Oncoshield AI · Architecture briefing · Draft for discussion

How Oncoshield works, end to end

One page for the call with the clinical team: who uses the system, how a person's answers become advice, how the medical rules are kept up to date, and what happens when two guidelines disagree. Items marked proposed are open for decision.

1 · The whole system

System map

Eight parts. The patient app, rules engine and evidence registry exist today as an MVP prototype, not yet built for production. The rest is what turns the prototype into a service hospitals and districts can run.

MVP works in the current prototype; not production-ready Planned phase 1–2 build Future phase 3 or open question

Oncoshield AI

A confidential risk check that turns a person's answers into a screening and referral plan, where every piece of advice traces back to a named guideline and a clinician's sign-off.

Patient app
  • MVPAdaptive questionnaire by age and sex
  • MVPSeparate pathway for under-18s
  • MVPPlan with actions first, alerts after
  • MVPMarathi and English
  • PlannedSave plan with phone OTP
  • PlannedReal booking and callback
Rules engine
  • MVPAnswers become risk variables
  • MVP64 rules, each with evidence links
  • MVPEvery result frozen and fingerprinted
  • MVP25 automatic test cases
  • PlannedRules stored in a database, not code
  • Planned"What changes" preview before publishing
Evidence registry
  • MVP45 guideline sources tracked
  • MVPPer-rule agreement and divergence notes
  • Planned231 extracted recommendations loaded
  • Planned7 more ICMR documents added
  • FutureAutomatic watch for new versions
Clinical governance
  • MVPRead-only rules console
  • PlannedReview cases by tumour panel
  • PlannedOnline voting with reasons
  • PlannedWritten precedence policy
  • PlannedSign-off and version release
Camp and doctor tools
  • MVPCamp mode (demo)
  • PlannedHealth-worker login, works offline
  • PlannedDoctor sees referred patients
  • PlannedDoctor records test outcome
  • PlannedDoctor rates the advice given
AI assistance
  • FutureSpots new or changed guidelines
  • FutureDrafts recommendation extracts
  • FutureSummarises doctor feedback
  • FutureFlags conflicts between sources
  • FutureDrafts patient wording and Marathi
Admin and partners
  • MVPProgramme dashboard (demo data)
  • PlannedUsers, camps, facilities
  • PlannedDistrict / hospital dashboards
  • FutureWhite-label brands per partner
Platform and compliance
  • PlannedLogins and role permissions
  • PlannedTamper-evident audit trail
  • PlannedConsent records (DPDP Act)
  • PlannedHosting in India
  • FutureHospital systems and ABHA link
  • FutureMedical-device regulatory path
2 · Logins and permissions

Who uses it, and what each person can do

The public never needs an account. Everyone else signs in, and sees only what their role needs. Every view of identifiable data is logged.

CapabilityPublic
no login
Patient
phone OTP
Camp / health workerDoctorClinical panelGovernance leadHospital / district admin
Take the risk check●●● for others––––
Save plan, see past checks–● own◐ own camp, time-limited◐ referred patients–––
Book, callback, referral slip●●●––––
Register people at a camp, offline––●––––
Record test outcome; rate the advice–––●–––
Review rules, vote with reasons–––◐ if on a panel●●–
Decide conflicts; approve and publish a ruleset–––––●–
Manage users, camps, facilities, branding––––––●
Dashboards (de-identified numbers only)––◐ own camp–●●●

● full access · ◐ limited as noted · – none. Oncoshield's own platform team sits above all of these for support and is audited the same way.

3 · Data flow

Patient journey: from answers to a plan, and back

The engine never guesses: answers are converted into variables, and only rules that a clinician has signed off can fire. The doctor's outcome closes the loop and feeds the review process.

Ways inanswersevidence linksadvicestoresreferraltest result + ratingfeeds review cases

QR at camp or hospital

Link or WhatsApp

Health worker, assisted

Consent and language

Adaptive questions
adult or under-18 set

Risk variables

Rules engine
signed-off ruleset only

Evidence registry
sources and versions

Personal plan
actions first

Frozen record
ruleset version + fingerprint

Booking, callback
or referral

Doctor

Outcomes

De-identified analytics

Clinical governance

Solid arrows run today in the prototype up to the booking step; the doctor, outcome and analytics steps are proposed. Dashed arrows are reference links, not patient data.
4 · The review loop

Keeping the medical rules current

This replaces emailing Word files. Each Word file we prepared becomes one review case per tumour type inside the doctor panel: the same questions, answered on screen, with every vote and reason kept.

noyespattern foundno: second roundwith results shownyesstill split after round 2rationale recorded

New or updated guideline

Recommendations extracted
AI drafts, a person verifies

Changes or contradicts
a current rule?

Update evidence links

Review case opened
for the tumour panel

Doctor ratings and
outcome data

Panel votes online
keep, change or remove + reason

Agreement above
the agreed threshold?

