How AI Medical Coding Works in Clinical Trials: From Verbatim to Standardized Terms

AI-assisted medical coding in clinical trials converts site-entered verbatim terms into standardized MedDRA or WHODrug terminology. It combines existing dictionary matching with language interpretation, candidate ranking, confidence-based routing, and medical coder review to make clinical trial coding faster while keeping decisions controlled and traceable.

Clinical trial data rarely enters a system in standardized language. An adverse event may be recorded as “severe headache,” “throbbing pain in head,” or “head pounding,” while medications may be entered using brand names, abbreviations, or misspellings. Before this data can be consistently reviewed and analyzed, verbatim terms must be mapped to standardized terminology such as MedDRA and WHODrug.

Traditional medical coding uses dictionary searches, exact matches, synonym lists, coding conventions, and coder review. AI-assisted medical coding adds an interpretation layer by normalizing language, identifying likely clinical concepts, ranking possible matches, and routing ambiguous terms for review or clarification.

The goal is not to replace medical coding judgment, but to move from site-entered verbatim data to consistent, traceable coded data more efficiently.

A verbatim is the original clinical information recorded by an investigator or site. Depending on the study, this may include adverse events, medical history, concomitant medications, indications, and other codable clinical information.

The challenge is that investigators do not necessarily describe the same concept in the same way.

For example:

• Headache
• Head pounding
• Severe pain in head
• Throbbing headache

Leaving these entries only as free text would make it difficult to consistently group and evaluate the same clinical concept across hundreds or thousands of subjects.

Medical coding provides that standardization.

Importantly, coding does not mean replacing the original information. The original verbatim remains traceable alongside the coded information. FDA study-data guidance, for example, describes the importance of providing both collected verbatim terms and coded terms in electronic study data submissions.

The terminology used depends on what is being coded, which is where MedDRA and WHODrug enter the workflow.

Traditional auto-coding is effective when a verbatim term already has a clear dictionary match, validated synonym, or predefined coding rule. It works well for straightforward and previously known terms where little interpretation is required.

AI medical coding extends this process by helping interpret terms that do not match cleanly. It can account for spelling variations, abbreviations, synonyms, contextual clues, and semantic similarity, then identify and rank the most relevant MedDRA or WHODrug candidates.

The key difference is not that traditional systems cannot automate coding. It is that AI can support the interpretation of less straightforward verbatim terms and help determine the most relevant coding options when exact matching is not enough.

In simple terms:

Traditional auto-coding identifies known matches.
AI medical coding helps interpret less straightforward terms and identify the most relevant coding candidates.

This distinction becomes clearer when we look at how AI medical coding works across the coding workflow.

AI medical coding works across the coding workflow, but its role is not identical at every stage. Some steps rely on structured rules and dictionary logic, while others use NLP, machine learning, or semantic interpretation to understand the verbatim and identify the most relevant coding candidates.

Together, these capabilities create a workflow that can handle straightforward matches efficiently while applying deeper interpretation when the reported term is less clear.

Step 1. Identify the Verbatim and Coding Context

The process begins by identifying both the source term and the context in which it was entered. The system may already know from the EDC field whether the entry represents:

• An adverse event
• A medical history condition
• A concomitant medication
• An indication

This context helps determine which coding dictionary and rules should apply.

Step 2. Check Exact and Existing Matches

The system first looks for reliable matches using available dictionary terms, validated synonyms, previously accepted mappings, and established coding rules.

If a clear match already exists, the term can move forward without unnecessary deeper interpretation.

Step 3. Normalize Unmatched Terms

When a verbatim does not match cleanly, the workflow may normalize the language before deeper interpretation. Depending on the system, this can involve predefined rules, NLP, AI, or a combination of these approaches.

Normalization can address:

• Misspellings
• Abbreviations
• Capitalization differences
• Word-order variations
• Synonyms
• Other linguistic variations

The purpose is to reduce language differences that may prevent a reliable match, while preserving the clinical meaning of the original verbatim.

Step 4. Interpret the Clinical Meaning with AI

For terms that remain unclear after normalization, AI can evaluate the wording and available context to identify the likely underlying clinical concept.

For example: “Throbbing pain in head” can be interpreted as a complete clinical expression rather than matched only word-for-word against dictionary terms.

For medication coding, other available context may also help distinguish between possible medicinal product matches.

Step 5. Search and Rank Relevant Dictionary Candidates

Once the likely concept has been identified, the system searches the relevant MedDRA or WHODrug terminology and ranks potential matches based on relevance or confidence.

This reduces the amount of manual dictionary searching required and helps direct coder attention toward the most relevant options.

Step 6. Decide What Happens Next

Identifying a candidate does not necessarily mean the coding process is complete. The next action depends on how clear the match is and whether the available source information supports it.

For example, a verbatim such as “heart problem” is too broad to identify a specific medical condition with confidence. AI may surface possible interpretations, but choosing one would introduce information that was not provided in the source. 

In such cases, the correct next step is clarification rather than forcing a coding decision.

AI can reduce searching and help prioritize likely candidates, but ambiguous clinical language still requires coding judgment.

