MedDRA coding automation works when you automate the routine, high-confidence verbatims, route the ambiguous ones to a trained coder, and govern the whole loop with documented conventions. It fails when a tool is dropped in "out of the box" and nobody owns the synonym list, the version updates or the review queue. Below is how to build the first kind of setup in a US sponsor or CRO environment.
TL;DR
- Published studies show AI coding can raise accuracy and cut cost, but the evidence is strongest when AI and human coders work together, not when humans are removed.
- Auto-code routine verbatims with high confidence. Send multi-concept, medication-error and off-label cases to an expert.
- A governed synonym list and a documented set of coding conventions matter more than the choice of model.
- Context (age, sex, trial details) must reach the coding system, or query loops with sites will erase your time savings.
- Treat each MedDRA version release as a controlled change to your automation, not a background event.
Table of Contents
- What MedDRA coding actually involves
- Where automation helps most
- Where it breaks
- A workable design: confidence tiers with expert review
- Governance: the part that decides success
- Our perspective
- How to start: a 30-day readiness check
- Build versus buy versus partner
- Working with HaiPhai on coding and safety operations
- FAQ
- Sources
What MedDRA coding actually involves
MedDRA, the Medical Dictionary for Regulatory Activities, is the ICH-standardized terminology for adverse events, medical history and related data in clinical research and post-marketing safety work. A coder reads what the site wrote (the verbatim), selects the most appropriate Lowest Level Term (LLT), and that LLT rolls up to a Preferred Term (PT) and the higher levels of the hierarchy. The MedDRA Introductory Guide describes this five-level structure and why it allows retrieval by specific or broad groupings.
That sounds mechanical, and for many verbatims it is. "Headache" becomes a headache LLT in seconds. The difficulty sits in the tail: abbreviations, misspellings, two events in one phrase, and text that only makes sense with the patient's context. The MedDRA Points to Consider (PTC) Companion Document is blunt that small deviations in coding can produce misleading safety analyses. That is the standard any automation has to meet.
The basic principles have not changed with AI. MedDRA's own training material on coding basics lists them: do not alter MedDRA, always select a lowest level term, select only current terms, use medical judgment, and know when to request a new term. An automated system that violates any of these is not saving time. It is moving the error downstream.
Where automation helps most
The business case is real. A 2025 study of adverse event coding in Alzheimer's disease trials, published in PubMed, reported roughly a 20% gain in accuracy and an 80% reduction in coding cost for AI-based coding compared with manual clinician-led coding, on a sample of 1,000 adverse events collected between 2005 and 2020. It is one study in one therapeutic area, so treat the figures as a signal, not a promise for your program.
In pharmacovigilance, a French validation study used machine learning models on 11,633 patient adverse reaction reports from 27 pharmacovigilance centres. It reported AUC values up to 0.97 in internal validation, with an F-measure of 0.80 for the best-performing model. Those are internal results on one national dataset, which is a reminder that performance on your own verbatims has to be measured separately.
Newer work points the same direction. A hybrid system for automated MedDRA coding of adverse drug events and an LLM-based reasoning framework presented in 2025 both aim at scalable, consistent coding. Earlier FDA-related work, such as an evaluation of NLP systems for annotating drug labeling with MedDRA terms, shows the regulatory community has studied this for years.
The gains concentrate in three places:
- Volume of routine terms. Common, unambiguous verbatims are where a model is both fast and reliable.
- Consistency. A rule or model applies the same term to the same text on Monday and Friday, across coders and across studies.
- Turnaround. Faster coding means faster safety review and cleaner data at interim locks.
Where it breaks
Context gets lost
MedDRA contains terms that are non-specific, and a model matching text alone can pick the wrong one among similar-sounding options. The PTC Companion Document stresses that coding decisions can depend on information such as patient age, sex and trial specifics. If your EDC, safety database and coding tool sit in separate silos, the coding system never sees that context.
The cost shows up later. Wrong or vague terms trigger queries to sites, and the query loops can consume the time you saved at the coding step. This is a data-flow problem, not an AI problem, and it is the most common reason we see automation projects disappoint.
Hard concepts stay hard
Practitioners report that AI struggles most with medication errors, off-label use and product defects. Multi-concept entries are another trap. A verbatim like "fall and contusion" contains two events that a coder may need to split and assign separately. A tool that forces one term onto the whole phrase is quietly losing information.
"Out of the box" does not exist
Many teams expect a tool to work immediately. In practice, your sites write verbatims in their own shorthand, and your sponsor or CRO has its own conventions. The PTC Companion Document says organizations should establish and document their own coding conventions and synonym lists. Automation does not replace that work. It depends on it.
A workable design: confidence tiers with expert review
The pattern that holds up is a tiered one. The researchers behind the Alzheimer's study frame the goal as combining AI and human expertise, not removing the human. A practical version looks like this:
| Tier | Typical verbatim | Handling | Human role |
|---|---|---|---|
| 1. High confidence | Clear, single-concept, matches an approved synonym | Auto-code | Periodic sampling and audit |
| 2. Medium confidence | Misspelled, abbreviated, or near-match | Suggest term, coder confirms | Fast confirmation |
| 3. Low confidence or complex | Multi-concept, medication error, off-label use, product defect | Route to expert coder | Full medical judgment, split and assign |
| 4. No match | New or unusual wording | Expert review, possible term request | Decide whether to request a new MedDRA term |
Two things make this table real. First, thresholds are set from your own data, by checking auto-coded output against expert-coded output on a representative sample before go-live. Second, every override by a coder feeds back into the synonym list so the system improves under governance, not on its own.
