Washington Health Summary (WAHS) · UW CIRG & WA DOH Collaboration
People receiving care across multiple clinics, hospitals, and pharmacies often end up with large, fragmented health exports filled with repeated lab panels, duplicate notes, outdated problem-list stages, and conflicting medication lists. The WA Health Summary workstream explores how clinical rules and open-weight AI can condense these exports into a clear, verifiable summary that patients control and clinicians can trust.
How it works
Multi-year FHIR R4 exports can exceed 18 megabytes and 7,000 entries—far too large for a clinician to scan or a language model to read at once. A progressive 5-tier condenser removes duplicate notes, groups repeated lab readings into longitudinal trends, and fits the record into a predictable size while keeping every allergy, active medication, and active condition intact.
When records arrive from separate clinics or registries, the same medication may appear under both brand and generic names, two drugs from the same class may be listed as active simultaneously, or earlier disease stages may remain marked active alongside newer stages. Deterministic clinical checks highlight these discrepancies so patients and clinicians can review them.
Health records frequently contain sensitive social, behavioral, or personal findings (such as domestic violence screening, justice involvement, or housing instability) stored alongside medical diagnoses. Patients can review their complete personal record in a self-view while automatically redacting sensitive entries when preparing a summary to share with an external provider.
Every summary statement, conflict alert, and Q&A response cites the exact FHIR resource identifier and clinical date so a patient or clinician can inspect the underlying record. When asked about facts absent from the chart, the assistant explicitly abstains rather than guessing.
What we evaluate
Every summary statement, conflict alert, and chat answer links directly to the underlying FHIR record identifier and date. When a question asks about a test, allergy, or procedure that is not in the patient’s chart, the system explicitly states that the information is not documented rather than guessing.
Tested across synthetic patient records representing an 84-year-old adult with complex heart, kidney, and diabetes care, a 74-year-old adult with multi-clinic specialty records, and a 17-year-old adolescent with school and state registry immunizations.
Walk through condensation, conflict reconciliation, privacy controls, and grounded Q&A step by step.
Launch interactive UX (wahs-demo.openclinai.org) Inspect the 32-case evaluation report Open the detailed design & architecture dashboardBackground & Standards Foundation
Our condensation, reconciliation, privacy segmentation, and human-in-the-loop review workflows are grounded in established health interoperability standards, clinical practice guidelines, federal and Washington state privacy frameworks, and human-AI interaction research. See the detailed design & research report for the full technical synthesis.
patient_self) from external provider sharing (external_share).
Critical Review & Phase 3 Roadmap
Our Phase 2 prototype and evaluation validate the deterministic pre-LLM foundation (condensing large FHIR bundles and running clinical rule checks), but important research and engineering questions remain before live clinical or patient-facing deployment. See Section 09 of the detailed design & architecture dashboard for the complete technical analysis, model inventory, and experimental design.
uwcirg/llm-configs).
ips-selective-disclosures (IETF SD-JWT salted sd-hash extensions on FHIR elements/entries) and SMART Health Links JWE encryption.
Explore the interactive patient summary prototype, read the recorded evaluation findings, or inspect the detailed technical design dashboard.
Experimental research and demonstration software using synthetic patient records. AI outputs and automated checks require human review and do not establish clinical safety or effectiveness.