# Connector health-canada-dpd — Health Canada Drug Product Database (DPD)

# Terms and licence (quoted)

The DPD web application carries no licence statement of its own (its footer links to the Canada.ca terms and conditions, verified 2026-09-11). The DPD data extract is published on the Open Government portal (dataset bf55e42a-63cb-4556-bfd8-44f26e5a36fe, "Drug Product Database - All Files") with "Licence: Open Government Licence - Canada". From that licence (https://open.canada.ca/en/open-government-licence-canada), stored verbatim in manifest.termsNotes:

The Information Provider grants you a worldwide, royalty-free, perpetual, non-exclusive licence to use the Information, including for commercial purposes, subject to the terms below. You are free to: Copy, modify, publish, translate, adapt, distribute or otherwise use the Information in any medium, mode or format for any lawful purpose. You must, where you do any of the above: Acknowledge the source of the Information by including any attribution statement specified by the Information Provider(s) and, where possible, provide a link to this licence. […] you must use the following attribution statement: Contains information licensed under the Open Government Licence – Canada.

CancerIndex shows every DPD record as a dated, sourced market authorization for one product (DIN) — never as an indication (the DPD publishes none) and never as a treatment recommendation.

Rate limits: none documented (2026-09-11). The connector self-limits to 2 requests/s, one at a time (rateLimits), i.e. ≈ 1,500 products × 4 requests ≈ 50 min per full pass; the pass is restartable per product (below). No 429 was observed during the first full pass.

# Verified endpoints (2026-09-11)

Schedule/form endpoints exist but are not fetched (4 requests per product: product, ingredients, status, routes).

# Ingestion

Scope (bounded, CLAUDE.md §227): the oncology therapeutic-class rows, sorted by drug_code. ctx.cursor.lastDrugCode is advanced after each fully persisted product (checkpoint every 50 records or 60 s, manifest.checkpointEvery), so a run stopped by the time budget (--max-minutes) or a signal resumes exactly where it stopped; a completed pass (cursor.completedAt) starts a fresh one on the next schedule (passCount increments; change events are recorded from pass 2).

1. Anomaly guard: ctx.guardCount('product', n) on the filtered bulk list (1,539) against the previous successful run; plus a hard floor of 10,000 bulk rows (anomaly: error) — a truncated list never triggers deletions (nothing is ever deleted by this connector).
2. Per product (fetchBundle): product → skip if missing; ingredients, status, routes. Each sub-object is validated (zod) and observed for schema drift (product, active_ingredient, status, route). One product source record per drug_code with the composite payload {therapeuticClass, product, ingredients, status, routes} — self-contained, so --mode backfill can replay it. Non-human classes (Veterinary, Disinfectant) are recorded and then skipped.
3. Ingredient → drug (deterministic, no LLM; parseIngredientName): the bracketed salt is split off ("IMATINIB (IMATINIB MESYLATE)" → molecule IMATINIB, salt form IMATINIB MESYLATE), a trailing salt token is stripped ("DOXORUBICIN HYDROCHLORIDE" → DOXORUBICIN; the salt list is SALT_TOKENS shared with trial reconciliation). Excipients listed as kit components (WATER, SODIUM CHLORIDE, BUFFER SOLUTION…) are skipped. Lookup in drug_aliases.normalized (index loaded once per run and kept in sync): the full name first, then the molecule. Exactly one drug → resolved (match_type ALIAS); several → the one whose alias is generic, else the one whose own name matches, else ambiguous → unresolved_labels (reason ambiguous_alias).
   • Unknown molecule, single-ingredient product → a drug is minted (mintId 'DRUG', name = molecule in Title Case — "Vinblastine", not "Vinblastine Sulfate" (CLAUDE.md §7: the base molecule is the entity, salts are aliases), slug = slugify(name) with a -hc-<drug_code> suffix on clash, kind null, developmentStatus 'marketed_ca' when the DPD status maps to approved), aliases generic (molecule), salt (salt form) and brand (cleaned brand), change event created.
   • Unknown molecule, multi-ingredient product → unresolved_labels (reason component_of_multi_ingredient_product, with DIN, brand, ATC); the approval row is linked to every component that did resolve.
4. Codes and aliases for every resolved drug: drug_codes atc (code tc_atc_number, label tc_atc), din (label = brand as published), hc_drug_code (label = brand), match_type EXACT_IDENTIFIER; a brand alias with the presentation noise removed (cleanBrandName: "PROCYTOX TABLETS 50MG" → "PROCYTOX", "THIO TEPA INJ 15MG/VIAL" → "THIO TEPA") and a salt alias when the ingredient carried one.
5. Approval row per (product, drug) — drug_approvals: jurisdiction 'CA', authority 'Health Canada', approval_type 'DIN', application_number = DIN (8 digits), approval_date = original_market_date (null when the product is approved but not marketed — never fabricated), status: Marketed / Approved / Authorized… → approved; Cancelled Post Market / Pre Market / (Safety Issue) / (Unreturned Annual) / Dormant → withdrawn with withdrawal_date = history_date; anything else → no row + warning (statusUnknown). The DPD status is kept verbatim in raw.dpdStatus and surfaced on the pages: a cancelled or dormant DIN is one product's status, not a withdrawal of the molecule (KEYTRUDA DIN 02441152 is "Cancelled Post Market" since 2019-12-04 while other pembrolizumab DINs are marketed). indication is honest text: "Marketed in Canada as KEYTRUDA (DIN 02441152) under ATC L01FF02 PEMBROLIZUMAB. Indications are not published in the Drug Product Database — see the Health Canada Product Monograph."; cancer_id NULL, tumor_agnostic false — no cancer is ever inferred from an ATC class. raw keeps drug code, DIN, brand, statuses/dates, ATC (code, label, group), class, company, descriptor, all ingredients with strengths, routes, last_update_date.
6. Provenance: one row per product and run when the source record changed (sourceUrl = public product page, dataset 'Health Canada DPD', datasetVersion dpd-<run date>, evidenceType regulatory, accessLevel open, geography 'CA', publishedAt = original market date, updatedAtSource = last_update_date, methodology text naming the ingredient and how it was reconciled). Unchanged product + existing row → the previous provenance row is reused (no duplicate evidence per weekly run).
7. Idempotency: rows are keyed by (source_id, drug_id, jurisdiction 'CA', application_number = DIN) — existing rows are updated in place (status changes are recorded as approval_status_changed from pass 2), new rows inserted (approval_added from pass 2). Nothing is deleted: a DIN that disappears from the bulk list keeps its last known status.

