on the efficacy dose response scatter plot, read each scatter point data point recorded in the efficacy dose response scatter plot one by one, take the drug dose value of each scatter point data point as the independent variable, and the efficacy result value as the dependent variable, and sort all data points in ascending order according to the drug dose value to form an efficacy dose response value set;

[0130] According to the efficacy dose response value set, the non-linear regression method is used to calculate the given drug dose value, and the calculation formula is:

[0131]

[0132] Wherein, ZY pred is the predicted efficacy result value under the given drug dose value X, X is the drug dose value, T is the maximum asymptote value of the efficacy dose response curve, i.e. the maximum efficacy value, B is the minimum asymptote value of the efficacy dose response curve, i.e. the basal effect value, EC 50 is the drug dose value required to produce 50% of the maximum efficacy result value, and H is the Hill slope value of the inflection point of the efficacy dose response curve.

[0133] A uniformly distributed drug dose value sequence is generated within the drug dose observation range, each dose value in the drug dose value sequence is substituted into the predicted efficacy result value calculation formula one by one to calculate, and all calculated predicted efficacy result values and corresponding drug dose values are connected in sequence to generate an efficacy prediction trend curve.

[0134] Specifically, based on the efficacy dose-response scatter plot, the system first accesses the data structure associated behind the chart, that is, the drug dose efficacy drawing data file, extracts the coordinate values of each scatter data point one by one, for each data point, the system records the horizontal coordinate value as the drug dose value as the independent variable, and records the vertical coordinate value as the efficacy result value as the dependent variable, and stores the pair of values together with the sample number as a tuple, for example (sample number: 'ID-101', drug dose: 50, efficacy result: 20), after extracting all the data of the scatter points, the system forms a temporary list of these tuples, then the system calls a sorting algorithm, for example, quicksort, to sort all tuples in the list in ascending order of drug dose value, this operation ensures that the data points are organized in order of dose from small to large, and the final ordered list is the efficacy dose-response value set.

[0135] Formula: The benefit of the formula is that in the low dose region, the effect is not obvious (the basic effect), with the increase of the dose, the effect is rapidly enhanced, and finally reaches a platform in the high dose region (the maximum effect).

[0136] X is the drug dose value, which is the independent variable of the model, representing the drug dose given to the subject. This value is directly derived from the dose data in the "efficacy dose-response value set" generated in the previous step. When performing nonlinear regression fitting, all drug dose values in this set will be used as input for parameter estimation. For example, in the "efficacy dose-response value set", there are a series of dose values such as 10mg, 25mg, 50mg, 100mg, 200mg, 400mg, etc.

[0137] T is the maximum asymptote value of the efficacy dose-response curve, that is, the maximum efficacy value. It represents the theoretical maximum effect that the drug can achieve when the drug dose tends to infinity. This parameter is not pre-set, but is estimated by the nonlinear regression algorithm fitting the data points in the "efficacy dose-response value set". The algorithm adjusts T and other parameters through iteration to minimize the sum of squared residuals between predicted values and actual observations. For example, after fitting a set of experimental data, the maximum tumor inhibition rate T is 98.5%.

[0138] B is the minimum asymptote value of the dose-response curve, i.e., the baseline effect value. It represents the baseline effect in the absence of drug administration (zero dose) or at very low doses, which can reflect the placebo effect or the natural progression of the disease. Like T, B is also a parameter estimated from the "set of dose-response values" by nonlinear regression fitting. For example, in the evaluation of an anticancer drug, the fitted baseline tumor growth rate (negative effect) B can be -5%, indicating that the tumor volume grows at an average rate of 5% without treatment.

[0139] EC 50 is the dose value of the drug required to produce 50% of the maximum effect value. This parameter is a key indicator of the potency of the drug, and EC 50 The smaller the value, the lower the dose required to achieve the same effect, and the stronger the potency. Its exact definition is the dose of the drug required to produce an effect at the midpoint between B and T, i.e., B + (T-B) / 2. This parameter is also obtained by nonlinear regression fitting of the "set of dose-response values". For example, the fitted EC 50 value for a drug is 120 mg.

[0140] H is the Hill slope value of the dose-response curve at the inflection point (i.e., X = EC 50 ). This parameter describes the steepness of the dose-effect curve. If H > 1, the effect changes dramatically with dose, showing a "switch" effect; if H < 1, the curve is relatively flat. The value of H is also fitted by the nonlinear regression algorithm based on the data characteristics of the "set of dose-response values". For example, for a typical drug interaction, the fitted value of the Hill slope H is 1.5.

