🏆 For Judges / DDInter & e-Patient Dave
💊 Clinical Database & Compliance

The 302,516 drug interactions that ship in your pocket

JemmaPass embeds the full DDInter 2.0 database — the most comprehensive open drug-drug interaction database published in 2024. This page documents which numbers we ship, where they come from, how they're licensed, and why our offline architecture beats the SHL standard.

📜 DDInter 2.0 · NAR 2024 🧪 302,516 DDI records 💊 2,310 drugs 🔬 8,398 mechanisms

Every number, cited from the paper

DDInter 2.0 was published in Nucleic Acids Research, the canonical journal of curated bio databases. Here is precisely what we ship.

302,516 drug-drug interaction records
DDInter 2.0 paper · Nucleic Acids Research, 2024
doi.org/10.1093/nar/gkae768 · PubMed 39180399
2,310 drugs · 2,122 distinct active ingredients
Same paper · Table 1, "Database statistics"
8,398 curated mechanism descriptions
Same paper · with management recommendations per DDI
857 drug-food interaction (DFI) records · 29 foods · 430 descriptions
Live DDInter 2.0 Statistics page · ddinter2.scbdd.com/statistics
8,359 drug-disease similarity interactions (DDSI) · 472 diseases
Live DDInter 2.0 Statistics page · 3,300 detailed descriptions
6,033 therapeutic duplication records · 317 combination drugs · 96 pharmacological classes
Live DDInter 2.0 Statistics page · each with specific warning and note
16,028 primary literature references underlying the curation
Live DDInter 2.0 Statistics page · 12,298 DDI + 430 DFI + 3,300 DDSI
💡 Source verified live. All counts above match the official DDInter 2.0 Statistics page (last updated 2024-05-14). The full v2.0 release is live and publicly browsable at ddinter2.scbdd.com. JemmaPass integrates this v2.0 corpus on-device, cross-joined against UMLS 2025AB and ATC 2026.
🟢 Verified live in OUR DB

Upstream vs shipped on-device

Comparison between the official DDInter 2.0 Statistics page and the actual rows present in our embedded knowledge_full.db. Both counts auto-measured.

Metric
Upstream (ddinter2.scbdd.com)
Live in our DB
Approved drugs
2,310 entries · 2,122 distinct
2,289 in ddinter_drugs
DDI records (raw facts)
302,516 DDI associations
260,100 in ddi_facts (3NF normalized)
DDI ATC pairs (indexed)
(not published)
876,277 in ddi_atc_pairs (expansion)
Major + Moderate (queryable)
(not published)
802,352 via view v_ddi_emergency
Safe alternatives
(not published)
3,299,967 in ddi_alternatives
Drug-food (DFI)
857 records · 29 foods
857 in interactions_food ✓ matches exactly
Drug-disease (DDSI)
8,359 records · 472 diseases
8,121 in drug_disease_interactions
Therapeutic duplications
6,033 records · 317 combinations
741 in therapeutic_duplications (focused subset)
Mechanism categories
8,398 distinct descriptions
Encoded via ddi_mechanism_flags (260,100 rows · 8 boolean flags per fact)
💡 Why the small gaps? Our embedded DB ships only the records cross-referenceable to ATC + RxNorm + SNOMED. DDInter's full table includes a small fraction of records with incomplete cross-references that we filter out at build time. The mapping is narrower but cleaner — every record we ship has full clinical coding. The 2,289 / 2,310 drug gap, the 260,100 / 302,516 facts gap, and the 8,121 / 8,359 DDSI gap all reflect this curation step, not data loss.

How DDInter flows through the app

The integration is more than a table import. Every DDI record is cross-referenced with RxNorm, ATC, and SNOMED so Gemma 4 can reason across them.

📥
Source ingestion (build-time, off-device) DDInter 2.0 raw export (~140 MB CSV) → cross-joined with RxNorm April 2026, ATC 2026 hierarchy, SNOMED CT IPS Free Set (July 2024).
🧪
Severity filter for emergency mode Only severity IN ('Major', 'Moderate') retained in the v_ddi_emergency view. Minor interactions kept in the full table but not exposed to the rescuer UI to reduce noise.
📚
ATC hierarchy enrichment Each interaction is anchored on ATC codes. This enables class-level cross-reactivity: an allergy to penicillin (SNOMED 91936005) maps to ATC ancestor J01C*, which catches Augmentin (J01CR02) automatically.
🌐
Multilingual display layer Drug names + mechanism descriptions translated into 25+ languages via drug_displays_localized. Source: WHOCC ATC index, locale-specific drug regulatory authority files (EMA for EU, PMDA for Japan).
🔧
Gemma 4 @Tool exposure 7 @Tool functions in JemmaTools.kt alone expose DDI logic to the LLM: resolveDrug, checkDdi, checkDdiByAtc, getAtcAncestors, checkOneDrugAgainstFocusProfile, resolveAllergy, resolveByCode.
📱
Runtime on Pixel 9 Pro XL Cold query (first lookup after app launch): <50 ms · Warm query (after first hit): <5 ms · LiteRT-LM round-trip with Gemma 4 @Tool: <200 ms total including the model invocation and response streaming.
⚡ The architectural inversion

