MedDRA coding consistency and signal detection in pharmacovigilance

Signal detection in pharmacovigilance is the search for new or changed associations between a medicine and an adverse event. The standard statistical method for spontaneous report databases is disproportionality analysis, used alongside review of individual cases: it compares how often an event is reported with one product against all others. Those counts are made per MedDRA term, usually the Preferred Term, so when one clinical event is coded to different terms across cases, the count splits and a real signal can sit below every threshold.

What is signal detection in pharmacovigilance?

Signal detection is the first step of signal management, followed by validation, confirmation, analysis and prioritization, assessment, and a recommendation for action.

EU law defines a signal as information suggesting a new potentially causal association, or a new aspect of a known one, between an intervention and an event, judged likely enough to justify verificatory action (Implementing Regulation (EU) No 520/2012, Article 19, in GVP Module IX Rev 1, 2017).

Signals come from spontaneous reports, active surveillance, studies and the literature, and detection may use case review, statistics, or both. EMA also screens a designated medical event (DME) list of MedDRA terms for serious, often medicine-related conditions, irrespective of statistical criteria. A case coded to a term outside that list escapes that review.

FDA's Adverse Event Monitoring System (AEMS, formerly FAERS), EudraVigilance and VigiBase all code events with MedDRA. FDA's postmarketing surveillance best practices (January 2024) treat data mining results as hypothesis generating, and the absence of disproportionality does not confirm the absence of a safety signal in postmarket surveillance.

How does disproportionality analysis work?

Every method starts from a two-by-two table for one drug and event pair: reports with both, either one, or neither. FDA's data mining white paper (2018) calls the proportional reporting ratio (PRR) the foundational concept for many disproportionality methods.

Measure	Compares	Where used, and how a result is read
Proportional reporting ratio (PRR)	The event's share of the product's reports against its share of all others	EMA 2008 EudraVigilance guideline, section 4.6: PRR above 2 with chi-square above 4, or a 95 percent confidence lower bound of at least 1, each with at least 3 cases
Reporting odds ratio (ROR)	Odds of the event with the product against all others	EMA screening of EudraVigilance (2016): 95 percent lower bound above 1, an event on the important medical event list, and at least 3 cases for a substance on the EU additional monitoring list, 5 for others
Information component (IC)	Observed against expected, Bayesian	Uppsala Monitoring Centre for VigiBase: a rising IC with a positive IC025 (lower limit of the 95 percent credibility interval)
Empirical Bayes geometric mean (EBGM)	Observed against expected, shrunk toward no association at small counts	FDA, read with EB05 (the lower 90 percent limit FDA reports); thresholds adjusted to the severity of event and condition

A drug and event pair that crosses a threshold is a signal of disproportionate reporting (SDR), which needs medical assessment before it is treated as a safety signal.

EMA's GVP Module IX Addendum I (2017) notes that the Preferred Term level has been shown to be a good choice for sensitivity and positive predictive value. Coding is one of several reasons a count moves with no change in patients; duplicate reports, underreporting, stimulated reporting and masking are the others (FDA white paper; EMA screening, 2016).

Why does MedDRA coding consistency matter for signal detection?

MedDRA coding consistency matters because disproportionality counts are made per Preferred Term: when coders choose LLTs under different PTs for one clinical event, its count splits and each piece can fall below a screening threshold. EMA's 2008 guideline names coding practices among the factors that shape a PRR.

MedDRA has five levels: System Organ Class (SOC), High Level Group Term, High Level Term (HLT), Preferred Term (PT) and Lowest Level Term (LLT). Version 29.1 holds 82,124 current LLTs, 27,504 PTs and 27 SOCs (What's New for MedDRA 29.1). Each LLT links to only one PT, and each PT has one primary SOC. The coder chooses the LLT; everything above follows (MedDRA coding for adverse events, verbatim to SOC).

Coder variance at that choice distorts counts in two directions:

• Splitting. One event reaches several PTs because coders chose different LLTs. Each PT count can stay below threshold, and nothing in the output shows the event was divided.
• Merging. Distinct events collect under one non-specific term. The count can flag a mixed case series while the specific events stay invisible.

A hypothetical example: six reports of one presentation are coded to three PTs, as 3, 2 and 1 cases. Under a 3-case minimum, as in the 2008 PRR criteria, the rows of 2 and 1 cannot flag whatever their ratio. Under the 5-case minimum EudraVigilance applies outside the additional monitoring list, no row qualifies. One PT holding all six clears either minimum and is judged on its ratio.

The ICH-endorsed MedDRA Data Retrieval and Presentation: Points to Consider (DRP:PtC, Release 3.26, March 2026) lists where counts change for reasons unrelated to patients:

Where variance enters	Effect on the count	Section
Related events land on more than one PT	The true rate is obscured unless PTs are grouped	2.5.2, 3.1
A diagnosis coded alone, or with its signs and symptoms	Counts move with the convention	2.4
A laboratory finding coded as a test result	It lands in SOC Investigations, which has no multiaxial links	2.5.3.2
Legacy data converted term to term	Specific searches miss cases	2.1
A new MedDRA version	LLTs move between PTs; primary SOCs change	2.6

The DRP:PtC gives an example: a thrombocytopenia search confined to SOC Blood and lymphatic system disorders misses PT Platelet count decreased, in SOC Investigations.

How do MedDRA and CTCAE versions change signal counts?

