Drug key quality attribute screening method based on serious adverse reaction data

Abstract

The invention relates to a drug key quality attribute screening method based on serious adverse reaction data, and the method employs the serious adverse reaction data to establish appropriate indexes, and is used for dividing the serious adverse reaction risk levels of different products of the same variety. And a drug key quality attribute screening method based on the adverse reaction data is established from the overall perspective, and key control points of processes related to drugs are explored. The established method can prompt the possible correlation between the quality attribute of the medicine and the occurrence probability of the serious adverse reaction, and provides a direction for further carrying out toxicological research and medicine quality standard research; the method is not influenced by the selection range of quality attributes, only the internal relevance is searched in the existing range, and the result reliability is good; the correlation is used for replacing the causal relationship, the result of logistic regression reveals possible objective phenomena, a further toxicity evaluation direction is given, and the method has guiding significance and application value for improvement of a drug process and improvement of a drug standard.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method for screening key quality attributes of drugs based on severe adverse reaction data. Background Art

[0002] Adverse reaction monitoring refers to the process of real-time, continuous and systematic monitoring, reporting and management of adverse reactions that may occur after the use of drugs. Adverse reactions refer to harmful reactions caused by qualified drugs under normal usage and dosage that are unrelated to the expected drug efficacy. The monitoring of adverse reactions of drugs can be traced back to the early 20th century. At that time, doctors and pharmacists began to realize that certain drugs would cause adverse reactions in patients and began to record and report these adverse reactions. Over time, adverse reaction monitoring has become increasingly important and has gradually become an important part of drug regulatory agencies around the world. Internationally, adverse reaction monitoring is led and guided by the World Health Organization (WHO). The World Health Organization collects, manages and analyzes adverse drug reaction reports worldwide through its international drug monitoring agency (Adverse Drug Reaction Monitoring Center). Drug regulatory agencies in various countries are responsible for domestic adverse reaction monitoring and monitor and report adverse drug reactions in accordance with the guidance and requirements of the World Health Organization. In my country, the National Center for Adverse Drug Reaction Monitoring is responsible for the technical work of reporting and monitoring adverse drug reactions nationwide. Drug manufacturers, drug dealers and medical institutions should establish a drug adverse reaction reporting and monitoring management system in accordance with the requirements of relevant laws and regulations, and assign full-time personnel to be responsible for the reporting and monitoring of adverse reactions. The local drug supervision and administration departments and health administrative departments at all levels are responsible for the management and guidance of adverse drug reaction reporting and monitoring work within their administrative regions, and at the same time, they must establish and improve adverse drug reaction monitoring institutions. In addition, the National Adverse Drug Reaction Monitoring Center is also responsible for the collection, evaluation, feedback, reporting of adverse drug reaction reports and monitoring, as well as the construction and maintenance of the adverse drug reaction monitoring information network.

[0003] At present, the adverse reaction monitoring data of the National Adverse Reaction Monitoring Center is counted by product using the one-case-one-time method. In addition to reporting the total number of adverse reaction data, the adverse reaction data are reported as general cases and serious cases according to the degree of clinical manifestation. The adverse reaction reporting mechanism is a combination of voluntary reporting and mandatory reporting. Among them, serious adverse reactions are generally mandatory reporting mechanisms, so their data integrity and representativeness are more comprehensive than general adverse reactions. Commonly used adverse reaction data analysis methods include the reporting proportion method (PRP), the comprehensive standard method (MHRA), the reporting odds ratio method (ROR), the information component method (IC) and cluster analysis. These methods are all for studying the univariate correlation between adverse reactions and drug varieties. For drug regulatory science, more research is done on the complex multivariate correlation of adverse reaction risks of different categories of drugs and different products of the same variety, which requires big data analysis. At present, the relevant data management system is more about collecting and managing data, and the big data analysis function based on drug quality laws is very scarce.

[0004] The current status of domestic pharmaceutical process development shows poor process control capabilities. The root cause is the lack of understanding of the process and the lack of understanding and attention to the relationship between key quality attributes, key material attributes, and key process parameters. According to the QbD principle, key quality attributes (CQAs) refer to the physical, chemical, biological, and microbiological characteristics of the final product, which should be within appropriate limits or distributions to ensure the quality of the drug.

[61] The quality attributes of drugs include identification, assay, content uniformity, degradation products, residual solvents, drug release or solubility, moisture content, microbial limits, etc. The criticality of an attribute depends mainly on the severity of harm to patients if the attribute exceeds the limit. Therefore, there may be a certain causal relationship between drug quality attributes and adverse reaction data, but this causal relationship is unknown. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for screening key quality attributes of drugs based on serious adverse reaction data.