dry_run walks the first 20 products of the list without writing (resolution against the alias index is logged as "→ CI-DRUG-…", "would mint …" or "AMBIGUOUS"). --mode backfill re-derives drugs, codes, aliases and approval rows from the composite payloads in the raw lake — no HTTP, cursor untouched.

Health check: therapeuticclass?id=92551 must contain L01FF02.

# Observed runs (local copy of production, 2026-09-11)

Counts after those 455 products (plus the 10-record smoke): 507 product source records; 52 drugs minted (Vinblastine, Busulfan, Dactinomycin, Procarbazine, Thiotepa, Bleomycin, Interferon Alfa-2b…); ingredients resolved by alias 303 + 106 after salt stripping; 502 CA approval rows on 161 drugs (165 approved, 337 withdrawn — old DINs of long-marketed molecules dominate the low drug codes); drug_codes 502 DIN + 502 hc_drug_code + 119 ATC; aliases 186 brand + 52 generic + 47 salt; 8 labels in unresolved_labels (components of multi-ingredient kits: RIBAVIRIN of REBETRON, ANCESTIM of STEMGEN, FORMESTANE, MELACINE lysate components…), 0 ambiguous. pnpm cix reconcile-drugs afterwards linked 69,207 → 74,435 trial intervention rows (+5,228; 5,155 of them to the minted drugs — e.g. Busulfan, Bleomycin, Vinblastine trials were unresolved before).

# Limitations / notes

• No indications: the DPD is a product registry; indications live in the Product Monograph (PDF, not in the API). Canadian rows therefore never carry a cancer and never feed cancer-scoped pipeline stages or approved_drug_count per cancer; they do make the unscoped pipeline row "approved" and appear in the approvals feed and on the drug page.
• DIN-level granularity: one molecule has many DINs (brands, generics, strengths); the feed counts records, not molecules. Status is per DIN.
• Dates: original_market_date is the first marketing date of that DIN, often a generic's date; approval_date on a CA row is therefore not "the Canadian approval of the molecule" — the earliest CA row of a drug is the best available proxy. Historic products carry year-end placeholders from the registry itself ("1954-12-31") — kept as published.
• Scope by ATC: products without an ATC assignment in the DPD, or classed outside L01/L02/L03/ V10 (e.g. supportive-care antiemetics, L04 immunosuppressants), are not ingested. L03 brings non-oncology immunostimulants (interferon beta for MS) — they are drugs, honestly recorded with their ATC group, but should not be read as oncology approvals.
• Minted drugs start with no kind, mechanism or targets; ChEMBL/openFDA enrichment and pnpm cix reconcile-drugs (trial interventions) run afterwards. Salt-form duplicates against existing entities are avoided by minting the base molecule; remaining duplicates are proposed by packages/ranking/src/drug-duplicates.ts (entity_merges, status proposed).
• Only human and radiopharmaceutical classes are ingested; veterinary/disinfectant products are recorded as source records and skipped.