📄 medicine

faers: A High-Fidelity Framework and R/Bioconductor Package for Precision Adverse Event Surveillance

The paper introduces **faers**, an open-source R/Bioconductor package that provides a standardized, high-fidelity framework for transforming raw FAERS data into analysis-ready formats through regulatory-compliant deduplication and automated terminology mapping, thereby enabling scalable, reproducible, and precise pharmacovigilance signal detection.

Original authors: Zhangyu Wang, Yun Peng, Jian-Guo Zhou, Xiaoyun Bu, Yankun Zhao, Zimo Li, Bin Yan, Yujun Sun, Chunyang Wang, Chenyang Shu, Yanru Cui, Shixiang Wang

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhangyu Wang, Yun Peng, Jian-Guo Zhou, Xiaoyun Bu, Yankun Zhao, Zimo Li, Bin Yan, Yujun Sun, Chunyang Wang, Chenyang Shu, Yanru Cui, Shixiang Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the FDA's Adverse Event Reporting System (FAERS) as a massive, chaotic library where millions of people drop off sticky notes about how they felt after taking medicine. Some notes are written in perfect English, some are scribbles, some are in different languages, and unfortunately, many people have dropped off the exact same note five times because they heard a friend do it or because the system glitched. Trying to find a real pattern in this mountain of sticky notes is like trying to find a specific needle in a haystack that's also on fire.

Enter faers, a new digital tool created by a team of researchers that acts like a super-smart, high-speed librarian for this chaotic library.

The Problem: A Messy Library

The paper argues that while FAERS is a goldmine for spotting drug safety issues, it's currently too messy to use easily. The data is inconsistent, full of duplicates, and hard to clean up. Existing tools are either too simple (like a basic search bar that can't do complex math) or require researchers to do all the messy cleaning work by hand before they can even start analyzing. The authors explicitly state that current "black box" platforms don't give researchers enough control to do the deep, precise detective work needed for modern drugs, especially complex ones like cancer immunotherapies.

The Solution: The "faers" Robot Librarian

The team built faers, an open-source software package (a tool for the R programming language) that automates the entire cleanup process. Think of it as a robot that doesn't just read the sticky notes; it organizes them, translates them into a standard language, and removes the duplicates before you even ask a question.

Here is how it works, step-by-step:

The Grabber: It automatically downloads the raw data from the FDA.
The Translator: It uses a massive medical dictionary (MedDRA) to ensure that "heart attack," "cardiac arrest," and "myocardial infarction" are all recognized as the same thing.
The Detective: This is the most important part. The tool uses a clever, multi-step rule system to find and remove duplicate reports. It checks if a report is a repeat by looking at eight different details (like age, gender, country, and when the drug was taken). It's like checking if two people telling the same story are actually the same person or just two different people telling the same story.
The Calculator: Once the data is clean, it runs advanced math to spot "signals"—patterns where a specific drug seems to be linked to a specific side effect more often than chance would allow.

What They Found (The Proof)

To prove their robot librarian works, the researchers ran three specific tests:

Test 1: The Heart Check
They asked the tool to look for heart problems linked to a popular class of cancer drugs called PD-1/PD-L1 inhibitors. They compared their results to a famous study done by other scientists. The results matched up almost perfectly.

The Numbers: The tool found 66,278 total cases after cleaning (compared to 60,766 in the original study) and identified 3,266 severe heart cases (vs 3,012 originally).
The Signal: The tool calculated a "Proportional Reporting Ratio" (a measure of how strong the signal is) of 1.40, which is nearly identical to the 1.39 found in the original study.
The Detail: It confirmed that these heart issues usually happen quickly, with a median time of 37 days after starting the drug, and mostly affect men (about 67.1% of cases).

Test 2: The Cancer Therapy Check
They tested the tool on a study about CAR-T cell therapy (a type of cancer treatment) and whether antibiotics increased the risk of new cancers later on.

The Numbers: The tool recreated the original study's patient group with 99.88% consistency. For example, it found 11,241 total cases compared to the original 11,255.
The Result: Just like the original study, the tool found that patients who took antibiotics had a significantly higher risk of developing new cancers, with a statistical difference so strong it had a p-value of less than 0.0001.