[0006] The technical solution adopted by the present invention is: a risk level classification method based on severe adverse reaction data, based on severe adverse reaction cases and clinical medication dosage data, and using severe adverse reaction weight coefficients and / or relative severe adverse reaction weight coefficients to measure the risk of severe adverse reactions; The formula is as follows: Formula 1; Formula 2; W sADR RW is the weight coefficient of serious adverse reactions. sADR is the relative serious adverse reaction weight coefficient.

[0007] The drug key quality attribute screening method based on serious adverse reaction data is based on the risk level classification method. The multinomial logistic regression method is used to analyze each quality attribute for each level, and the correlation between the quality attribute and the serious adverse reactions is judged by the probability distribution corresponding to each quality attribute.

[0008] Preferably, the specific steps are as follows: Step 1: Based on the serious adverse reaction data, measure the risk of serious adverse reactions using the serious adverse reaction weight coefficient and / or the relative serious adverse reaction weight coefficient; Formula 1; Formula 2; W sADR RW is the weight coefficient of serious adverse reactions. sADR is the relative serious adverse reaction weight coefficient; Step 2: Select the quality attributes related to the occurrence of serious adverse reactions as the key points of process control, take the quality attributes as the independent variables and the incidence of adverse reactions as the dependent variable, and use the multinomial logistic regression method to mine the key points of process control; according to the risk level divided, perform multinomial logistic regression analysis for the same risk level; (Formula 4); Among them, drug quality attributes QA1, QA2, ...QA m , the probability of serious adverse reactions is p, then the probability of no serious adverse reactions is 1-p; β 0 is a constant, β 1 ，β 2 ，…，β m Quality attributes QA 1 , QA 2 , …, QA m The logistic regression coefficient of Step 3: Pass β j (j=1,2,…,m) The output value of analyzes the contribution of quality attributes between different risk levels. β j for β 1 ，β 2 ，…，β mOne of the quality attributes is used to determine the correlation between the quality attribute and the serious adverse reactions through the probability distribution of the corresponding quality attribute.

[0009] Preferably, in step one, based on the number of serious adverse reactions and clinical drug dosage data, a logistic regression curve of relative serious adverse reaction weight coefficients is established, risk levels are divided according to the serious adverse reaction weight coefficients, and n+1 levels are obtained by dividing the risk levels using n serious adverse reaction weight coefficient division points.

[0010] Preferably, in step 2, the quality attributes include at least two of pH value, polymer content (PIMP), single impurity content, total impurity content (IMPT), moisture value, and API content value based on anhydrous matter; wherein the single impurity content includes a combination of one or more of IMPA, IMPB, IMPD, IMPE and IMPS.

[0011] Preferably, in step 2, the probability p of occurrence of serious adverse reactions for each dividing point value is calculated respectively, and multiple serious adverse reaction-quality attribute logistic regression models are fitted.

[0012] Preferably, in step three, β j The larger the output value of j The more serious adverse reactions occur.

[0013] The advantages and positive effects of the present invention are: defining a "relatively serious adverse reaction weight coefficient" and establishing an adverse reaction analysis method suitable for group analysis, which is used for the classification of different process risk levels under the same variety; and further establishing a screening method for key quality attributes related to adverse reactions based on quality data and adverse reaction data; The established method can indicate the possible association between the quality attributes of drugs and the probability of serious adverse reactions, and provide direction for further toxicological research and drug quality standard research; the method is not affected by the selection range of quality attributes, and only seeks intrinsic correlation within the existing range. The results are reliable, and correlation is used instead of causal relationship. The results of logistic regression reveal possible objective phenomena and provide direction for further toxicity evaluation, which has guiding significance and application value for drug process improvement and drug standard improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Method steps for drug critical quality attribute screening based on serious adverse reaction data; Figure 2 Adverse reaction data of Cefathiamidine for injection from 2015 to 2019; Figure 3 Frequency of severe adverse reactions to Cefathiamidine for injectionsADR RW sADR Relationship diagram; (a): N sADR With RW sADR Relationship diagram; (b): ln(N sADR ) and RW sADR Relationship diagram; Figure 4 R sADR Normal fitting and logistic fitting results; Figure 5 R sADR The cumulative probability distribution and tangent point of ; Figure 6 Normal probability plot of the Pearson fit residuals of the logistic regression model for cefathiamidine for injection. DETAILED DESCRIPTION

[0015] The embodiments of the present invention are described below with reference to the accompanying drawings.