Governance lead decides
using the precedence policy

Shown to patients as
'guidelines differ, discuss with a doctor'

Draft rule tested
+ preview of who is affected

Sign-off

New ruleset version live

App uses it; older results
stay reproducible

Two voting rounds with results shown between them is a modified Delphi process, the usual way guideline panels reach consensus.
5 · The hard question

When two guidelines disagree

The Word files collect opinions; they do not decide conflicts. What decides them is a short written precedence policy that the clinical team agrees once, then applies case by case. A draft is below.

symptom orwarning signroutine screening,average riskyesnoyesnohigh risk orinheritedyes

Two sources disagree

What kind of advice?

Use the most protective
referral threshold

Indian national
guidance exists?

Follow MoHFW / ICMR;
show others as 'discuss'

WHO guidance for
similar settings?

Follow WHO, adapted

Strongest international
source after tie-breakers

Specialist societies
NICE, ESMO, NCCN
adapted by the panel

Still tied?

Panel vote; no consensus
means shared decision

Draft precedence policy. Symptom advice errs toward referral because a missed cancer costs far more than an extra visit.

Tie-breakers, in order

  1. Certainty of evidence — graded systematic review beats expert consensus.
  2. Fit for Indian patients — incidence, age at diagnosis, TB prevalence, tobacco and areca use.
  3. Can the partner site deliver it — a mammogram we recommend must be available to book.
  4. Harm of over-testing — false positives, overdiagnosis, cost to the family.
  5. Recency — newer version of an equally strong source.
Worked example: PSA testing

ICMR (2023) says no population PSA screening in India. USPSTF says men 55–69 may choose after discussion. Rule: routine screening, Indian guidance exists, so ICMR leads. The app would stop prompting PSA for all men and keep symptom-led advice; the panel confirms or overrides with a recorded reason.

This approach follows the published method for adapting international guidelines to a local setting (GRADE-ADOLOPMENT and the Evidence-to-Decision framework): adopt, adapt or set aside each recommendation, and write down why.

6 · Scope of AI

Where AI helps, and where people decide

Proposed rule: AI prepares and flags; clinicians decide; the engine only runs signed-off rules. This keeps every piece of advice explainable and makes the regulatory case simpler.

AI assists

  • Watches guideline sites and flags new versions
  • Drafts recommendation extracts with page locators
  • Flags where a new source contradicts a current rule
  • Summarises doctor comments and ratings per rule
  • Drafts plain-language and Marathi wording for review
  • Answers "why did I get this advice?" from approved text only

People decide

  • Which recommendation becomes a rule
  • Thresholds: ages, intervals, symptom triggers
  • How conflicts are resolved, with a written reason
  • When a ruleset version goes live
  • Final patient-facing wording
  • Nothing changes in the engine without sign-off

Instead of an AI rating: a transparent confidence score per rule

Your idea of rating recommendations from doctor feedback works best as a score anyone can recompute, shown in the governance console. A drop below a set level opens a review case automatically.

Rule confidence (0–100) = evidence grade of the leading source (up to 30) + share of sources that agree (up to 20) + panel agreement at last vote (up to 25) + doctor "appropriate" ratings, 6 months (up to 15) + screening yield from recorded outcomes (up to 10)

Weights are a starting proposal for the panel to set. Every input is stored, so the score can always be explained.

7 · Sequence

Roadmap

Phases, not dates; timing depends on the pilot partner and team size.

Now

Prototype v4

  • Single-page app, brand applied
  • 64 rules, 45 sources, 25 tests
  • Read-only rules console
Phase 1

Foundation

  • Backend, logins, roles
  • Rules and evidence in a database
  • Review cases and online voting
  • Audit trail, consent, India hosting
Phase 2

Pilot

  • Camp app that works offline
  • Doctor portal with outcomes and ratings
  • Real booking, SMS/WhatsApp
  • Hospital or district dashboard
Phase 3

Scale

  • AI assistants and confidence score
  • White-label partners
  • Hospital system and ABHA links
  • Regulatory submission if required
8 · For the call

Decisions to take with the clinical team

Who sits on the clinical panel?

One reviewer per tumour group (9 groups, some can be combined), one governance lead, and the agreement threshold (for example 75%).

Adopt the precedence policy?

Indian national guidance first for routine screening; most protective threshold for symptoms; specialist societies for inherited risk. Edit, then sign.

Baseline review: on screen or on paper?

Online review needs phase 1. If the panel wants to start sooner, the Word files serve as the one-time baseline and are loaded into the system afterwards.

Do doctors rate advice per patient?

A one-tap "appropriate / not appropriate" on each referral. Agree how ratings are used: as input to review cases, never as automatic changes.

Scope of AI

Confirm "AI assists, people decide" as the rule for version 1.

Login for the public

No login to take the check; optional phone OTP to save the plan and get reminders.

Regulatory and data

Get advice on whether CDSCO treats Oncoshield as software as a medical device, and confirm DPDP consent wording and India hosting.

Pilot site

Which hospital or district, which camp, and who closes the loop on outcomes.

Suggested 45-minute agenda

  1. System map and who uses it (sections 1–2) — 8 min
  2. Patient journey (section 3) — 7 min
  3. Review loop and the precedence policy, with the PSA example (sections 4–5) — 15 min
  4. AI scope and confidence score (section 6) — 7 min
  5. Decisions and owners (section 8) — 8 min