Adverse-Event Coding & Causality

Clinical / PharmacovigilanceDrug-safety physician (PV)

Given an individual case safety report (ICSR) narrative — a patient on a study drug who experienced one or more adverse events — extract every reportable event, code each to the correct MedDRA Preferred Term and System Organ Class using the provided dictionary, determine ICH E2A seriousness, and assess WHO-UMC drug causality (temporality, dechallenge/rechallenge, alternative etiology, label). The agent has read-only tools — the case narrative, a MedDRA lookup, and the drug-class label — and must reason itself. Uses a teaching MedDRA subset, not a licensed distribution.

Why this is fundable

Scarce expert who grades this: Drug-safety physician / pharmacovigilance specialist (MD or PharmD, ~$200–400/hr loaded)
What one decision is worth: Mis-coding or under-calling causality on a serious AE can miss a safety signal (patient deaths, black-box warnings, withdrawals) or trigger regulatory findings. PV is volume-heavy and gradable.
Real-world data sources: ICSR/FAERS case narratives, MedDRA dictionary, drug labels (SmPC/USPI). Teaching MedDRA subset here; production requires a MedDRA license and maps onto real safety-database cases.

Agent tools

list_casesget_casemeddra_lookupget_drug_label

Expert grading rubric

Dimension	5 (excellent)	1 (poor)
Event extraction completeness	Identifies every reportable adverse event in the narrative — including a serious event that is unrelated to the drug and secondary/lab events — without lumping distinct events or inventing ones.	Misses reportable events (e.g. overlooks the neutropenia behind a febrile-neutropenia admission, or drops the unrelated fracture), or merges separate events into one.
MedDRA coding accuracy	Maps each lay event to the correct Preferred Term and its System Organ Class using the dictionary lookup (e.g. 'low white count' -> Neutropenia / Blood and lymphatic system disorders; 'shortness of breath with infiltrates' -> Pneumonitis / Respiratory). PT and SOC are consistent with the dictionary, not guessed from memory.	Wrong PT or mismatched SOC, codes to a symptom when a diagnosis PT exists (or vice versa), or fabricates a code never returned by meddra_lookup.
Seriousness determination (ICH E2A)	Correctly classifies each event as serious/non-serious and names the right ICH E2A criterion (death, life-threatening, hospitalization, disability, congenital anomaly, medically important) — e.g. flags the hospitalized CRS/pneumonitis/febrile-neutropenia events as serious and the asymptomatic resolved lab abnormality as non-serious.	Calls a clearly serious (hospitalized/life-threatening) event non-serious or vice versa, or cites the wrong/no criterion, or conflates severity grade with seriousness.
Causality assessment quality (WHO-UMC)	Assigns a defensible WHO-UMC category with sound reasoning: weighs temporality (onset vs dosing), dechallenge/rechallenge, alternative etiologies (confounding meds, comorbidity, trauma), and whether the event is labeled — e.g. probable/certain for step-up-dose CRS, unlikely for the antibiotic-confounded transaminitis, unrelated for the traumatic fracture.	Reflexively blames or exonerates the drug, ignores a strong alternative etiology or the temporal mismatch, conflates seriousness with causality, or misuses the WHO-UMC categories.
Evidence faithfulness	Grounds every PT/SOC in an actual meddra_lookup result and every causality factor in the narrative or the drug label; no invented events, codes, lab values, or label claims; states uncertainty where the narrative is genuinely ambiguous.	Invents events or MedDRA codes, asserts label content not returned by get_drug_label, or contradicts the narrative (wrong timing, fabricated dechallenge).

Example queries

Code the adverse events to MedDRA PT/SOC and assess seriousness and drug causality for case ICSR-001 (a patient on a DLL3 T-cell engager). Justify each causality category.
For case ICSR-002 (a patient on a PD-1 checkpoint inhibitor), extract and code the adverse events to MedDRA PT/SOC, determine ICH E2A seriousness, and give a WHO-UMC causality assessment for each.
Review case ICSR-003 (a patient on a HER2 antibody-drug conjugate with abnormal liver labs). Code the events to PT/SOC, assess seriousness, and carefully assess drug causality given the patient's concomitant medications and history.
Assess case ICSR-005 (a patient on oral fluoropyrimidine chemotherapy). Code each adverse event to MedDRA PT/SOC, determine seriousness per ICH E2A, and assign WHO-UMC causality — be explicit about which event is drug-related and which is not.

Trajectories

model panel (compare side by side)

Model	Provider	Tier	Judge 1–5	Verdict
Claude Opus 4.8	anthropic	frontier	5.0	strong
GPT (frontier)	openai	frontier	4.4	acceptable
Claude Haiku 4.5	anthropic	small	3.8	acceptable
GPT-4o mini	openai	small	2.8	flawed

batch 20260606_021639

Query	Model	Tool calls	Time	Status
Code the adverse events to MedDRA PT/SOC and assess seriousness and drug causality for c	claude-haiku-4-5-20251001	6	13.3s	ok
For case ICSR-002 (a patient on a PD-1 checkpoint inhibitor), extract and code the adver	claude-haiku-4-5-20251001	5	10.3s	ok

batch 20260606_021404

Query	Model	Tool calls	Time	Status
Code the adverse events to MedDRA PT/SOC and assess seriousness and drug causality for c	claude-opus-4-8	10	31.6s	ok
For case ICSR-002 (a patient on a PD-1 checkpoint inhibitor), extract and code the adver	claude-opus-4-8	6	26.6s	ok
Review case ICSR-003 (a patient on a HER2 antibody-drug conjugate with abnormal liver la	claude-opus-4-8	7	27.2s	ok
Assess case ICSR-005 (a patient on oral fluoropyrimidine chemotherapy). Code each advers	claude-opus-4-8	4	21.0s	ok