[0016] In the existing statistical methods, the quality of drugs is measured by the proportion of serious adverse reactions in the total batches, among which the data of serious adverse reactions are based on the number of serious cases reported by the manufacturer. Although this frequency of serious adverse reactions (cases / 10,000 units) can reflect the relationship between adverse reactions and manufacturers to a certain extent, it is an absolute value calculated from the data of each manufacturer (process) and is independent. The correlation and comparability between the results of each manufacturer are weak and cannot be directly used to study the correlation analysis between adverse reactions and drug processes. In order to further mine the critical control point information of process correlation, it is necessary to quantitatively or semi-quantitatively characterize the risk level of serious adverse reactions.

[0017] The present invention relates to a risk level classification method based on serious adverse reaction data and a drug key quality attribute screening method. First, serious adverse reaction data are used to establish appropriate indicators for classifying the serious adverse reaction risk levels of different products under the same variety, and then a drug key quality attribute screening method based on adverse reaction data is established from an overall perspective, and the key control points of the drug-related processes are explored. Through adverse reaction data analysis, correlation information mining is carried out with drug quality attributes, and the correlation between adverse reactions and drug key quality control attributes is studied using big data thinking, and key process control point information is screened and obtained. The factors that affect the production of serious adverse reactions of drugs are known through data analysis, which is conducive to the improvement of drug preparation technology, thereby promoting the continuous improvement of drug production quality.

[0018] First, we use the severe adverse reaction data to establish appropriate indicators for the classification of severe adverse reaction risk levels for different products under the same category. The severe adverse reaction ratio and clinical drug use ratio are used to represent the relative value of the severe adverse reaction incidence, which are defined as the weight coefficient of severe adverse reaction (sADR, W sADR ) and relative weight coefficient of sADR (RW sADR ), the formula is as follows: (Formula 1); (Formula 2); Secondly, from an overall perspective, a screening method for drug key quality attributes based on adverse reaction data is established to explore the key control points of the drug process. Specifically, the key quality attributes are explored by seeking the relationship between the probability of serious adverse reactions and the quality attributes of the drug, that is, those quality attributes related to the occurrence of serious adverse reactions (such as certain impurities, pH value, API purity, etc.) are considered to be the key points of process control. Here, the quality attributes may be continuous or discrete, but the risk level of serious adverse reactions is discrete. With quality attributes as independent variables and the incidence of adverse reactions as dependent variables, the polynomial logistic regression method is used to mine the key points of process control.

[0019] The steps of drug critical quality attribute screening based on serious adverse reaction data are as follows: Step 1: Use serious adverse reaction data to establish appropriate indicators for classifying the serious adverse reaction risk levels of different products under the same variety; Step 2: Select the quality attributes related to the occurrence of serious adverse reactions as the key points of process control, take the quality attributes as the independent variables, and the incidence of adverse reactions as the dependent variable, and use the multinomial logistic regression method to mine the key points of process control; specifically, according to the risk level divided, multinomial logistic regression analysis can be performed for the same risk level; (Formula 4); in, β 0 is a constant, β 1 ，β 2 ，…，β m Quality attributes QA 1 , QA 2 , …, QAm The logistic regression coefficient of Step 3: Determine the key quality attributes that cause each level of serious adverse reaction risk through the corresponding probability distribution; specifically, β j (j=1,2,…,m) The output value of analyzes the contribution of quality attributes between different risk levels, where β j for β 1 ，β 2 ，…，β m one of the.

[0020] For step 2, the specific method of mining the key points of process control through the polynomial logistic regression method is as follows.

[0021] Assume that in the drug quality attributes QA1, QA2, …, QA m The probability of a serious adverse reaction under the action of a drug is p, and the probability of no serious adverse reaction is 1-p. The ratio of the probability of a serious adverse reaction to the probability of no serious adverse reaction (p / 1-p) is defined as the odds of sADR (Odds of sADR, Odds of sADR). sADR ), then the logit transformation of p is: (Formula 3); Quality attributes that may be related to serious adverse reactions were selected, and a serious adverse reaction-quality attribute logistic regression model (Formula 4) was established to perform critical screening of quality attributes.