Test 3: The New Discovery (Age and Gender)
This is where the tool showed its power to find things others missed. The researchers used faers to look at how age and gender interact when reporting side effects from immune therapies.

The Finding: They discovered a "statistically significant" interaction. Younger women (ages 18–44) reported side effects much more often than men of the same age. However, as people got older, this gap got smaller. By the time patients were 75 or older, the reporting rates between men and women were almost the same.
The Confidence: The math showed this wasn't a fluke; the interaction had a p-value of less than 0.001.

How Fast and Big Can It Go?

The team tested how fast the tool is by feeding it huge amounts of data.

Speed: It processed a full year of data (the entire 2015 dataset) in just 2.39 minutes.
Scalability: When they fed it 32 quarters (eight years) of data, it finished in 22.5 minutes. The speed didn't slow down; it stayed steady at about 1.91 quarters per minute.
Memory: Even with all that data, it only used about 17.26 GB of computer memory, which is manageable for standard research computers.

What the Paper Says It Is NOT

The authors are very careful to say what this tool doesn't do.

It doesn't prove cause and effect. Just because the tool finds a link between a drug and a side effect doesn't mean the drug caused it. The data comes from spontaneous reports, which can be biased (people might report scary things more often than boring ones).
It's not a magic wand for every drug. If a drug description is written in a very weird way, the tool might still need a human to check it manually.
It doesn't replace doctors. It's a tool to help researchers and regulators spot potential risks so they can investigate further.

The Bottom Line

The faers package is a high-speed, transparent, and reproducible way to clean up the messy FAERS database. It suggests that by using this tool, researchers can find safety signals faster and more accurately than before, uncovering subtle patterns—like how age and gender change risk—that might otherwise be hidden in the noise. It doesn't solve the problem of drug safety on its own, but it gives scientists a much better microscope to look at the data they already have.

Technical Summary: faers: A High-Fidelity Framework and R/Bioconductor Package for Precision Adverse Event Surveillance

1. Problem Statement

The FDA Adverse Event Reporting System (FAERS) is a cornerstone of post-marketing pharmacovigilance but presents significant barriers to reproducible, large-scale analysis. The data is characterized by:

Structural Complexity: Multi-table relational structures with inconsistent formats across reporting periods.
Data Redundancy: Pervasive reporting redundancies and duplicate reports that distort signal detection.
Terminological Inconsistency: Unstandardized medical terminology that hinders semantic analysis.
Tooling Gaps: Existing solutions (e.g., FAERS Public Dashboard, openFDA) function as "black boxes" with limited programmatic flexibility, while existing R packages (e.g., PhViD, openEBGM) often require pre-processed inputs, lacking robust capabilities for raw data parsing, iterative deduplication, and terminology standardization.

These limitations impede the transition from detecting dose-dependent non-specific toxicities to the precise identification of mechanism-driven specific toxicities, particularly for complex therapies like immune checkpoint inhibitors (ICIs) and CAR-T cell therapies.

2. Methodology

The authors developed faers, an open-source R/Bioconductor package designed as a high-fidelity, end-to-end pipeline for adverse event surveillance. The framework utilizes an object-oriented paradigm (R S4 class system) to encapsulate FAERS data into unified container objects, ensuring data integrity and traceability.

The pipeline consists of four core functional modules:

Data Acquisition and Parsing: Automatically retrieves raw data from official FDA sources (ASCII/XML), performs concurrent downloads with integrity checks, and parses heterogeneous files into standardized S4 objects. It manages field mapping and enriches temporal metadata for data dating back to 2004.
Clinical Terminology Standardization: Harmonizes records using the Medical Dictionary for Regulatory Activities (MedDRA) hierarchy via a two-tier matching mechanism (Lowest Level Terms to Preferred Terms). It integrates the Athena drug knowledge base for bulk drug name normalization and synonym resolution.
Multi-level Deduplication Strategy: Employs a rule-based algorithm executing record linkage across eight critical dimensions (sex, age, reporting country, event date, therapy start date, indication, drug name, and adverse event). It utilizes a six-round iterative strategy that enforces exact matching for primary fields (drug, event) while allowing tolerance for auxiliary fields, using unique encoding for missing values to prevent spurious clustering.
Pharmacovigilance Signal Detection: Based on a unified contingency table framework, it integrates a comprehensive suite of disproportionality methods:
- Frequentist: Reporting Odds Ratio (ROR), Proportional Reporting Ratio (PRR), and associated statistical tests.
- Bayesian: Bayesian Confidence Propagation Neural Network (BCPNN) with normal approximation and Markov Chain Monte Carlo (MCMC) variants, and Empirical Bayes Geometric Mean (EBGM).

