MedDRA Coding Assist — Verbatim Term to Preferred Term by Meaning

Objective

Coding a reporter's free-text verbatim ("sore aching muscles in my legs") to the right MedDRA Preferred Term is slow and coder-dependent, and keyword lookups break on paraphrase. This recipe ranks candidate PTs by meaning, so the coder gets a short, confident shortlist to confirm — consistent coding without hard-coding a synonym dictionary.

Step 1: A small MedDRA PT dictionary with embeddings

In production this table is your licensed MedDRA PT list; here we use a handful. Each PT name is embedded once with EMBED().

DROP TABLE IF EXISTS meddra_pt;

CREATE TABLE meddra_pt (
    pt_code TEXT,
    pt_name TEXT,
    emb     VECTOR(384)
);

INSERT INTO meddra_pt VALUES
 ('10028411','Myalgia',        EMBED('Myalgia')),
 ('10019211','Headache',       EMBED('Headache')),
 ('10028813','Nausea',         EMBED('Nausea')),
 ('10013573','Dizziness',      EMBED('Dizziness')),
 ('10039020','Rhabdomyolysis', EMBED('Rhabdomyolysis'));

Step 2: Rank Preferred Terms for a verbatim

SELECT pt_code,
       pt_name,
       COSINE_SIMILARITY(emb, EMBED('sore aching muscles in my legs')) AS sim
FROM   meddra_pt
ORDER  BY sim DESC
LIMIT  3;

Result — Myalgia tops the shortlist for the leg-muscle verbatim:

Why it matters

The coder confirms the top suggestion instead of searching the dictionary — faster, and more consistent across coders. The absolute score matters less than the ranking: the correct PT is cleanly separated from the rest. Swap in your full PT list and the same query scales to the whole terminology.

Log the accepted code to a tamper-evident audit ledger so coding decisions are reconstructable at inspection.