Ranking Drugs in Spontaneous Reporting System by Naive Bayes
A. Bazila Banu,
S. Appavu Alias Balamurugan
P. Thirumalaikolundu Subramanian
In this study detection of association between drugs and Adverse Drug Reactions (ADRs), is carried out by using Naive Bayes method. Adverse event reports submitted to the United States Food and Drug Administration (FDA) were reviewed to find top 10 drugs causing frequent ADRs for a particular period. The main objective of this paper is to evaluate drugs associated with list of outcomes provided by FDA. For a particular category of disease, drugs creating outcomes are ranked using Naive Bayes method. FDA represents ADRs in Preferred Terms(PT) by referring Medical Dictionary for Regulatory Activities (MedDRA).To create conceptual hierarchy System Organ Class (SOC) present in MedDRA is mapped with low level Preferred Terms (PT) in FDA dataset. For each SOC the drugs are ranked based on posterior probability obtained by Naive Bayes method. Data mining model has been built to analyse drugs associated with outcome for a disease category in SOC level. The newly designed tool is user friendly and applicable to pharmaceutical industries, policy makers and practitioners.
Received: March 13, 2013;
Accepted: April 04, 2013;
Published: July 10, 2013
Adverse Drug Reaction (ADR) is defined as all toxic and inadvertent response
to a drug with any dosage level. ADRs remain an important health issue to be
investigated in earlier stage to avoid accidental effects. There are two types
of ADR.Type A or dose related creates highest percentage of adverse
reactions than Type B reactions (Talbot and Waller,
2005). Type A reactions are considered for this study. To detect
ADR signals, data mining methods like Proportional Reporting Ratio and Naive
Bayes are used in the Spontaneous Reporting System (SRS) (Wang
et al., 2011; Hauben et al., 2007).
United States Food and Drug Administration (US FDA) plays an important role
in pharmacovigilance activity especially in providing drug safety. Since 1969
FDA has maintained a computerized repository for storing and retrieving all
ADR reports (USFDA, 2011).This repository can also be
referred as SRS. It stores all reported ADRs by adopting the protocols of standardized
international terminology, Medical Dictionary for Regulatory Activities (MedDRA).It
represents disease names in Preferred Terms (PT) level (Pearson
et al., 2009). MedDRA is a medical dictionary for describing adverse
events, with five levels: the highest level is System Organ Class(SOC), followed
by High Level Group Term (HLGT), Higher Level Term (HLT), Preferred Term (PT)
and Lowest level Term (LLT) (Henegar et al., 2006).
List of outcomes as given by FDA like Life-Threatening (LT), Hospitalization
- Initial or Prolonged (HO), Disability (DS), Congenital Anomaly (CA), Required
Intervention to prevent permanent impairment/damage (RI), Death (DE) are ranked
by Naive Bayes method based on the attributes drug, category of the diseases.
In data mining, supervised learning methods such as decision trees, association
rules, Naive Bayes and neural networks are used in the field of ADR to analyse
the reports (Chazard et al., 2011). The Naive
Bayes classifier (NB) is one among the statistical classifier used to predict
class membership probability. It detects the class based on the maximum probability
obtained for the given tuple to a particular class. It has been widely used
by researchers for classification. It assumes all variables participating in
the classification as independent and produces good results for prediction.
Zhang and Su (2004) have proved that apart from classification
Naive Bayes can be used for ranking and it outperforms other traditional decision
tree algorithms like c4.5. The limitations intrinsic in SRS like FDA are studied
by Sakaeda et al. (2011a, b).
Aim of this study is to design an user friendly tool to be used by pharmaceutical
industries and practitioners to analyse ADRs based on disease category.
To identify the drugs causing outcomes of an ADR, this paper aims at ranking
drugs associated with list of outcomes for a particular disease category. To
design a user friendly tool for analyzing ADRs.To Map the adverse event representation
of FDA from PT level to MedDRA SOC level. Researchers suggested that it may
be more beneficial to perform data mining, using highest level adverse event
representation like SOC than the MedDRA PT level (Pearson
et al., 2009). The depth of the PT level is high and hence difficult
to evaluate the disease associated with particular outcome. Mapping of PT with
SOC will produce better clarity in evaluating the drug outcome association.