[0022] (Formula 4); Among them, β0 is a constant, β 1 ，β 2 ，…，β m Quality attributes QA 1 , QA 2 , …, QA m In some embodiments of the present invention, the analysis objects are graded according to the relative severe adverse reaction weight coefficients, and n+1 grades are obtained by dividing the n severe adverse reaction weight coefficient division points. Different groups in the multiple grades form multiple model numbers, and the number of models is n.

[0023] Definition: Relative risk of serious adverse reactionssADR The relative sADR risk is the ratio of the incidence of serious adverse reactions at different risk levels to describe the effect of quality attributes on the probability of adverse reactions. sADR It can be approximately estimated as: (Formula 5); Among them, p0 and p1 are the probability of occurrence of serious adverse reactions of different risk levels.

[0024] According to formula 5, we can get: (Formula 6); Then RR sADR With logistic regression coefficient β j The relationship is: (Formula 7); Among them, QA j0 , QA j1 QA j The average level under the probability of serious adverse reactions p0 and p1.

[0025] In step three, you can β j The output value of the QA j The probability of serious adverse reactions corresponding to the risk level of serious adverse reactions p1 is increased by exp(β j ) times. The key quality attributes that cause each level of risk of serious adverse reactions are determined by the corresponding probability distribution.

[0026] R sADR For discrete continuous variables, first according to RW sADR The values ​​of are converted into ordered variables, and then the multinomial logistic regression method is used to screen the quality attributes and obtain a reasonable explanation for the analysis results.

[0027] In order to verify the rationality of the obtained data, the probability distribution of Pearson residuals is used to check the quality of model fitting. The Pearson residual calculation formula is as follows: (Formula 8); Among them, i is the observation probability, e i is the predicted probability, is the estimate of the standard deviation of the predicted probability, and N is the number of independent samples multiplied by the number of models n.

[0028] Since the Pearson residuals are the original residuals normalized by the estimated standard deviation and have the same variance, they are roughly normally distributed when the model fits the data reasonably. The present invention uses the normal probability plot of the Pearson residuals to compare the distribution of the Pearson residuals with the normal distribution to check whether they follow the normal distribution. The established method is verified using the adverse reaction data of each manufacturer. The serious adverse reaction weight RW is calculated using a multinomial logistic regression model based on the average quality attribute level of each manufacturer. sADR The distribution range is predicted.

[0029] This method analyzes adverse reaction data and then mines the correlation information with drug quality attributes, and uses big data thinking to establish a method for obtaining key process control point information based on correlation research. It is a group analysis method for the variety-manufacturer model. The manufacturer group is representative enough and is less affected by the size of individual samples. Big data thinking can be used for the analysis of group data of any size. Adverse reactions are direct evaluation indicators of drug quality and safety. The "relative serious adverse reaction weight coefficient" is defined and an adverse reaction analysis method suitable for group analysis is established for the classification of different process risk levels under the same variety; and then a screening method for key quality attributes related to adverse reactions based on quality data and adverse reaction data is established. The established method can indicate the possible association between the quality attributes of drugs and the probability of serious adverse reactions, provide direction for further toxicological research and drug quality standard research, and have guiding significance and application value for drug process improvement and drug standard improvement.

[0030] The scheme of the present invention is described below in conjunction with specific embodiments.

[0031] Example: Take the statistical data of various adverse reactions of cefathiamidine for injection from 2015 to 2019 as an example (such as Figure 2 ), according to the classification statistics of the involved systems, among which skin reactions such as rash and itching account for the majority, and some gastrointestinal reactions such as nausea, vomiting, abdominal pain, diarrhea, and cardiovascular and cerebrovascular reactions such as chest tightness, dizziness, and palpitations are also seen. The adverse reaction reporting mechanism is a combination of voluntary reporting and mandatory reporting, among which serious adverse reactions are generally mandatory reporting mechanisms, so their data integrity and representativeness are more comprehensive than general adverse reactions. Therefore, the present invention uses serious adverse reaction data for correlation analysis. The serious cases in the serious adverse reaction data are based on the corresponding cases reported by the manufacturer.

[0032] 1. Use the adverse reaction data of cefathiamidine for injection from 2015 to 2019 to establish indicators and classify the risk levels of serious adverse reactions The data on the number of serious adverse reactions and the number of clinical drugs used by different manufacturers are shown in Table 1. W is calculated according to Formula 1 and Formula 2. sADR and RW sADR , calculate the frequency of serious adverse reactions N sADR , see Table 1 for details.

