Proof of concept / synthetic data only

Connected Microsoft IQ architecture. Research AMC proof of concept. Human review stays in control.

AMC IQ is a proof of concept for a Microsoft IQ intelligence layer in a research academic medical center. The primary story is a connected/live architecture that brings institutional knowledge, structured clinical data, care-team workflow, and external evidence into one cited assessment. A deterministic local fallback remains available for reproducible walkthroughs and repeatable testing.

Explore the source Read the architecture

Synthetic only. This proof of concept uses synthetic data, contains no PHI, and is not clinical decision support.

AMC IQ proof-of-concept interface with an editable free-text question, four IQ providers queued in parallel, and the Building Assessment workflow in progress. — Start with a curated question-bank prompt, edit it, or type any free-form question. The four IQ providers then retrieve connected or deterministic fallback evidence in parallel.

4IQ context layers

25Synthetic patients

10Fictional trials

0PHI or clinical use

The scenario

A research-AMC trial-readiness question that crosses data boundaries.

"Patient with metastatic NSCLC, EGFR exon 20 insertion, prior platinum therapy, ECOG 1, CrCl 48. The open trial mentions renal thresholds and prior-therapy exclusions. Is the patient eligible, and what should the care team do next?"

A useful answer needs more than search. The proof of concept reconciles protocol language, patient facts, institutional policy, current workflow, and public evidence while surfacing uncertainty instead of hiding it.

Concept walkthrough

One question becomes a governed, multi-source retrieval problem.

These diagrams show the idea behind AMC IQ: identify the evidence domains implicit in the question, retrieve across them deliberately, reconcile the results, and keep the final assessment subject to human review. Select a diagram to open it at full resolution.

A typical oncology trial-readiness question that requires multiple grounding sources. — **01 / The question** Concept framing for a cross-domain trial-readiness question. The abbreviated illustration label NCT-4324 maps to synthetic trial NCT99004324 in the runnable demo.

The clinical research knowledge ecosystem surrounding an oncology question. — **02 / The ecosystem** People and teams are context participants; EHR, genomics and pathology, protocols, and guidelines and policy are retrievable knowledge sources.

Question clauses mapped to the clinical and research evidence domains needed to answer them. — **03 / The evidence map** Each clause points to a different supporting domain. The abbreviated illustration label NCT-4324 maps to synthetic trial NCT99004324 in the runnable demo.

An agentic retrieval engine planning queries, selecting sources, and merging evidence. — **04 / The retrieval pattern** Iterative planning, knowledge-source selection, evidence merging, and a governed human-review task. Enabled integrations vary by deployment.

Four layers, one proof of concept

Each source answers a different part of the question.

Foundry IQ

Grounds the assessment in protocols, amendments, pathology, genomics, consent policy, and institutional operating procedures.

Institutional knowledge

Fabric IQ

Retrieves structured patient, laboratory, treatment, trial, enrollment, and scheduling facts through a Fabric Data Agent.

Structured clinical context

Work IQ

Adds coordinator ownership, tumor-board context, open tasks, PI availability, and the care team's current operating state.

Workflow context

Web IQ

Supplies current public evidence through a Bing-backed Azure AI Search knowledge source. It is a scenario stand-in for the limited-access Web IQ product.

External grounding

How the proof of concept works

Connected retrieval. Explicit uncertainty. Human review.

01
Interpret the question
Resolve the patient, trial, requested decision, and missing facts.
02
Retrieve in parallel
Dispatch live or deterministic Foundry, Fabric, Work, and Web context requests concurrently.
03
Match evidence to criteria
Separate met criteria from blockers, uncertainty, and required follow-up.
04
Draft a governed next action
Return citations, an owner, and a task that remains drafted and unsubmitted.

Evidence you can inspect

01 / Trace

Citations and assessment steps stay visible.

Findings link back to their source material, while the assessment trace shows the specialist queries, deepening, reconciliation, and governed composition steps.

Source-level evidence for every material finding
Visible retrieval and specialist activity
Human-readable impact and unresolved issues

AMC IQ evidence packet with Foundry, Fabric, and Work IQ citations beside a completed multi-agent assessment trace. — Every material claim remains connected to evidence and an auditable assessment step.

02 / Govern

The output ends in review, not automation.

The final assessment distinguishes potential match from confirmed eligibility, makes open issues prominent, and assigns a next action without submitting a task or sending a message.

Potential match Not ready for screening Ready for PI review Drafted, not submitted

Completed AMC IQ trial-readiness assessment showing criteria, open issues, next action, reviewer assignment, patient snapshot, and safety notice. — A cited recommendation for a coordinator and principal investigator to review.

Designed to be honest about the proof-of-concept boundary

Connected architecture first

The primary story is a connected Microsoft IQ architecture for a research AMC, with live Foundry, Fabric, Work, and external evidence integrations where access exists.

Deterministic local fallback

The full local walkthrough can run against deterministic synthetic data without an Azure account, authentication, or live service access so the experience stays reproducible.

Human review, no clinical claim

Tasks remain drafted, not submitted, and the cohort, protocols, identities, and assessments are fictional. AMC IQ is not validated for patient care, trial enrollment, or clinical decision support.

Open, inspectable, reproducible

Inspect the connected architecture or run the reproducible local fallback.

The repository includes the React interface, FastAPI orchestration, synthetic cohort, ontology, IQ adapters, provisioning scripts, tests, and architecture documentation for connected and deterministic local operation.

View the repository Review limitations