A new MedDRA version can move an LLT to a different PT, so a count run across two versions can change with no new case. MedDRA updates twice a year; an X.1 release such as 29.1 (September 2026) carries simple changes only (DRP:PtC section 2.6).

Simple changes still move counts. Version 29.0 moved medication error LLTs from "error" PTs to "issue" PTs, and 29.1 reversed it: LLT Incorrect storage of drug went from PT Product storage issue back to PT Product storage error (What's New for MedDRA 29.1, section 3.2).

Oncology trials add a second dictionary: NCI states that each CTCAE v6.0 term is a MedDRA v28.0 LLT. CTCAE v5.0 grades a low platelet count as Platelet count decreased, in SOC Investigations. CTCAE v6.0 grades it as Thrombocytopenia, in SOC Blood and lymphatic system disorders. A program with studies on both records one laboratory finding under two LLTs in two SOCs (Thrombocytopenia and Platelet count decreased, side by side; every CTCAE v6.0 change).

That is the DRP:PtC example in practice. A pooled thrombocytopenia query across v5.0 and v6.0 studies needs both PTs, or an SMQ whose term list holds both in the data's MedDRA version.

How does coding consistency affect DSURs and IND safety reports?

Trial aggregate review counts by Preferred Term too. ICH E2F asks for DSUR line listings and summary tabulations that present MedDRA Preferred Terms by SOC, and for the dictionary version used. Under 21 CFR 312.32(c)(1)(i)(C), an aggregate analysis showing specific events more often in the drug group than in a control group can make them reportable in an IND safety report. A split lowers the drug arm's count there too.

The clinical database term and the safety database MedDRA term can also differ for one case (SAE reconciliation before database lock). See IND safety reports under 21 CFR 312.32 and what a DSUR contains.

What is a MedDRA SMQ, and does it repair a split signal?

A Standardised MedDRA Query (SMQ) is a grouping of MedDRA terms, ordinarily PTs, for a defined medical condition or area of interest (SMQ Introductory Guide 29.1). MedDRA 29.1 carries 110 Level 1 SMQ topics and 230 SMQs counting sub-SMQs (What's New 29.1). Grouping recovers a split count; it cannot recover a case coded to the wrong concept.

A narrow search favors specificity; a broad one, sensitivity. The DRP:PtC (section 4.9) says SMQs were developed to address MedDRA's granularity.

Pearson and colleagues (2009) mined FDA spontaneous reports from 2001 to 2005 for 26 drugs. The share of unlabeled, literature-supported SDRs was higher at the HLT and SMQ levels than at the PT level, at the cost of less specific language and an extra step to examine the PTs inside each flagged grouping.

How do you keep MedDRA coding consistent?

Six practices keep MedDRA coding consistent across coders, partners and versions:

1. Write conventions down and follow the ICH-endorsed MedDRA Term Selection: Points to Consider (Release 4.26, March 2026): always select an LLT, and never add information the source lacks (where AI helps at the coding step).
2. Code from the verbatim, not a legacy term, and keep the verbatim beside the code.
3. Decide once whether to code a diagnosis alone or with its signs and symptoms, and check a partner's convention before pooling.
4. Record the MedDRA version, and the CTCAE version for graded oncology events, and query in that version (DRP:PtC sections 2.2 and 2.6).
5. Read SOC Investigations alongside the condition's SOC before trusting a primary SOC table.
6. Keep a record of verbatim-to-LLT decisions, and have a qualified reviewer check term selection, autoencoder output included, on a fixed schedule (Term Selection PtC section 2.2).

Where Burna AI fits

Consistency is cheapest where an event is first recorded: in an oncology trial, the clinical note at the site.

Burna AI grades each adverse event in the note against the CTCAE version the protocol names, with v5 and v6 live on one engine. Every graded event carries its MedDRA LLT code and source sentence, from the same pass as the grade. For a v6.0 grade, that code is the MedDRA v28.0 LLT the CTCAE term itself is, so grade and code name the same term. Every grade traces to the line in the medical record that produced it (what Burna AI does today). AI suggests, clinicians decide.

For a sponsor, the trial safety data the label rests on arrives with each code tied to its source sentence, which helps elucidate the safety profile while the trial runs. See Burna AI for pharmacovigilance.

Common questions about signal detection and MedDRA coding

What methods are used for signal detection in pharmacovigilance?

Spontaneous report databases are screened with disproportionality measures: the PRR and ROR in EudraVigilance, the information component in VigiBase, and the EBGM at FDA. Case review runs alongside, and a flagged pair is medically assessed before it is treated as a signal.

Is a signal of disproportionate reporting the same as a safety signal?

No. A signal of disproportionate reporting is a statistical result; it becomes a safety signal only when medical review judges the association likely enough to justify verificatory action. EMA's guideline on statistical signal detection says SDRs should always be medically assessed, and FDA treats data mining output as hypothesis generating.

What MedDRA level is used for signal detection?

Usually the Preferred Term, which EudraVigilance uses for routine screening. Grouping terms and SMQs serve specific questions.

Can we change an SMQ to fit our product?

Yes, but once its terms or structure change, the DRP:PtC (section 5.1) calls it a modified MedDRA query based on an SMQ, not an SMQ. Its example excludes PT Syncope from a QT prolongation search for an antipsychotic.

If your next aggregate safety review pools studies graded on CTCAE v5.0 and v6.0, book a 15-minute conversation.