Drug-drug interaction data trust: what the FDA adverse event database actually contains
The FDA Adverse Event Reporting System contains over 29 million reports, but duplicate entries, missing drug fields, and inconsistent coding make raw FAERS data unreliable for drug-drug interaction analysis. Most AI models trained on FAERS never quantify these data quality failures before inference.
The FDA Adverse Event Reporting System (FAERS) contains over 29 million adverse event reports spanning more than five decades. Researchers, pharmaceutical companies, and AI developers routinely treat this database as ground truth for drug-drug interaction signal detection. It is not ground truth. It is a passive surveillance system with structural data quality problems that most users never quantify before building models on top of it.
The gap between what people assume FAERS contains and what it actually contains is where drug-drug interaction data trust breaks down.
What is the FDA database for adverse events?
FAERS is the FDA's primary post-market safety surveillance database. It collects voluntary reports of adverse events, medication errors, and product quality complaints associated with drugs and therapeutic biologics. Healthcare professionals, consumers, and manufacturers submit these reports through MedWatch, the FDA's safety reporting portal.
FAERS replaced the older Adverse Event Reporting System (AERS) in 2012, consolidating data into a more accessible structure. The database organizes reports into seven linked data files: demographics, drugs, reactions, outcomes, report sources, therapy details, and indications. Each quarterly data release adds roughly 500,000 to 700,000 new reports.
The critical distinction most users miss: FAERS is a hypothesis-generating system, not a hypothesis-confirming one. The FDA explicitly states that FAERS data alone cannot establish causation between a drug and an adverse event. Yet drug-drug interaction AI models routinely train on FAERS as though causation is embedded in the data.
What is the FDA's new adverse event database?
The FDA has been modernizing its adverse event infrastructure through the openFDA initiative and the Sentinel System. openFDA provides API access to FAERS data in JSON format, making programmatic queries easier than downloading flat quarterly files. But openFDA does not fix the underlying data quality problems; it simply makes flawed data more accessible.
The Sentinel System represents a fundamentally different approach. Rather than relying on voluntary reports, Sentinel queries electronic health records and claims data from a distributed network covering over 100 million patients. Sentinel uses active surveillance, meaning it can systematically look for safety signals rather than waiting for someone to file a report.
For drug-drug interaction research, Sentinel offers a stronger evidentiary foundation because it captures actual prescription fills, diagnosis codes, and clinical outcomes in structured formats. But Sentinel access is limited, and most commercial AI developers still default to FAERS because it is free and publicly available.
What is the MDR database?
The Medical Device Reporting (MDR) database is a separate FDA system that collects adverse event reports specifically for medical devices rather than drugs. Device manufacturers, importers, and user facilities are required to submit MDR reports when they become aware of information suggesting a device may have caused or contributed to a death or serious injury.
The MDR database is maintained through the Manufacturer and User Facility Device Experience (MAUDE) system. While MDR and FAERS share some structural similarities as passive surveillance systems, they serve different product categories and have different reporting requirements. MDR data is not relevant for drug-drug interaction analysis, but the two databases are frequently confused in searches about FDA adverse event reporting.
Both databases share a common problem: voluntary reporting rates are low. The FDA estimates that FAERS captures only 1% to 10% of actual adverse drug events. MDR has similar underreporting challenges for devices.
The seven data quality failures in FAERS
FAERS data quality problems are well documented in pharmacovigilance literature but rarely quantified by teams building drug-drug interaction models. Here are the specific failures that degrade trust.
Duplicate reports. The same adverse event frequently appears multiple times in FAERS. A patient, their physician, and the drug manufacturer may all submit separate reports for the same event. The FDA provides a caseid and primaryid system to help identify duplicates, but deduplication is imperfect. Studies have found duplicate rates ranging from 4% to over 20% depending on the drug and time period.
Missing concomitant medication data. Drug-drug interaction analysis requires knowing all drugs a patient was taking. In FAERS, reporters often list only the suspected drug and omit concomitant medications. One analysis of FAERS data found that approximately 30% of reports list only a single drug, making drug-drug interaction signal detection impossible for those cases.
Inconsistent drug naming. FAERS uses free-text drug name fields. The same drug can appear as a brand name, generic name, abbreviation, or misspelling. Metformin might appear as "metformin," "Glucophage," "metformin HCl," "metfromin," or dozens of other variants. The FDA maps some entries to standard terms, but many remain unmapped.
