Information Sources & Citations

An honest description of where the information shown in the HealthLens app comes from, what is verifiable to a primary source, and what is generated by an AI model whose training data is not directly traceable on a per-statement basis.

Effective date: 9 May 2026 · Last updated: 9 May 2026 · Version 1.0

1. Source map at a glance

2. AI-generated content (no per-statement citation)

The plain-language explanations, drug summaries, interaction warnings, symptom assessments and chat answers are produced by a general-purpose Large Language Model (LLM) hosted on Microsoft Azure OpenAI Service. The model was trained by OpenAI on a broad corpus of public text. Two consequences follow:

• We cannot give you a citation for an individual sentence. The model does not retrieve from a known database when it answers; it generates text token-by-token from its parameters. We do not run a retrieval-augmented generation (RAG) pipeline against a curated medical knowledge base in this version of the app.
• The model's knowledge has a training cut-off and may not reflect the latest dosing changes, recalls, or guideline updates. The model may also produce confident-sounding hallucinations — statements that read like fact but are wrong.

For these reasons, every AI-generated statement in the app must be treated as a starting point for a conversation with a clinician, not as a clinical fact you can rely on. See the Medical Disclaimer.

3. OCR text extraction

Optical character recognition is performed by Google ML Kit's on-device text-recognition model. The image and the OCR happen entirely on your device; the image is not uploaded for OCR. Citation:

• Google ML Kit Text Recognition v2 — developers.google.com/ml-kit/vision/text-recognition/v2

OCR errors are common on handwritten, low-contrast, or rotated reports. The app does not auto-correct values that look medically implausible.

4. Translation

Short user-interface strings and short fragments of report text are translated using the MyMemory Translation API by Translated S.r.l. Longer summaries are translated by Azure OpenAI. Citation and attribution:

• MyMemory — mymemory.translated.net
• Translation memories accessible through MyMemory's free API are made available under a Creative Commons Attribution-ShareAlike licence. We do not republish raw translation memory entries; we only render the translated string back to you in-app.

5. Reference ranges & lab values

Reference ranges (the "normal/abnormal" markers next to a value) are taken from your own report as printed by the originating lab. The app does not impose a global reference range. This matters because:

• different labs use different methods, instruments, and population calibrations — a "high" at one lab can be "normal" at another;
• units differ between countries (mg/dL vs mmol/L for glucose; ng/mL vs nmol/L for vitamin D);
• special populations (paediatric, pregnant, athletes) have different normal ranges that a generic LLM cannot infer from a single lab printout.

6. Authoritative sources to verify against

For any value, dose, interaction, or interpretation you are about to act on, verify against an authoritative primary source. The list below is non-exhaustive but covers the most common starting points:

Drug information

• India — Central Drugs Standard Control Organization (CDSCO): cdsco.gov.in
• India — National Formulary of India (Indian Pharmacopoeia Commission): ipc.gov.in
• India — MIMS India / IDR / KIMS Drug Reference (subscription)
• USA — FDA DailyMed (full prescribing information): dailymed.nlm.nih.gov
• USA — openFDA (drug labels, recalls, adverse events): open.fda.gov
• USA — National Library of Medicine RxNorm: nlm.nih.gov/research/umls/rxnorm
• USA — MedlinePlus: medlineplus.gov
• UK — British National Formulary (BNF) / NICE: bnf.nice.org.uk
• UK — Electronic Medicines Compendium (emc): medicines.org.uk/emc
• EU — European Medicines Agency (EMA): ema.europa.eu
• WHO Model List of Essential Medicines: who.int — Essential Medicines
• WHO ATC/DDD index: atcddd.fhi.no
• DrugBank Online (mixed open / commercial): go.drugbank.com
• The patient information leaflet (PIL) printed inside your medication's box — the most reliable, country-specific source for that exact product.