Performance Optimization: The package leverages the data.table library for memory-efficient data manipulation and integrates parallel computing strategies to handle large-scale datasets.

3. Key Results

The authors evaluated the package through performance benchmarking and three distinct case studies:

Performance and Scalability:
- Efficiency: Processing the full 2015 FAERS dataset (approx. 1.5 million records) took 2.39 minutes. The deduplication stage accounted for 50.2% of the runtime, while signal detection required only 3.46 seconds (<3%).
- Scalability: Stress tests scaling from 1 to 32 quarters (8 years of data) demonstrated near-linear growth in both processing time ( $R^2 = 0.9811$ ) and memory allocation. Throughput remained consistent at 1.91 quarters per minute.
- Resource Usage: Peak memory footprint was 17.26 GB, manageable on standard research workstations without specialized High-Performance Computing (HPC) resources.
Case Study I: Replication of PD-1/PD-L1 Cardiotoxicity:
- The package successfully replicated findings from Cheng et al. (2024) regarding severe cardiac adverse events (scAEs).
- Results were consistent across data foundation, signal strength (PRR 1.40 vs. 1.39 in original), and patient demographics (median age 69, 67.1% male).
- The analysis identified myocarditis, pericarditis, and heart failure as top signals, with a median time-to-onset of 37 days.
Case Study II: Validation of CAR-T Therapy and Secondary Primary Malignancies (SPMs):
- Replicating Peng et al. (2025), the package confirmed a significant association between antibiotic exposure and SPM risk in CAR-T therapy patients.
- Cohort construction and baseline characteristics showed >99% consistency with the original study.
- Cumulative hazard analysis confirmed significantly higher risk in antibiotic-exposed groups ( $p < 0.0001$ ).
Case Study III: Sex-Age Interaction in irAE Reporting:
- An exploratory analysis of 2015–2025 data revealed a statistically significant interaction between sex and age in immune-related adverse event (irAE) reporting ( $p < 0.001$ ).
- Females showed higher reporting rates than males, but this gap attenuated with age, converging in the $\ge$ 75 age group.
- Stratified analysis showed the Reporting Odds Ratio (ROR) for females vs. males decreased from 1.11 (18–44 age group) to 1.06 ( $\ge$ 75 age group).

4. Significance and Claims

The paper positions faers as a critical tool for advancing precision pharmacovigilance by addressing the structural and technical bottlenecks of FAERS data analysis.

Standardization and Reproducibility: By providing a transparent, scriptable, end-to-end workflow, the package eliminates the "black box" nature of existing tools and the methodological heterogeneity caused by manual preprocessing. It ensures that data cleaning, deduplication, and signal detection are fully reproducible.
Data Fidelity: The multi-level deduplication strategy is highlighted as a key differentiator. The authors note that by purifying the denominator through rigorous deduplication, the package can correct signal-dilution effects, potentially revealing more robust risk profiles (as seen in the higher PRR values for cardiac toxicity compared to some prior studies).
Democratization of Analysis: By optimizing for memory efficiency and parallel processing, the package enables high-volume data processing on standard workstations, removing the barrier of requiring specialized HPC infrastructure.
Precision Surveillance: The framework supports complex, multi-confounder adjusted interaction models (e.g., sex-by-age interactions), moving beyond basic signal detection to uncover nuanced demographic and mechanistic signals often overlooked by conventional approaches.

The authors conclude that faers provides a robust technical foundation for generating high-quality, real-world evidence, thereby strengthening drug safety monitoring and supporting evidence-based clinical decision-making. They acknowledge limitations inherent to spontaneous reporting systems (e.g., under-reporting, causality inference) and note that full utilization of semantic mapping requires a MedDRA license. Future work aims to integrate Natural Language Processing (NLP) and Large Language Models (LLMs) to extract insights from unstructured narrative text.