[0033] Table 1 Cases and frequencies of serious adverse reactions to cefathiamidine injection, and weights of serious adverse reactions

[0034] Weight coefficient W of severe adverse reactions of cefathiamidine for injection sADR The results are shown in Table 1. The severe adverse reaction coefficient of Cefathiamidine for injection produced by M3 is the smallest, which is -0.19; the severe adverse reaction coefficient of Cefathiamidine for injection produced by M1 is the largest, which is 1.97. The results are consistent with the characterization method of severe adverse reaction frequency. The relationship between severe adverse reaction frequency and severe adverse reaction weight coefficient is shown in Figure 3 As shown in (a), W sADR The value of is evenly distributed in the range of [-0.5, 2.0], indicating that W sADR The frequency of more serious adverse reactions is more conducive to the definition of the level of adverse reaction risk and further analysis; Figure 3 (b) It can be seen that ln(N sADR ) and W sADR There is a linear relationship with a correlation coefficient of 0.9999. Data fitting obtained N sADR , W sADR and RW sADR The relationship is: (Formula 9); (Formula 10); (Formula 11); 2. Using the 2019 national evaluation sampling data, combined with the risk analysis of adverse reactions of cefathiamidine for injection in the past five years, the multinomial logistic regression method was used to explore the key points of process control Cefathiamidine for injection RW sADR As dependent variables, quality attributes (QA) such as pH value, polymer content (PIMP), individual impurity content (IMPA, IMPB, IMPD, IMPE, IMPS (other maximum)), total impurity content (IMPT), moisture value, and API content value based on anhydrous matter were selected as independent variables. These quality attributes were screened to find influencing factors related to serious adverse reactions, and then the degree of influence of these factors on serious adverse reactions was analyzed.

[0035] Continuous variable RW sADRThe normal fitting and logistic fitting results are as follows Figure 4 As shown, the normal fitting has a kurtosis of 0.6460, a skewness of 0.4198, and a variance of 0.1762; the logistic fitting has a kurtosis of 0.6888, a skewness of 0.2292, and a variance of 0.1728. It can be seen that the logistic fitting and the normal fitting are highly consistent, and the cumulative probability distributions of the two are basically the same, as shown in Figure 5 As shown in Figure 2, three “cut points” C1, C2, and C3 are obtained at cumulative probabilities of 30%, 60%, and 90%, respectively, and the corresponding RW sADR The values ​​are 0.426, 0.752, and 1.184, respectively, dividing the distribution into four risk areas for logistic regression classification: RW sADR ≤0.462 is level 1, 0.462<RW sADR ≤0.752 is level 2, 0.752<RW sADR ≤1.184 is level 3, RW sADR >1.184 is level 4.

[0036] Based on the national evaluation sampling data in 2019, the quality attribute (QA) risk area classification of pH value, polymer content (PIMP), individual impurity content (IMPA, IMPB, IMPD, IMPE, IMPS (other maximum)), total impurity content (IMPT), moisture value, and API content value was obtained, as shown in Tables 2 and 3.

[0037] Table 2

[0038]

[0039]

[0040]

[0041] Table 3 Cefathiamidine for injection W ADR , risk level and quality attributes

[0042] A multinomial logistic regression model was fitted using the quality attributes and adverse reaction data of 148 batches of Cefathiamidine for Injection: like Figure 6As shown in the figure, the fitting residuals of the adverse reaction-quality attribute model basically conform to the normal distribution, and there are obvious differences in the distribution trends of each risk level. Therefore, the constructed model has a good fit and can be used for the screening and analysis of quality attributes. The model has 406 degrees of freedom and an overall fitting deviation of 262.40, which is relatively close to the chi-square distribution of the 95% confidence interval. , the model is significant.

[0043] The model parameters are shown in Table 4. The coefficient estimate of IMPE is the largest, which is 16.21, followed by pH value 7.87 and PIMP 2.43. The p-value matrix results of the significance test show that p IMPE 、p pH 、p PIMP All of them were less than 0.05, indicating that IMPE, pH value, and PIMP all had significant effects on the risk of serious adverse reactions. Therefore, IMPE, PIMP, and water content were significantly correlated with the incidence of serious adverse reactions.