Klementiev et al. (2007) has proposed an unsupervised
learning algorithm for rank aggregation. Zhang et al
(2005) has done extensive work in applying Naive Bayes for ranking. He has
proposed a method named as Augmenting Naive Bayes, suitable for applying in
limited training data. Jiang et al. (2005a, b)
has proposed K-nearest neighbor Naive Bayes for ranking, however the performance
is unknown for ranking items. Jiang et al. (2005a)
also proposed Tree Augmented Naive Bayes for ranking, but the performance is
poor when compared to Naive Bayes.
MATERIALS AND METHODS
ADRs in FDA 2011 are considered as case set for the study. PTs used in FDA
are mapped with SOC of MedDRA by referring cancer therapy evaluation program
simplified disease classification v4.0 (MedDRA v 12.0) (National cancer institute,
United States). Data is extracted from SRS provided by FDA. Then the duplicate
reports were deleted according to FDAs suggestion of using recent case
number as described in the file Asc-nts.doc from the website of
the FDA. SRS such as US FDA provide an opportunity to study drugs causing side
effects. Once the schema is designed and the associated database is constructed,
the data is loaded from FDAs text file to oracle database, using ETL (Extract,
Transform and Load) tools. Indices are constructed using patient identifier.
Xml Mapping is created by constructing SOC as the root node and PT as the child
nodes. Based on XML Sql Utility (XSU), the root nodes are injected into the
database for the equivalent PTs to create disease categorical mapping. Figure
1 represents the hierarchical design of XML for SOC and PT. XML Sql Utility
(XSU) in oracle 11 g is used to extract the data from an x mL document and to
inject the data in to the database (Yu and Stahnikam, 2011).
First step is to establish Java Database Connection (JDBC) then an instance
of OracleXmlQuery is created to pass sql query which contains the table name
and column name to be updated from xml data. Next step is to obtain Document
Object Model (DOM) by calling get method of OracleXmlQuery.XSU converts the
elements to sql types and binds them to the appropriate statement. At this step
the disease category of SOC level is obtained by mapping the PT of FDA with
MedDRA PT. This step creates concept hierarchy. For ranking the drugs based
on outcome Naive Bayes method is used because it is based on the assumption
of class conditional independence of all attributes. If a learning algorithm
is able to estimate accurate class probability, it certainly produce precise
ranking (Gouthami et al., 2012).
Data mining model
Schema: Description of database schema is shown in Fig. 2.
The central entity is that of a patient. Patient is uniquely identified by ISR
(unique number for identifying an AERS report). For mapping PT with SOC, each
ISR is assigned to only one SOC. All other data in the schema is adverse event
|| Hierarchical design of XML for SOC and PT
|| Description of database schema
For data mining phase, the data have to be simplified. For instance the duration
and dose of medications have been ignored. Three distinct entities considered
are drug, preferred term and outcome.
Data mining prototype: Computational steps for data storage and retrieval
is shown in Fig. 3. A web based system has been designed to
analyse the various levels of disease categories for a particular period. First
step of Fig. 3 describes how the data from FDA are transformed
into analytical schema using ETL workflows. The data presented in the. txt format
are transformed and loaded in to the schema by using SQL Loader. PTs used in
FDA are mapped with SOC of MedDRA to create ADR schema.
Implementation: Naive Bayes algorithm is used to obtain the top 10 drugs
for each year based on the attributes drug, category and the outcome. Using
Naive Bayes, model can be built with different prior probability assumptions.
The outcome of the adverse events is considered as major class and the other
attributes like category and drugs are considered as subset for ranking. First,
the fundamental assumption of attribute independence is considered for this
study. Naive Bayes theorem given in formula 1 is used to calculate the probability
of an outcome:
where, P(H), the probability that the hypothesis H holds for the observed data
tuple X. This is the prior probability that any patient may get an outcome regardless
of the drug and category. The posterior probability P(H/X) is based on more
patient information like drug and disease category. All the possible outcomes
given in FDA were considered for the P(H/X). Where, X = (Disease Category =
Gastrointestinal disorders |Any of 26 disorders, Code of drug=1|2, outcome =
LT| HO| DS| CA| RI |DE).Code 1 denotes valid trade name of the drug and 2 denotes
verbatim name of the drug.
|| Computational steps for data storage and retrieval
The outcomes LT, HO, DS, CA, RI and DE are used for computing the values of
P(X), where P(X) denotes prior probability of X. Based on the outcome, the probabilities
of category and drug are evaluated. By using the posterior probabilities for
corresponding category and outcome, the top 10 drugs are listed for ranking.
