kalamri2 / faers-and-cms-data-mining

FAERS and CMS Data Mining

Python and Bash pipeline to extract and clean data for analysis

The objective is to identify adverse drug events (ADEs) in the CMS claims data with the aid of information from the US Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) database. This effort will evaluate the utility of CMS claims data for identifying potential adverse drug events and if successful will improve the timeliness of ADE detection. Some of the studies such as detecting association between Statin Use and Cancer (Mai Fujimoto, 2015), Retrospective Detection of Drug Safety Signals and Adverse Events in Electronic General Practice Records (Andrew Tomlin, 2012), Drug safety surveillance using de-identified EMR and claims data (PM, 2010) and Leveraging Food and Drug Administration Adverse Event Reports for the Automated Monitoring of Electronic Health Records in a Pediatric Hospital (Tang H, 2017) present several approaches that can be used.

Data Used

• CMS 2008-2010 Data Entrepreneurs’ Synthetic Public Use Files (DE-SynPUF) for Beneficiary Summary, Inpatient Claims and Prescription Drug events downloaded from https://www.cms.gov/Research-Statistics-Data-and-Systems/Downloadable-Public-Use-Files/SynPUFs/DE_Syn_PUF.html
  The pdf Centers for Medicare and Medicaid Services (CMS) Linkable 2008–2010 Medicare Data Entrepreneurs’ Synthetic Public Use File (DE-SynPUF) has information regarding the files, data elements and their descriptions.
• FAERS quarterly ASCII files with demographic, drug and reaction information for the years 2014 – 2017 downloaded from the following link. https://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Surveillance/AdverseDrugEffects/ucm082193.htm The pdf ASC_NTS has information regarding file and data element descriptions.

Methods

The CMS files are in csv format and the FAERS are "$" delimited text files. The preprocessing is done using a sequence of bash commands. These are included in cms_cmds.txt and faers_cmds.txt. In brief, preprocessing for each category includes

• Extracting the necessary features
• Merging the individual files, for example (all drug files to one final drug file)
• Sorting and eliminating duplicates (row wise)
• Joining the relevant files to get one file each in CMS and FAERS
• Sample files cms_diab_100.csv and faers_100.csv are in the data folder.

Processing for individual diseases

The NDC codes list for different disease consitions can be obtained from data.gov. ndcdisease.csv contains the diabetes NDC codes. The CMS master file (output of cms_cmds.txt) is processed using cms_diab_cmds.txt. This gives a sunset of CMS master file, that has beneficiaries who have been prescribed at least one of the diabetes drugs from the data.gov diabetes NDC codes list. The same process can be followed to get data corresponding to several other conditions like, hypertension, depression, anxiety etc. The CMS file that results from this process and the final FAERS file become input to the next step.

Creating files for data analysis

These files are then read in python via pandas. The code/pipeline is in preprocess_cms.py and preprocess_faers.py. The goal here is to match the two files on drug codes. The drug information in CMS/PDE file is represented as NDC codes whereas in FAERS, text strings are used to represent drug names. The conversion of NDC codes into Rx concepts/rxcui is achieved as below,

• For CMS data, the api used is NDC/RXNORM Transformer (https://ndc.terminology.tools/) from West Coast Informatics LLC. The code/pipeline is in preprocess_cms.py. A unique list of NDC codes is created and multiple api calls are made to get the corresponding rxcuis. The output of this pipeline is a csv file containing beneficiary id and rxcuis.
• For FAERS data, the api used is Rxnav Approximate Matching is used. More information is available in https://rxnav.nlm.nih.gov/RxNormApproxMatch.html. The code/pipeline is in preprocess_faers.py. From FEARS data,the dosage and unit information is extracted from dose_vbm column using regex. This information is concatenated with drugname and then api calls are made to the corresponding rxcui. The output is a csv file that has primaryid and corresponding rxcuis. A sample output file, final_faers.csv is in the data folder.

Data Analysis

• The two files can be joined to get matching patients with respect to drug combinations using bash join commands.
• This file can then be input to any downstream analysis such as clustering, to see if the CMS beneficiaries get clustered along with the FAERS patients that had an ADE.