Clinical guidelines & lab interpretation

• NICE Guidance (UK): nice.org.uk
• UpToDate (subscription, used by clinicians worldwide): uptodate.com
• Cochrane Library: cochranelibrary.com
• WHO Guidelines: who.int/publications/who-guidelines
• PubMed (peer-reviewed primary literature): pubmed.ncbi.nlm.nih.gov
• ICMR (India): icmr.gov.in
• Indian clinical society guidelines — API, ICOG, IAP, CSI, IDA — for condition-specific Indian-context recommendations.

7. Drug-drug and drug-food interactions

Interaction warnings shown by the app are generated by the LLM at the moment you ask. They:

• do not consult a versioned interaction database (e.g. Lexicomp, Micromedex, Stockley's, DrugBank Interactions);
• do not consider your dose, renal/hepatic function, age, weight, pregnancy, lactation, pharmacogenetic profile, or smoking/alcohol exposure;
• may miss minor or moderate interactions;
• may flag spurious interactions for medications that share a name fragment.

Always confirm any interaction warning with a licensed pharmacist or physician, with the BNF / Stockley's / DailyMed entry for the specific products you take, and with your country's regulator.

8. Pill identifier

The pill identifier sends an OCR'd text snippet (and optionally the cropped image) to the LLM and asks it to identify the product. There is no lookup against an imprint-database such as the FDA's "Pillbox", Drugs.com Pill Identifier, RxList Pill Identifier Tool, or the IDA database. Therefore:

• identification is a best-effort guess and can be wrong;
• generic and counterfeit products with similar imprints can be confused;
• the dose strength shown is the LLM's guess, not a database fact.

Suggested verification sources for an Indian-market pill:

• the dispensing pharmacist;
• the printed strip / blister and the patient information leaflet;
• the manufacturer's website;
• the Indian Pharmacopoeia & CDSCO product-approval database;
• the original prescription.

9. Symptom checker

The symptom checker is an LLM-driven triage helper. It does not replace established triage tools such as the NHS 111 service, the Manchester Triage System, the Emergency Severity Index (ESI), or the WHO IMCI for paediatrics. The output should not be used as the sole basis for a decision to seek (or to avoid) emergency care.

10. AI Health Hub chat

The chat sees only the report content you have explicitly opened in the session, plus optional health-profile fields you have entered. It does not have:

• access to your wider electronic health record;
• access to medical literature beyond the LLM's training cut-off;
• any tool-use that retrieves real-time drug, lab, or guideline data;
• knowledge of your personal allergies or contraindications unless you have told it.

11. How information stays current

Because the app does not pull from a versioned drug or guideline database, the only ways the information improves over time are:

• upgrading the underlying LLM to a newer Azure OpenAI model with a later training cut-off;
• tightening the system prompts that instruct the LLM to recommend authoritative verification;
• adding, in a future release, a retrieval pipeline that grounds responses in a curated, versioned medical knowledge base. Until that ships, treat AI output as ungrounded by design.

12. Reporting an incorrect statement

If the app shows information you believe is medically wrong, please email contact@fcappstudio.tech with subject [Content Correction] and include:

• the exact wording shown by the app;
• a screenshot if possible;
• the authoritative source you would prefer it to align with (a DailyMed link, BNF entry, EMA SmPC, etc.).

We tune our prompts in response to credible reports.

13. Third-party attribution

• Google ML Kit — © Google LLC. On-device text recognition. Used under the ML Kit Terms of Service.
• Microsoft Azure OpenAI Service — © Microsoft Corporation; underlying models © OpenAI, Inc. Used under our enterprise agreement.
• MyMemory — © Translated S.r.l. Translation memory entries available under Creative Commons Attribution-ShareAlike; we attribute MyMemory whenever a translated string from their public memory is shown.
• Plus Jakarta Sans — bundled UI font, used under the SIL Open Font Licence 1.1.
• Authoritative-source links in section 6 are provided for the user's convenience. We are not affiliated with, endorsed by, or sponsored by any of those organisations, and trademarks remain the property of their respective owners.