The algorithm applies disease category as a variant. Disease categories are
chosen for an outcome and the posterior probabilities are estimated for ranking
The experimental result helps the medical practitioners to identify the drugs
creating particular outcome based on the category of the disease in the SOC
level. It is tedious to evaluate the outcomes based on PT level. Posterior probability
obtained from Naive Bayes is used to rank the drugs .Among all the drugs Aspirin
occur in four outcomes like DE, DS, HO and LT. It has the probability of 1 for
HO. Lasix is the second drug to have three outcomes like HO, LT, DE. It produces
the probability of .6158(HO), .0009(LT), .1410(DE). Drugs like Humira and Remicade
occur in multiple outcomes like HO and LT with probability of .8908(HO), .7598(HO),
.0006(LT), .0009(LT).Drugs like Yaz and Yasmin occur in HO with probability
of, .9142 and .9216. Heparin Sodium Injection and Revlimid occurs in multiple
outcomes with less probability of .0008(LT), .0009(LT), .0309(DE), .0181(DE).Other
drugs like Trasylol, Dianeal, Avandia, Tracleer, Acetaminophen, Heparin, Dexamethasone
and Coumadin produces DE alone. Outcomes like CA and RI are very less. Drugs
like Zolosoft have the probability of .000006(CA) and Actos have the probability
of .00013(RI). List of SOCs are given in Table 1.
|| Representation of disease categories interns of SOC level
|SOC: System organ class
Table 2-5 represent the ranking of the
drugs for the disease category Gastrointestinal Disorder, with respect to outcomes
like HO, LT, DE and DS. As per the results of Table 2 and
3, it is observed that Aspirin holds the 1st rank to cause
Gastrointestinal Disorder with outcome as HO and LT. For outcome HO, drugs like
Yasmin, Yaz and Humira acquire 2nd , 3rd and 4th ranks with probability greater
||Top 10 drugs for the period 2011(Q1-Q4) with outcome as HO
in gastrointestinal disorder
||Top 10 drugs for the period 2011(Q1-Q4) with outcome as LT
in gastrointestinal disorder
|LT: Life threatening
||Top 10 drugs for the period 2011(Q1-Q4) with outcome as DE
in gastrointestinal disorder
||Top 10 drugs for the period 2011(Q1-Q4) with outcome as DS
in gastrointestinal disorder
For outcome, LT drugs like Prednisolone, Revlimid, Lasix acquire 2nd, 3rd and
4th ranks with minimal probability. Table 4 represents one
of the severe outcomes DE and the drug Revlimid acquires the 1st rank. Other
drugs like Trasylol, Dianeal and Aspirin acquires 2nd, 3rd and 4th ranks with
nominal probability. Table 5 represents hazardous outcome
DS and the drug Metoclopramide acquires the 1st rank and Asprin acquires the
4th position but the probability is nominal.
Detection of ADR depends on FDAs assessment of patient cases. Determining
precise knowledge from large volume of medical prescription dataset has come
to realism for decision making as well as to avoid hazards in the drugs (Xiuzhen
et al., 2011). Post marketing surveillance system is responsible
for providing drug safety. As a result it may necessitate years to identify
difficult drugs from the promoters (Ji et al., 2011).
A list of outcomes and their link between drug and categories is used for analysis.
In this paper 6 outcomes enable us to trace ADRs. Among the 26 categories listed
in the Table 1, Disease Category Gastrointestinal Disorder
(SOC) is taken to evaluate the six outcomes. The drugs are ranked based on the
probabilities obtained by Naive Bayes method. Performing causality assessment
in pharmacovigilance may help decision making in single ADR pertaining to the
In this study we presented our knowledge of designing an user friendly environment
for conducting Adverse Drug Reaction studies (ADR) based on mining large scale
primary care database FDA. We have used Naive Bayes theorems posterior
probabilities to rank the drugs. This approach serves as a reusable environment
for ranking drugs based on disease categories and outcomes. In this paper the
calculations of Naive Bayes theorem are presented for ADR study. In Future the
SOC level of MedDRA can be further tuned to lower levels like HLGT and HLT for
analysis and decision making at different levels of disease classification.
