Signal Detection Of Adverse Drug Reaction Based On Spontaneous Reporting System And Evidence-based Medicine

Background: With the development of biomedicine, the safety profile of drug is attracting more and more attention. The drug safety issues in recent years damaged the health of human being severely. According to a statistic from World Health Organization (WHO), the patients due to adverse drug reaction (ADR) accounts for 5-10% of the inpatients. Meanwhile, 10-20% of all the inpatients suffer from ADR and the mortality is as high as 0.24-2.9%. ADR poses great challenge to the public health. Pharmacovigilance (PV) is defined as the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or any other drug-related problem by WHO. Signal is defined as reported information on a possible causal association between an adverse event and a drug.By the end of 2002, the monitoring system for ADR has been established in China. However, the statistical function does not perform well enough and the function of automated signal detection is still lacking. A great deal of issues needs to be solved greatly. It is of great importance to analyze the ADR data and establish effective data mining algorithms for drug risk assessment so as to detect the potential harmful drugs and prevent the harms timely and accurately.Aim: This thesis aims to detect ADR signals from Shanghai spontaneous reporting system by data mining algorithms and develop a computerized system to help this process automatically.Currently several studies have explored the relationship between proton pump inhibitors and hip fracture. Conflicting results from the original studies needs further evidence-based medicine methods to confirm this relationship.Methods: Association rule is one of the most classic data mining algorithms, with a power to handle the incomplete data and discover the hidden patterns. It could help the researchers to understand and further analyze the data. We screened the interesting indices by different rules so as to further explore the possible relationships, such as drug-ADR combination, drug-drug-ADR combination, drug-multiple ADR combination.Given the fact that there is no golden standard for evaluating ADR, we employed Monte Carlo method to simulate the data for comparing the sensitivity and specificity of different data mining algorithms. Taking the chi-square test as the golden standard statistical method, we compare different algorithms, such as proportional odds ratio, proportional reporting ratio, bayesian confidence propagation neural network, association rule. We compared their sensitivity and specificity respectively by receiver operating curve. Meanwhile, we explored the feasibility of association rules to detect signals due to drug-drug interaction and multi-level relationship between drug and ADR.We downloaded the reports from Shanghai spontaneous reporting system in 2009 and set up ADR database for analysis. We cleaned the database and investigated its characteristic, and later we tried to detect the possible ADR signals by association rules. We also developed a web-based system for signal detection by software Visual Studio 2008 and SQL Server 2005.Evidence-based medicine is a subject following scientific evidence and develops rapidly in recent years. Meta analysis could improve the power of studies, enhance the credibility and objective, and solve the discrepancy in studies. Individual pharmacoepidemiology study could not get the credible conclusion due to the low incidence of ADR. We searched different digital libraries and investigated the relationship between proton pump inhibitors and hip fracture by meta analysis.Results: The study summarized the basic theory, main algorithms and interesting measure rules of association rules. We found that minimum support, minimum confidence and lift were key indices for signal generation. The threshold of these indices could influence the number of signals. Lower threshold could lead to more rules, with a high sensitivity and low specificity, and vice versa.Lift is the most important index in the process of ADR signal detection. We observed that the principle of lift is similar to the proportional reporting ratio in disproportional methods. We also found that proportional reporting ratio method employed not only the point estimate but also 95% confidence interval to generate signals, taking the intensity and variability in the sampling process. We elevated the lift threshold and took 1.2 as the threshold. Though the area under the receiver operation curve was less than other algorithms, it was greater than 0.7, reaching the qualification of diagnosis methods. Considering the advantage of handling large database, we believed that association rule algorithms could be used in the process of signal detection.Altogether 30105 reports were downloaded from Shanghai spontaneous reporting system in the year 2009 and 40407 reports were observed after cleaning for further analysis. We found several signals by association rules, such as abatacept- herpes zoster, trastuzumab- pericardial effusion, prednisone & abatacept- herpes zoster, erythropoietin & peritoneal dialysis- eosinophilia.In addition, we also developed a web-based system for signal detection, with a specific website (http://statadr.smmu.edu.cn). Any computer with the access to the internet can use this system. This system could display the data mining results by statistical figures and tables in a vivid way.9 studies met the inclusion criteria. PPI therapy was associated with clinically and statistically risk of hip fracture (pooled OR=1.24, 95%CI 1.15-1.34, p<0.00001) under a random model. Meanwhile, we found that the effect of PPIs on hip fracture differs in different duration groups.Conclusions: Association rule, as a data mining algorithms, could be applied in the ADR signal detection with its unique advantage. The web-based system could detect signals timely and protect the health of the human beings.The meta-analysis indicated that the proton pump inhibitor therapy increase the risk of hip fracture. PPI should be cautiously prescribed in clinical use with proper dosage. Different effects on hip fracture in the subgroup analysis do not support a causal relationship between PPIs and hip fracture. Whether the risk exists warrants further investigation.