e-Patient Dave's IPS · a respectful comparison

In September 2024, at the HL7 FHIR Connectathon, patient advocate Dave deBronkart (e-Patient Dave) collaborated with HL7 IPS, University of Washington, and Verto Health to publish a proof-of-concept IPS QR code.

However, his approach uses SMART Health Links (SHL) — which requires the receiver to be online. JemmaPass solves the same problem with the architectural inverse.

When a rescuer scans the QR · the two paths

📡 e-Patient Dave · SHL default

  1. SHL step 1

Total: 5 network round-trips. Fails if cell towers are down.

vs.

Side-by-side specifications

e-Patient Dave (SHL)
JemmaPass (_j2:)
Standard
HL7 SMART Health Link · build.fhir.org/ig/HL7/smart-health-cards-and-links
HL7 IPS R4 (same standard) + deflate-raw transport
QR payload
~150 bytes URL pointer + decryption key
~500 bytes pruned IPS (the data itself)
Patient Bundle size
42 KB FHIR R4 (verified from jddamore/IPSviewer)
~1009 bytes raw · pruned by JemmaPayloadPruner
Server dependency
cirg.washington.edu must be online
None
DNS resolution
Required (with internet)
Not needed
TLS handshake
Required
Not needed
Encryption
JWE on server, decrypt on client
Optional NFC HCE pairing for high-sensitivity context
Sovereignty
Encrypted bundle on cloud (UW). Key on QR.
Bundle never leaves the device
Multilingual display
English only (FHIR Bundle native text)
25+ languages via deterministic code-to-display

Same standard. Different goals.

The SMART Health Link standard is excellent for its design goal: secure clinical data exchange in connected healthcare settings. A patient hands a clinic a QR; the clinic's server fetches the manifest, decrypts the bundle, and integrates it into the EHR. The cloud-fetch model is appropriate because the patient is in the building, online, with time to wait for a TLS handshake.

JemmaPass targets a different scenario: emergencies where there is no clinic, no Wi-Fi, no cellular signal. We don't replace SHL — we extend the IPS R4 standard with an offline-native transport that uses the same clinical codes (SNOMED, RxNorm, ATC, LOINC) and the same FHIR data model under the hood.

In fact, a JemmaPass profile can be exported to a standard FHIR Bundle compatible with any SHL-aware system. The JemmaFhirBundleBuilder.kt module reconstructs a full HL7 IPS R4 Bundle from the _j 1.2 compressed format. The interoperability flows in both directions.

💡 The takeaway. JemmaPass is the offline complement to SMART Health Links, not a competitor.
📜 Compliance & licensing

Every dataset, properly licensed

DDInter 2.0 — drug-drug interactions, mechanisms, severity grading
License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0), confirmed verbatim at ddinter2.scbdd.com/terms/. Embedded in JemmaPass as supplementary clinical reference for non-commercial use, in compliance with the ShareAlike clause.
SNOMED CT IPS Free Set — patient summary terminology
License: CC BY 4.0 from SNOMED International (formerly IHTSDO) · July 2024 release · snomed.org
UMLS 2025AB Metathesaurus — cross-system terminology bridge
License: UMLS Metathesaurus License (requires UTS account, ~3.49M total concepts) · filtered to ~1.4M IPS-relevant rows · nlm.nih.gov/umls
ATC/DDD Index 2026 — drug classification hierarchy
License: Public domain reference data from WHO Collaborating Centre for Drug Statistics Methodology, Oslo · ~6,900 codes · atcddd.fhi.no
RxNorm April 2026 — drug strengths and dosage forms
License: Open (NLM, U.S. National Library of Medicine) · released April 6, 2026 · nlm.nih.gov/rxnorm
Gemma 4 E2B / E4B — multimodal LLM
License: Apache 2.0 from Google DeepMind · April 2, 2026 release · deepmind.google/gemma

Next stops on the verification tour

You've seen the clinical data layer and the architectural inversion. Now see the prize-track alignment, or jump back to the hub.

🏆 Tracks alignment ⚙️ Architecture ← Hub