A medical coder may need to determine:

• Whether the proposed term accurately preserves the verbatim meaning
• Whether the suggestion introduces information that was not reported
• Whether the proposed level of specificity is justified
• Whether the verbatim contains more than one medical concept
• Whether study-specific coding conventions have been followed
• Whether further clarification is required
• Whether the correct dictionary and version are being used

This is particularly important when the original information is incomplete.

For example, selecting a specific diagnosis when the site has reported only symptoms can alter the clinical meaning of the source information. MedDRA term-selection guidance explicitly cautions users not to add information that was not reported.

Human review is therefore not a failure of automation. It is one of the controls that allows AI to be used without turning an uncertain prediction into an unsupported coding decision.

Looking at the complete path makes the distinction between simple matching and AI-assisted coding clearer.

Human oversight remains an important part of AI medical coding, particularly for ambiguous or complex terms. Alongside human review, the workflow also needs controls that keep AI recommendations consistent, traceable, and appropriately managed throughout the study.  

Confidence-Based Routing

Not every AI suggestion should be treated the same way. High-confidence matches may be eligible for greater automation, while lower-confidence terms should be routed for coder review or clarification based on predefined rules.

Coding Decision Traceability 

The workflow should preserve a clear link between the original verbatim, the AI-generated recommendation, the coder decision, and the final coded term. This makes it possible to understand how a coding decision was reached.

Controlled Dictionary and Version Management

AI recommendations must be generated against the correct MedDRA or WHODrug version. When dictionaries are updated, previously coded terms may need to be reassessed or recoded, with any changes remaining traceable.

Audit Trail for Coding Decisions 

Changes to coding decisions, coder overrides, approvals, and recoding activities should be recorded so that the history of the coding process can be reviewed when needed.

Validated Coding Workflow

AI should operate within a controlled and validated coding process, with defined rules for how recommendations are generated, reviewed, accepted, or escalated.

Data Security and Access Control

Because medical coding uses clinical trial data, appropriate access controls, privacy safeguards, and security measures should apply throughout the workflow.

Together, these controls ensure that AI does not simply generate coding suggestions, but supports a workflow in which those suggestions can be consistently reviewed, governed, and traced. 

The operational value of AI medical coding comes from changing how coding work is distributed. Instead of medical coders spending the same level of effort on every verbatim, the workflow can separate straightforward terms from those that genuinely require interpretation.

This can help teams:

• Move clear and high-confidence terms through coding faster
• Reduce manual dictionary searching for unmatched terms
• Bring ambiguous or incomplete verbatims to attention earlier
• Direct coder effort toward lower-confidence and clinically complex cases
• Maintain coding more continuously as study data is collected
• Reduce the volume of unresolved coding work accumulating later in the study

The result is not simply more automation. It is a coding workflow in which human effort is concentrated where medical judgment provides the most value.

When evaluating AI medical coding software, sponsors and CROs should look beyond whether the system can simply suggest a code. The more important question is how well the complete coding workflow is controlled, reviewed, and connected to the underlying clinical data.

Key areas to evaluate include:

• Dictionary support: Does the system support the required MedDRA and WHODrug dictionaries and versions?
• Matching logic: Can users distinguish exact matches, synonym-based matches, and AI-generated recommendations?
• Confidence handling: How are confidence or relevance scores used to determine whether a term is auto-coded or routed for review?
• Coder review: Can medical coders easily view, confirm, change, or reject suggested terms?
• Alternative candidates: Does the system surface other relevant coding options when the top suggestion is uncertain?
• Query management: Can ambiguous or incomplete source information be queried directly from the coding workflow?
• Dictionary versioning: How are updates, up-versioning, and recoding handled?
• Traceability: Can the original verbatim, AI recommendation, coder decision, and final coded term be followed through the workflow?
• Audit trail: Are overrides, approvals, and subsequent coding changes recorded?
• EDC integration: Does coding remain connected to the source record, or does it require separate exports, transfers, and reconciliation?
• Source-data changes: What happens if the underlying clinical data changes after a term has already been coded?

The strongest AI medical coding systems are therefore not defined only by how quickly they generate suggestions, but by how reliably those suggestions can move through a controlled clinical trial coding workflow.

AI medical coding is most valuable when clinical trial data does not fit neatly into an exact dictionary match. It helps interpret inconsistent verbatim terms, narrow the most relevant MedDRA or WHODrug options, and direct unclear cases to the right next step.

The real improvement is not simply faster coding. It is a more structured way to handle both clear and uncertain terms, so automation can increase without losing clinical meaning, coder oversight, or traceability.

For sponsors and CROs, that is what makes AI medical coding useful in practice: not replacing the coding process, but making it more efficient, focused, and easier to control.

AI Medical Coding with Clinion

Clinion integrates AI-assisted medical coding directly within its EDC, keeping site-entered verbatim terms, coding suggestions, coder review, queries, and final coded data connected in the same clinical data workflow. AI and NLP help interpret terms and surface relevant MedDRA and WHODrug candidates, while medical coders retain control over ambiguous or complex decisions. This integrated approach reduces separate data transfers and reconciliation, while maintaining traceability from the source entry through the final coding decision.

What Is Artificial Intelligence (AI) in Medical Coding? - AAPC 
(PDF) AI in Medical Coding: Transforming the US Healthcare System
MedDRA
WHODrug Global | UMC