[0044] Table 4 Parameters of adverse reaction-quality attribute logistic regression model

[0045] The adverse reaction data of each manufacturer were used to verify the established method. According to the average quality attribute level of each manufacturer in Table 3, the adverse reaction-quality attribute model was used to weight the serious adverse reaction RW. sADR The prediction results and their 95% confidence intervals are shown in Table 5. The results show that except for M6, the prediction results of other manufacturers are consistent with the actual RW sADR The prediction results of M6 show that there is a 10% probability that it is consistent with the risk level in Table 3, and a 26.7% and 58.5% probability of risk level 2 and 3, respectively, which reveals that this manufacturer has the potential to develop into a higher risk, and its serious adverse reaction data should continue to be paid attention to.

[0046] Table 5 Prediction results of adverse reaction-quality attribute model and their 95% confidence intervals

[0047] According to the adverse reaction-value attribute model, IMPE, pH value, and PIMP correspond to the relative risk RR of serious adverse reactions sADR They are 1.098E+07, 220.8 and 86.91 respectively, that is, when all other conditions remain unchanged, with the increase of IMPE, pH value and PIMP, the possibility of serious adverse reactions increases by 1.098E+07 times, 220.8 times and 86.91 times respectively. Therefore, attention should be paid to the three key process control points of impurity E, pH value and polymer, and the limits of impurity E and polymer should be controlled, and the pH value should be controlled within a reasonable range.

[0048] The embodiments of the present invention are described in detail above, but the contents described are only preferred embodiments of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A risk classification method based on severe adverse reaction data, characterized in that: Based on the number of serious adverse reactions and clinical drug dosage data, the risk of serious adverse reactions is measured by the serious adverse reaction weight coefficient and / or the relative serious adverse reaction weight coefficient; The formula is as follows: Formula 1; Formula 2: W sADR RW is the weight coefficient of serious adverse reactions. sADR is the relative serious adverse reaction weight coefficient.

2. A method for screening drug key quality attributes based on serious adverse reaction data, characterized in that: Based on the risk level classification method described in claim 1, the polynomial logistic regression method is used to analyze each quality attribute for each level, and the correlation between the quality attribute and the serious adverse reactions is judged by the probability distribution corresponding to each quality attribute.

3. The method for screening drug key quality attributes based on serious adverse reaction data according to claim 2, characterized in that: The specific steps are as follows: Step 1: Based on the serious adverse reaction data, measure the risk of serious adverse reactions using the serious adverse reaction weight coefficient and / or the relative serious adverse reaction weight coefficient; Formula 1; Formula 2: W sADR RW is the weight coefficient of serious adverse reactions. sADR is the relative serious adverse reaction weight coefficient; Step 2: Select the quality attributes related to the occurrence of serious adverse reactions as the key points of process control, take the quality attributes as the independent variables and the incidence of adverse reactions as the dependent variable, and use the multinomial logistic regression method to mine the key points of process control; according to the risk level divided, perform multinomial logistic regression analysis for the same risk level; (Formula 4) Among them, drug quality attributes QA1, QA2, ...QA m , the probability of serious adverse reactions is p, then the probability of no serious adverse reactions is 1-p; β 0 is a constant, β 1 ，β 2 ，…，β m Quality attributes QA 1 , QA 2 , …, QA m The logistic regression coefficient of Step 3: Pass β j The output value of analyzes the contribution of quality attributes between different risk levels. β j for β 1 ，β 2 ，…，β m One of the quality attributes is used to determine the correlation between the quality attribute and the serious adverse reactions through the probability distribution of the corresponding quality attribute.

4. The method for screening drug key quality attributes based on serious adverse reaction data according to claim 3, characterized in that: In step one, based on the number of serious adverse reactions and clinical drug dosage data, a logistic regression curve of the relative serious adverse reaction weight coefficient is established, and the risk level is divided according to the serious adverse reaction weight coefficient. n+1 levels are obtained by dividing the n serious adverse reaction weight coefficient division points.

5. The method for screening drug key quality attributes based on serious adverse reaction data according to claim 3, characterized in that: In step 2, the quality attributes include at least two of pH value, polymer content PIMP, single impurity content, total impurity content IMPT, moisture value, and API content value based on anhydrous matter; wherein the single impurity content includes a combination of one or more of IMPA, IMPB, IMPD, IMPE and IMPS.

6. The method for screening drug key quality attributes based on serious adverse reaction data according to claim 3, characterized in that: In step 2, the probability p of occurrence of serious adverse reactions at each dividing point value is calculated respectively, and multiple serious adverse reaction-quality attribute logistic regression models are fitted.

7. The method for screening drug key quality attributes based on serious adverse reaction data according to claim 3, characterized in that: In step three, β j The larger the output value of j The more serious adverse reactions occur.

Images