Absent or incorrect dose information. Dose fields in FAERS are frequently blank or contain non-standard entries. Without dose data, distinguishing between therapeutic use and overdose becomes impossible. This matters for interaction analysis because many drug-drug interactions are dose-dependent.
Reporter qualification variability. Reports from physicians, pharmacists, consumers, and lawyers all enter the same database with the same weight. A consumer report based on a Google search carries the same structural status as a physician report based on clinical assessment. The role_cod field identifies reporter type, but most models ignore this distinction.
Temporal imprecision. FAERS captures the date of the report, but the actual date of the adverse event is often missing or approximate. For drug-drug interaction analysis, knowing whether two drugs were taken simultaneously, sequentially, or months apart is essential. FAERS frequently cannot answer this question.
Indication bias. Drugs prescribed for serious conditions are over-represented in FAERS because sicker patients experience more adverse events and generate more reports. This creates confounding: an apparent drug-drug interaction signal may actually reflect the underlying disease severity rather than a pharmacological interaction.
Key statistics
FAERS contains over 29 million adverse event reports as of 2024, with 500,000 to 700,000 new reports added each quarter.
The FDA estimates FAERS captures only 1% to 10% of actual adverse drug events, meaning 90% to 99% of real-world drug interactions are never reported.
Approximately 30% of FAERS reports list only a single drug, rendering them useless for drug-drug interaction signal detection.
Duplicate report rates in FAERS range from 4% to over 20% depending on the drug and reporting period analyzed.
The Sentinel System covers over 100 million patients through distributed data, offering active surveillance compared to FAERS passive voluntary reporting.
What is the most accurate drug interaction checker?
No single drug interaction checker is universally the most accurate, and the answer depends on what accuracy means in context. Clinical decision support tools like Lexicomp, Clinical Pharmacology, Micromedex, and Epocrates each use different source databases, classification systems, and severity ratings.
A 2019 study in the Journal of the American Pharmacists Association compared five major interaction checkers and found significant disagreement. For the same drug pair, one tool might flag a "major" interaction while another lists it as "moderate" or misses it entirely. Agreement rates across all five tools were below 50% for many drug combinations.
The accuracy problem is not the algorithm. It is the data. Every interaction checker is only as reliable as the evidence base it draws from. Tools that rely heavily on FAERS-derived signals inherit all the data quality problems described above. Tools that incorporate controlled clinical trial data, pharmacokinetic modeling, and structured post-market surveillance produce more reliable results, but they also cover fewer drug combinations.
For AI-driven drug interaction prediction, the accuracy ceiling is set by training data integrity. A model trained on unscored FAERS data with 20% duplicates, 30% single-drug reports, and inconsistent naming conventions cannot produce reliable interaction predictions regardless of how sophisticated the architecture is.
Why FAERS data quality matters for AI drug interaction models
The pharmacovigilance community has used FAERS for decades with a clear understanding of its limitations. Trained pharmacoepidemiologists apply disproportionality analysis methods like the Reporting Odds Ratio (ROR) and the Information Component (IC) while accounting for known biases. They treat FAERS signals as hypotheses requiring confirmation through other data sources.
AI developers often skip this step. Machine learning models trained on raw FAERS data treat every report as an equally weighted observation. The model has no way to distinguish a duplicate from a unique case, a physician report from a lawyer-generated report, or a confirmed interaction from a coincidental co-prescription.
This problem is compounded when drug-drug interaction AI products are marketed to clinicians. A recent MedPageToday article highlighted physician wariness toward AI-driven prescription systems, noting that doctors question whether AI tools have sufficient data quality to make safe medication recommendations. That skepticism is warranted. If the training data cannot be trusted, the model output cannot be trusted.
As we covered in our analysis of why AI models trained on unscored health data fail in production, the failure mode is not dramatic. Models do not crash. They produce plausible-looking outputs that are subtly wrong, and those subtle errors compound in clinical settings.
The drug-drug interaction data trust gap
Drug-drug interaction data trust requires more than verifying that a data source exists. It requires scoring the underlying data across dimensions that matter for the intended use case.
Consider what a trust assessment of FAERS data for drug-drug interaction analysis would need to evaluate:
Provenance. Who submitted each report? Was it a healthcare professional with direct patient access, a consumer self-reporting, or a legal firm filing reports for litigation purposes? The reporting source fundamentally changes the evidentiary weight.
