io.github.neptun2000/heor-agent

Compute disproportionality statistics (PRR, ROR, IC/BCPNN, MGPS/EBGM) on user-supplied drug-AE case counts and decide a signal verdict per EMA GVP Module IX rev 2. Returns the verdict (no/strengthening/confirmed/previously known/refuted), workflow recommendations, and canonical RMP signal-section text. To classify a finding as previously_known_signal, supply BOTH prior_known_signals AND reported_event so the tool can match the disproportionality to a specific known event. Optionally layers in GVP Considerations P.III pregnancy follow-up (birth/3mo/12mo) when both pregnancy_exposure and rmp_has_pregnancy_concern are true. Pairs with pv_classify. ⚠️ MGPS uses single-stratum gamma-Poisson shrinkage in v1 — when database counts are confounded by sex/age strata, EBGM/EB05 may be inflated; stratified MGPS is planned for v2. v1 takes user counts; v2 will also integrate EVDAS programmatic access per Reg. 2025/1466.

Triage already-collected social-media posts for pharmacovigilance reportability per EMA GVP Module VI. For each post the caller supplies a structured assessment of which of the four ICSR elements are present (identifiable_reporter, identifiable_patient, suspect_product, adverse_event) plus optional sentiment / themes / adverse_event_terms. The tool applies the deterministic four-element validity test (all four → valid reportable ICSR with a 15-day serious / 90-day non-serious clock; product+event only → follow-up needed; no AE → qualitative insight only), tallies sentiment/themes/AE terms, and surfaces the reporting obligations and validity caveats. It does NOT scrape and does NOT run NLP — the calling model extracts the per-post fields, keeping the output auditable. Pairs with rwe.social_listening_protocol (study design) and pv.signal_workflow (signal scoring). ⚠️ Social data is the lowest-validity RWE source — sentiment/theme output is hypothesis-generating only. Pure logic, no external calls.

Look up current regulatory approval status for a drug from primary sources (OpenFDA for US, EMA EPI for EU). Returns approved indications, label text verbatim, age/weight/sex constraints, black-box warnings, REMS status, contraindications, source URLs, and fetch timestamp. Refuses to assert 'not approved' on database absence — emits 'unknown' instead. 24h cache with force_refresh flag. Use BEFORE any claim about approval status, pediatric indication, or regulatory restrictions. Design log #25.

Select an appropriate real-world-evidence (RWE) study design for a research question. Ranks the five core RWE methodologies — retrospective database analysis, survey, literature review, chart review, social-media listening — against your research objective, the data you can access, the decision context (HTA/payer, regulatory, clinical guideline, internal strategy, exploratory), and the rigour required (exploratory / supportive / submission-grade). Returns a primary recommendation + ranked alternatives, each with its results-validity tier, bias caveats, and the downstream tools in this server that operationalise it (e.g. literature.search, pv.signal_workflow, evidence.population_adjusted). Flags when no feasible method can meet the requested rigour and suggests triangulation across complementary designs. Pure decision logic, <200ms, no external calls. Enum values are case-insensitive.

Generate a social-listening (social-media listening) study protocol + compliance checklist for an RWE study. Given the drug, indication, objectives (patient_experience, sentiment, adverse_event_monitoring, unmet_need, treatment_adherence, disease_awareness), platforms, privacy jurisdictions, and seed keywords, it produces a full protocol: objectives & scope, search strategy, inclusion/exclusion, analysis plan (sentiment + thematic + MedDRA-coded AE extraction), data-governance/privacy/ethics (GDPR Art. 6/9 and HIPAA items switch on jurisdiction), a MANDATORY pharmacovigilance-handling section (GVP Module VI applies to any systematic digital-media review, regardless of objective or channel ownership — escalated when is_mah_managed_channel is true), deliverables, a mandatory/recommended compliance checklist, and limitations. Pairs with pv.social_listening_triage (execution) and irb.review. ⚠️ Social-listening is the lowest-validity RWE method — outputs are qualitative/hypothesis-generating; triangulate with higher-validity designs. Pure logic, no external calls; enum values case-insensitive.

Generate a formal PRISMA-ready systematic-review search strategy from a research question / PICOS. Returns numbered Embase + PubMed syntax blocks per review strand (economic, clinical, HRQoL, epidemiology). Live PubMed counts via NCBI E-utilities: per-concept lines, concept-only combined, and per-strand full queries (concept AND SIGN filter NOT exclusions AND limits). Embase blocks are paste-ready — Embase hit counts NOT computed. When a single strand is requested, pipeline_inputs.pubmed_query is the full strand query; with multiple strands it stays concept-only (see pubmed_strand_counts). Embase Emtree terms are DRAFT — verify in Emtree.

Fetch enrollment/eligibility criteria for a clinical trial from ClinicalTrials.gov by NCT number. Returns structured eligibility criteria (inclusion/exclusion), primary outcomes, study design, and population description. Use this BEFORE building any trial-vs-real-world-population gap analysis. Always call this tool instead of using training-data knowledge about trial enrollment.

Validate URLs by making HEAD requests and checking HTTP status codes. Returns categorization: working (200), browser_only (401/403/407 from bot-blocking sites that work in browsers), broken (404/410), or timeout/error. ALWAYS use this before presenting reference links to users — broken links destroy trust. Pass all URLs you plan to cite.

Verify DOI/PMID citations against Crossref + NCBI E-utilities. Returns per-citation status: verified (resolved), fabricated (format-valid identifier that resolves to NOTHING — a likely invented DOI), unverifiable (network error or malformed identifier — never asserts fabrication), or url_only (a URL with no resolvable DOI/PMID). Use this to catch hallucinated references BEFORE presenting them — general LLMs fabricate ~36-40% of DOIs. Pass up to 100 citations.

Orchestrate the end-to-end living-evidence pipeline — AI-augmented SLR → living knowledge base → JCA/HTA deliverables ('from review to reimbursement'). Returns the ordered runbook of tool calls to make: for stage='baseline', the full chain (literature.search → screen → risk_of_bias → triangulation → network/indirect → cost_effectiveness → budget_impact → gap_analysis → claim_registry → dossier/JCA → consistency_check → living_review init); for stage='refresh', only the steps triggered by the signals you pass (material_change + recommended_downstream from literature.living_review, drifting_claims from evidence.consistency_check) so an unchanged refresh is a near no-op. This is a deterministic runbook generator — it does NOT execute the steps or hold state (the calling agent runs them; the host owns persistence + the refresh schedule). Use it to drive or document the living pipeline. Enum values case-insensitive.

Run the canonical MAIC discovery+screening pipeline in one call: ITC feasibility + parallel literature_search (broad + per-trial) + PICO screening + evidence_network. Does NOT auto-run risk_of_bias (it needs named trials; the report directs you to run risk_of_bias on the screened set) and stops short of running MAIC/Bucher itself (those require IPD or trial-level effect estimates). Produces a structured 9-section report with explicit Next Steps. Use this as a one-shot orchestrator instead of asking Claude/ChatGPT to chain the underlying tools manually.