[0141] Calculation process:

[0142] First, the "set of dose-response values" is fitted by a nonlinear regression algorithm to obtain a set of optimal parameter estimates. In this example, the parameter values obtained are:

[0143] Maximum effect value T = 98.5 (%);

[0144] Baseline effect value B = -5.0 (%);

[0145] Half-effective dose EC 50 = 120 (mg);

[0146] Hill slope H = 1.5;

[0147] Now, using these determined parameters, the drug effect ZY pred is predicted when the given drug dose X = 150 mg.

[0148] The calculation process is as follows:

[0149]

[0150]

[0151]

[0152]

[0153] ZY pred = -5.0 + 43.169;

[0154] ZY pred = 38.169;

[0155] The result shows that according to the established dose-effect model, when a drug dose of 150 mg is given, the predicted pharmacodynamic result (such as tumor inhibition rate) is 38.169%. This calculation process shows how to use the fitted model for single-point prediction.

[0156] Within the observed range of drug doses, the system first extracts the minimum and maximum drug dose values from the pharmacodynamic dose response value set, for example, 10 mg and 500 mg, to determine the dose interval for generating a smooth curve, then the system generates a drug dose value sequence containing 200 equidistant points in this interval [10, 500], that is, starting from 10 mg, increasing by a step of (500-10) / 199 until 500 mg, then the system traverses each dose value in the newly generated drug dose value sequence and substitutes it as X into the predicted pharmacodynamic result value calculation formula of all parameters determined in the previous step, that is, The calculation is performed to obtain a corresponding predicted pharmacodynamic result value. After the traversal is completed, the system obtains a new sequence containing 200 predicted pharmacodynamic result values, which corresponds one-to-one to the drug dose value sequence. Finally, the system takes the two sequences as input, calls the line drawing function of the graphics library, and connects all (drug dose, predicted pharmacodynamic) data points in order with a smooth line segment, which is the final pharmacodynamic prediction trend curve.

[0157] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments still belongs to the protection scope of the present application technical solution.

Claims

1. A method for predicting drug efficacy trends based on a regression model, characterized in that: The following steps are involved: Based on the drug efficacy query request and drug target input by the user, access the data source directory and retrieve the access cost of each data source, select the optimal option, and establish the preferred data source execution path; Based on the preferred data source execution path, a search is initiated to the target data source to obtain a candidate data set and count the number of samples in the age and disease stage dimensions to obtain an initial sample distribution map. Based on the initial sample distribution map, the map is compared with the set target distribution to obtain a quantitative index of candidate data deviation; Based on the candidate data deviation quantification index, a dominant group with a positive deviation value and a disadvantaged group with a negative deviation value are identified, a data group division identifier is generated, and based on the data group division identifier, information entropy weighted undersampling is performed on the dominant group and physiological indicator interpolation oversampling is performed on the disadvantaged group associated with drug interactions to establish a balanced drug efficacy feature data set; Based on the balanced pharmacodynamic characteristic data set, the drug dosage characteristics are correlated with the pharmacodynamic results to generate a pharmacodynamic dose-response scatter plot. Based on the pharmacodynamic dose-response scatter plot, a continuous change trajectory is fitted by connecting all data points in the scatter plot to obtain a pharmacodynamic prediction trend curve.

2. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the preferred data source execution path are: Based on the drug efficacy query request and drug target input by the user, the drug target parameter values ​​described in the drug efficacy query request are parsed word by word, the drug target parameter values ​​recorded in the data source directory are called and matched item by item, and the corresponding data accuracy values, data completeness, and data update time from the successfully matched data source records are extracted to generate a data source quality normalization sub-item set; Calculating a comprehensive quality score of the data source based on the data source quality normalization sub-item set; According to the comprehensive quality score, the data source record with the highest comprehensive quality score is selected as the optimal data source item to form a preferred data source execution path.

3. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the initial sample distribution map are: Based on the preferred data source execution path, retrieval requests are initiated to the target data source in sequence, the age parameter value and the disease stage parameter value of each sample recorded in the target data source are extracted, and the number of samples in each age category and the number of samples in each disease stage category are counted one by one to generate a sample dimension statistical set; Calculating a sample distribution balance index based on the sample dimension statistical set; Based on the sample distribution balance index, the number of category samples in each dimension in the sample dimension statistical set is plotted in the form of a bar graph, and the sample distribution balance index of the corresponding dimension is marked in the graph to generate an initial sample distribution map.

4. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the candidate data deviation quantitative index are as follows: Based on the initial sample distribution map, the sample quantity values ​​of each category in the age dimension and the disease stage dimension and the sample distribution balance index calculated for the corresponding dimension are extracted respectively, and the proportion of the number of samples of each category to the total number of samples in the respective dimension is calculated one by one to generate an initial sample distribution ratio set; Based on the initial sample distribution ratio set, call the category ratio in the dimension corresponding to the set target distribution, calculate the absolute difference between each category ratio value in the initial sample distribution ratio set and the category ratio value corresponding to the set target distribution, and accumulate the absolute differences of each category ratio value to generate a cumulative difference value of the sample distribution in each dimension; Based on the sample distribution cumulative difference value and combined with the sample distribution balance index, the sample distribution cumulative difference value of each dimension is divided by the sample distribution balance index of the corresponding dimension, and the standardized deviation value of each dimension is calculated respectively. The standardized deviation values ​​of all dimensions are added and averaged to obtain the candidate data deviation quantification index.

5. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the data group classification identifier are: Based on the candidate data deviation quantitative index, the proportion of the number of samples in each category in the age dimension and the disease stage dimension to the target distribution is extracted respectively, and the difference between the proportion value of each category and the proportion value of the corresponding category in the target distribution is calculated one by one. The category with a positive difference is defined as the advantage category, and the category with a negative difference is defined as the disadvantage category, thereby generating an initial category deviation identifier; According to the initial category deviation identifier, the category to which each sample belongs is determined one by one. If the category to which the sample belongs is the dominant category, it is marked as a dominant group sample. If the category to which the sample belongs is the disadvantaged category, it is marked as a disadvantaged group sample. The sample is marked one by one to obtain the data group division identifier.

6. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the balanced drug efficacy characteristic data set are: Based on the data group division identifier, the category identifiers and corresponding information entropy values ​​of the dominant group samples are extracted in sequence, the information entropy values ​​are sorted from large to small, and an information entropy threshold is determined. It is determined one by one whether the information entropy value of each dominant group sample is lower than the information entropy threshold, and samples with information entropy values ​​lower than the information entropy threshold are eliminated to generate an information entropy weighted undersampling result set; Extracting the category identifiers and corresponding physiological index values ​​of the disadvantaged group samples one by one based on the information entropy weighted undersampling result set, screening the disadvantaged group samples associated with drug interactions, and performing interpolation calculations with a fixed step size within the range of physiological index values ​​to amplify the samples, thereby generating a physiological index interpolation oversampling result set; Based on the physiological index interpolation oversampling result set, the information entropy weighted undersampling result set and the physiological index interpolation oversampling result set are integrated to generate a balanced drug efficacy feature data set.

7. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the drug effect dose response scatter plot are: Based on the balanced drug efficacy characteristic data set, the corresponding drug dosage characteristic values ​​and drug efficacy result values ​​in each sample record are extracted in sequence, and the drug dosage characteristic values ​​and drug efficacy result values ​​are combined into data pairs according to the sample numbers in a one-to-one correspondence to generate a drug dosage and drug efficacy combination data set; According to the drug dosage-pharmacodynamic combination data set, checking data pairs of drug dosage characteristic values ​​and pharmacodynamic result values ​​one by one, verifying that the drug dosage characteristic values ​​and pharmacodynamic result values ​​in the data pairs are both valid values ​​and are both within a predefined value range, and generating a drug dosage-pharmacodynamic drawing data file; Based on the drug dose and pharmacodynamics plotting data file, each data pair in the file is read one by one, the drug dose characteristic value is plotted as the horizontal axis of the scatter plot, the pharmacodynamics result value is plotted as the vertical axis of the scatter plot, and the scatter data points are numbered in the plotting order to generate a pharmacodynamics dose response scatter plot.

8. The method for predicting drug efficacy trends based on a regression model according to claim 1, wherein: The steps for obtaining the drug efficacy prediction trend curve are: Based on the pharmacodynamic dose-response scatter plot, reading each scattered data point recorded in the pharmacodynamic dose-response scatter plot one by one, using the drug dose value of each scattered data point as an independent variable and the corresponding pharmacodynamic result value as a dependent variable, and sorting all data points in ascending order according to the drug dose value to form a pharmacodynamic dose response value set; Calculating a given drug dose value using a nonlinear regression method based on the set of drug efficacy dose response values; Generate a uniformly distributed drug dose numerical sequence within the drug dose observation range, substitute each dose value in the drug dose numerical sequence into the numerical calculation formula for predicting drug efficacy results one by one, connect all calculated predicted drug efficacy result values ​​and corresponding drug dose values ​​in sequence into a continuous change trajectory, and generate a drug efficacy prediction trend curve.

Images