Governance: the part that decides success
Synonym lists as controlled documents
A synonym list maps your verbatim language to MedDRA terms. Treat it like any controlled document: an owner, a change log, review by a qualified coder or safety physician, and a rule for retiring entries. It must be maintained alongside MedDRA releases so nobody inadvertently uses non-current terms.
Version management
MedDRA is updated twice a year. Version 29.0 was made available in English on 1 March 2026, and the Points to Consider documents are synchronized with MedDRA versions. MedDRA's coding and versioning guidance tells users to select current LLTs. For automation, that means each release triggers a defined routine:
- Load the new version into a test environment.
- Check which synonym list entries point to terms that were demoted, moved or made non-current.
- Re-run your validation sample and compare results with the previous version.
- Document the outcome and approve before production use.
Teams that skip this end up with a fast tool producing quietly outdated coding.
Documentation and auditability
For any auto-coded term you should be able to answer: which version, which rule or model version, which confidence score, and who reviewed it if anyone did. If a regulator or sponsor audit asks why a term was chosen, "the tool did it" is not an answer. Keep the model or rule configuration under change control, and record validation results the way you would for any system touching safety data.
Our perspective
We work with biotech and life sciences teams on operations, and coding is a good example of a pattern we see everywhere. The software is rarely the bottleneck. The bottleneck is that nobody mapped the workflow from the site's verbatim to the safety review, so the tool speeds up one step while the queries, handoffs and version changes stay manual.
Our approach starts from the outcome the team needs, such as a clean database at lock without a coding backlog, and works backward to find where time actually leaks. Sometimes that is the coding step. Often it is the missing context, the unowned synonym list, or a review queue with no clear routing rules.
We also hold a firm view on the human role. Automation should take the repetitive, high-confidence work so that your best coders and safety physicians spend their time on the ambiguous cases where medical judgment matters. A governed model with expert review is faster than a manual process and safer than an unsupervised one. A tool with no governance is neither.
That is why we embed with the team and design the adoption and governance model alongside the workflow, instead of handing over a tool and leaving. We describe the same principle for other functions in our guides on governed AI in clinical workflow redesign and aligning AI tools with clinical goals.
— John
How to start: a 30-day readiness check
Before buying or building anything, answer these questions with data from your own studies.
- Measure the baseline. How many verbatims per study, what share is coded without query, and how long does coding take from entry to approval?
- Sample the tail. Pull 200 to 500 coded verbatims and have a senior coder label them as routine, ambiguous or complex. This tells you how much is realistically automatable.
- Audit the data flow. Does the coding system receive age, sex, indication and study context? If not, fix that first.
- Inventory your conventions. Do you have a written coding conventions document and a synonym list with an owner? If not, create them before any model is trained or configured.
- Define success. Pick measures that include quality, such as agreement with expert coding and query rate, not just speed.
- Plan the version routine. Write down who does what when the next MedDRA release arrives.
If the audit shows that coding is only one of several slow steps, widen the lens. Regulatory writing and site activation often carry larger delays, and we cover those in our pieces on AI in regulatory document drafting and where AI helps in regulatory affairs.
Build versus buy versus partner
There are broadly three routes. Off-the-shelf coding tools built into EDC or safety platforms are quick to switch on and fit standard workflows, but they need your synonym list and review rules to perform well. In-house builds, such as SAS macros or custom models, give control, but they put validation, maintenance and version management on your own staff. An embedded operational partner designs the workflow, configures or integrates the tooling, and helps you stand up the governance model.
Which is right depends on your volume, your in-house data science capacity and how many other processes you want to redesign at the same time. We lay out the trade-offs in our decision framework on a fractional AI partner versus an internal team.
Working with HaiPhai on coding and safety operations
If your coding backlog is a symptom of a wider operational problem, our AI Velocity Diagnostic and operating partnership are built for that. We map where time is lost between the site and the database lock, redesign the workflow around your regulatory and scientific needs, integrate your institutional knowledge such as synonym lists and conventions, and put governed automation and monitoring in place. You can see the areas we cover on our solutions page, or reach out through the partnership page to discuss your situation.
FAQ
What is MedDRA coding used for?
MedDRA coding translates free-text adverse events, medical history and similar entries into standardized terms. That lets sponsors and regulators group, count and compare safety data across patients, studies and products. It is used in both clinical trials and post-marketing surveillance.
How do you automate MedDRA coding?
Start with a documented synonym list and coding conventions, then use rules or a trained model to auto-code high-confidence verbatims. Route ambiguous or multi-concept entries to an expert coder. Validate the output against expert coding on your own data before go-live, and re-validate after each MedDRA version update.
Can AI replace human MedDRA coders?
Not safely, based on the current evidence and guidance. Studies report real gains in accuracy and cost, but the researchers describe the best results as a combination of AI and human expertise. Coders shift toward review, ambiguous cases and governance.
What happens when MedDRA releases a new version?
Your automation has to be checked against it. Synonym list entries may point to terms that are no longer current, so you should test the new version, review affected entries, re-run a validation sample and document approval before production use.
Sources
- MedDRA Points to Consider Companion Document
- MedDRA Introductory Guide
- Coding with MedDRA: Basic Principles
- MedDRA Coding and Versioning
- English MedDRA Version 29.0 now available
- Artificial intelligence-enabled safety monitoring in Alzheimer's disease clinical trials
- Validation of AI to support automatic coding of patient adverse drug reaction reports
- A Hybrid System for Automated MedDRA Coding of Adverse Drug Event
- LLM-based reasoning framework for MedDRA adverse event coding
- Evaluation of NLP systems to annotate drug product labeling with MedDRA terminology