Chazard, E., G. Ficheur, S. Bernonville, M. Luyckx and R. Beuscart, 2011. Data mining to generate adverse drug events detection rules. IEEE Trans. Inf. Technol. Biomed., 15: 823-830.
Gouthami, S., G.J. Mary and P.S. Rao, 2012. Ranking popular items by naive Bayes algorithm. Int. J. Comput. Sci. Inf. Technol., 4: 147-163.
Direct Link |
Hauben, M., S. Horn, L. Reich and M. Younus, 2007. Association between gastric acid suppressants and Clostridium difficile colitis and community-acquired pneumonia: Analysis using pharmacovigilance tools. Int. J. Infect. Dis., 11: 417-422.
Henegar, C., C. Bousquet, A. Lillo-Le Louet, P. Degoulet and M.C. Jaulent, 2006. Building an ontology of adverse drug reactions for automated signal generation in pharmacovigilance. Comput. Biol. Med., 36: 748-767.
CrossRef | PubMed |
Ji, Y., H. Ying, P. Dews, A. Mansour, J. Tran, R.E. Miller and R.M. Massanari, 2011. A potential causal association mining algorithm for screening adverse drug reactions in postmarketing surveillance. IEEE Trans. Inform. Technol. Biomed., 15: 428-437.
Jiang, L., H. Zhang and J. Su, 2005. Learning k-nearest neighbor naive Bayes for ranking. Proceedings of the 1st International Conference on Advanced Data Mining and Applications, July 22-24, 2005, Wuhan, China, pp: 175-185.
Jiang, L., H. Zhang, Z. Cai and J. Su, 2005. Learning tree augmented naive Bayes for ranking. Proceedings of the 10th International Conference on Database Systems for Advanced Applications, April 17-20, 2005, Beijing, China, pp: 688-698.
Klementiev, A., D. Roth and K. Small, 2007. An unsupervised learning algorithm for rank aggregation. Proceedings of the 18th European Conference on Machine Learning, September 17-21, 2007, Warsaw, Poland, pp: 616-623.
Pearson, R.K., M. Hauben, D.I. Goldsmith, A.L Gouldf and D. Madigan et al., 2009. Influence of the MedDRA® hierarchy on pharmacovigilance data mining results. Int. J. Med. Inf., 78: e97-e103.
Sakaeda, T., K. Kadoyama and Y. Okuno, 2011. Adverse event profiles of platinum agents: Data mining of the public version of the FDA adverse event reporting system, AERS and reproducibility of clinical observations. Int. J. Med. Sci., 8: 487-491.
CrossRef | Direct Link |
Sakaeda, T., K. Kadoyama and Y. Okuno, 2011. Statin-associated muscular and renal adverse events: Data mining of the public version of the FDA adverse event reporting system. PLoS ONE, Vol. 6. 10.1371/journal.pone.0028124
Talbot, J. and P. Waller, 2005. Stephens Detection of new Adverse Drug Reactions. 5th Edn., John Wiley and Sons, New York., USA., pp: 173-175.
USFDA, 2011. FDA adverse event reporting system. U.S. Food and Drug Administration. http://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Surveillance/AdverseDrugEffects/default.htm.
Wang, W., K. Haerian, H. Salmasian, R. Harpaz, H. Chase and C. Friedman, 2011. A drug-adverse event extraction algorithm to support pharmacovigilance knowledge mining from pubmed citations. Proceedings of the Annual Symposium on Improving Health: Informatics and IT Changing the World, October 22-26, 2011, Washington, DC., USA., pp: 1464-1470.
Xiuzhen, F., H. Xiaohong and F. Bianling, 2011. A study on application of multidimensional association rule mining in adverse drug reactions. Energy Procedia, 11: 1720-1726.
Direct Link |
Yu, T. and B. Stahnikam, 2011. Developing XML applications with oracle XML DB and Oracle XML developer's kit. http://www.oracle.com/technetwork/Database/features/xmldb/s311509-developingxmlapplicationswi-132942.pdf.
Zhang, H. and J. Su, 2004. Naive Bayesian classifiers for ranking. Proceedings of the 15th European Conference on Machine Learning, September 20-24, 2004, Pisa, Italy, pp: 501-512.
Zhang, H., L. Jiang and J. Su, 2005. Augmenting naive Bayes for ranking. Proceedings of the 22nd International Conference on Machine Learning, August 7-11, 2005, Bonn, Germany, pp: 1020-1027.