Recency. When was the adverse event reported relative to when it occurred? FAERS contains reports from the 1960s alongside reports from last quarter. Drug formulations, prescribing patterns, and patient populations change over time, and stale reports may describe interactions that no longer reflect current clinical reality.
Quality. Are the required fields populated? Is the drug name standardized? Is the dose recorded? Is the outcome specified? A FAERS report with missing concomitant medications, no dose information, and a vague outcome description scores far lower than a complete report.
Concordance. Does this report align with other reports for the same drug combination? A single outlier report for a drug pair with millions of prescriptions and no other signals carries different weight than a cluster of consistent reports.
Validation. Has the reported interaction been confirmed through controlled studies, pharmacokinetic data, or Sentinel System analysis? An unvalidated FAERS signal is a hypothesis, not evidence.
Without scoring these dimensions, every FAERS report gets treated identically. That is not a data quality problem. It is a data trust problem.
How trust scoring changes drug interaction analysis
When you apply trust scoring to FAERS data before it enters an AI training pipeline, three things change.
First, you eliminate noise. Duplicate reports, incomplete records, and low-provenance submissions get scored below the trust floor and excluded from model training. This does not reduce your dataset arbitrarily; it reduces it precisely, removing records that would degrade model performance.
Second, you weight evidence appropriately. A physician-reported adverse event with complete drug lists, dose information, and a confirmed outcome contributes more to the model than a consumer report with a single drug name and no dose data. Trust scoring makes this weighting explicit and auditable.
Third, you create an audit trail. When the FDA asks how your AI drug interaction model reached a specific recommendation, and they will ask, you can trace the output back through scored training data with documented provenance. As we detailed in how the FDA will audit your health AI's training data, regulatory expectations for data provenance are increasing, not decreasing.
The RxNorm problem inside FAERS
FAERS drug naming inconsistency is not just an annoyance. It creates false negatives and false positives in drug-drug interaction detection.
False negatives occur when the same drug appears under different names across reports, preventing aggregation. If half the reports for a dangerous interaction use the brand name and half use the generic, and your normalization pipeline misses some variants, you split the signal and may fail to detect the interaction.
False positives occur when different drugs or formulations get mapped to the same normalized term incorrectly. Combination products are particularly problematic. A report for "lisinopril/hydrochlorothiazide" might get mapped only to lisinopril, attributing hydrochlorothiazide-related interactions to lisinopril alone.
We covered the medication data standardization challenge in depth in RxNorm drug data integrity: what medication reconciliation errors mean for AI. The FAERS context makes this problem worse because there is no reconciliation step. In clinical settings, pharmacists and physicians can verify medication lists. In FAERS, whatever the reporter typed is what the database contains.
What FDA expects from post-market drug interaction data
The FDA's expectations for real-world evidence quality have tightened considerably. The 2023 draft guidance on using real-world data for regulatory decisions specifies that data relevance, reliability, and completeness must be demonstrated for each intended use.
For drug-drug interaction analysis specifically, the FDA expects:
These expectations apply whether you are submitting a supplemental New Drug Application with updated interaction labeling or developing a clinical decision support tool that flags potential interactions.
The gap between what FDA expects and what most teams actually do with FAERS data is substantial. Most published FAERS analyses describe their deduplication approach in a methods paragraph and move on. Few quantify the data quality of the reports they retained after deduplication.
Building trust into drug interaction intelligence
The path forward is not abandoning FAERS. It is scoring FAERS data before use, combining it with higher-provenance sources, and documenting the trust characteristics of the combined dataset.
A scored approach to drug-drug interaction data would:
This is what the Data Trust Index framework was designed to do. Not for FAERS specifically, but for any health data source where the intended use requires a level of reliability that the raw data cannot guarantee.
The DTI Engine scores every health data record 0 to 100 across 8 trust dimensions before your AI model sees it. If your team is building drug interaction models on FAERS data, evaluating real-world evidence for regulatory submission, or deploying clinical decision support tools that flag medication interactions, you need to know the trust characteristics of your training data before inference. Contact Louis Simeonidis at louis@supertruth.ai or (215) 918-4140.
Further reading:
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
DTI scores the record, not the patient.
8 dimensions. 0–100. Travels with every record permanently.