Using precision pharmacoviglance to detect developmentally-regulated adverse drug reactions: a case study with antiepileptic drugs

Code and documentation accompanying the poster "Using precision pharmacoviglance to detect developmentally-regulated adverse drug reactions: a case study with antiepileptic drugs" presented at the 2018 OHDSI symposium in Bethesda, MD, USA

This repository documents the code and analyses to reproduce the results of the above poster. The main purpose is to increase transparency in the methodology used in this project. Subsequent work will employ better software development practices.

Introduction

This poster outlines a first-pass pharmacovigilance methodology for detecting ADRs that disproportionally affect those within younger age categories relative to adults. The following briefly describes the code used to generate the findings. Refer to the Rmarkdown notebooks for either descriptive visualizations of the dataset or for creating the poster figures/other visualizations.

In subsequent work, a systems pharmacology approach would ideally be taken which would include both pharmacovigilance statistics with biological/chemical knowledge to ground the expected results in a biologically plausible hypothesis.

If you find this useful, would like to discuss this work, or would like to collaborate, please contact me at nick.giangreco@gmail.com. If you would like to improve the code and documentation, feel free to do a pull request!

Coding Workflow

This section outlines the workflow taken for producing the results of this poster. The most detail are in the code itself. But the boxes (indicating the script type) contain the scripts that are used in the methodology. Most of the scripts were run on a private cluster to take advantage of parallelism over multiple cores. The data files in the data directory are not included just due to size. So by just cloning from github, the code couldn't be run without the data files (though I'd be happy to share the data directory with you or share applicable data files). However, hopefully this description will make the analysis for this poster transparent.

Legend:

• Thick Arrow: 'is used in'
• Thin arrow: 'output of previous is funneled into'
• Trapezoids: 'top script uses the bottom script'

Constructing Vocabulary Taxonomies & Hierarchies

These scripts use the OMOP CDM files (downloaded from Athena, but the directory with the files is ignored in this repository) to construct the taxonomic hierarchies used in the analysis.

Helper Functions & Filters

These scripts contain functions, filters, and data processing to support the main ADR detection and evaluation scripts. These are used in the analysis

Preprocessing AEOLUS

The AEOLUS¹ dataset was downloaded², curated and loaded onto our lab's mysql servers. I further processed the dataset so that all reports of ADRs had

• SNOMED encoded outcomes
• RxNorm encoded drugs
• Reported age

Further processing details can be found in the Real Data script.

A synthetic dataset was constructed (see Synthetic Data script), where the formatting is exactly the same, but the only difference is that the distribution of ages for 4/1000 sampled ADRs were redistributed to simulate the disproportionality in age categories we want and don't want to detect.

Adverse drug reaction detection methods

This is the main analysis where pharmacovigilance is performed. These scripts are meant to be used on a cluster (for processing the real data; the synthetic data can be run locally in a small amount of time). I ran each PhV script on our cluster using 25 cores, and each PhV script took about 3 hours to complete. The other scripts either took very little time or, in the case of prepare_aeolus_ATC.R, took about 20-30 minutes (this script could be optimized with using data.table but I didn't get around to doing that).

The three types of ADRs, as outlined in the poster, are computed in the respective scripts in Real Data ADR detection.

The ADRs, with controls, in the synthetic dataset are computed in Synthetic Data ADR detection. The ADR type is the same as the 'ADRs' type in the poster i.e. coded in the PhV_of_ADRs.R script.

Real Data ADR Evaluation

The positive control ADRs were assessed whether they could be detected and whether the negative control ADRs could not be detected. Fortunately, all positive controls are detected, and no negative controls are detected. This is shown in a confusion matrix in the synthetic data log file (see select_young_age_category_biased_ADRs_control_detection.log).

Real Data ADR Evaluation

This is where the candidate ADRs are detected, using the same filtering as the synthetic dataset. Numbers of candidate ADRs are outputted into a log file (see PhV_of_ADRs*biased_ADRs.log).

Data Availability

Because of the size of the processed AEOLUS dataset ( ~ 2 Gb), vocabulary files (> 100 Mb), and analysis results (> 100 Mb), they are ignored in this repository. But if you would like access to it, please let me know at nick.giangreco@gmail.com, and I'd be happy to share it!

References

1. Banda JM, Evans L, Vanguri RS, Tatonetti NP, Ryan PB, Shah NH (2016) A curated and standardized adverse drug event resource to accelerate drug safety research. Scientific Data 3: 160026. https://doi.org/10.1038/sdata.2016.26
2. Banda JM, Evans L, Vanguri RS, Tatonetti NP, Ryan PB, Shah NH (2016) Data from: A curated and standardized adverse drug event resource to accelerate drug safety research. Dryad Digital Repository. https://doi.org/10.5061/dryad.8q0s4