Skip to main content

Parthenon v1.0.9 — Protocol-to-Publication, FHIR Interoperability, and Copilots Everywhere

· 6 min read
Sanjay M. Udoshi MD
Creator, Parthenon

v1.0.9 — Protocol-to-Publication, FHIR Interoperability, and Copilots Everywhere​

Where v1.0.8 grew the research surface (Publish, Library Lifecycle, the first copilots), v1.0.9 is the release that makes that surface trustworthy and reproducible. It lands the deterministic protocol-to-publication pipeline (ADR-0020) — provenance, empirical calibration, scientific gates, and manuscript synthesis — brings FHIR ingestion to Medgnosis parity and proves it against the live Epic sandbox, puts the Abby copilot on every page with a local-model backend so Community Edition can run it on-prem, and closes a long list of production-readiness items: RBAC on every PHI route, R/HADES analytics proven end-to-end, no "coming soon" in primary workflows, and a fresh-machine install that starts cleanly from docker compose up.

Protocol-to-Publication pipeline (ADR-0020)​

The single largest feature block. A study now moves through a deterministic finite-state machine from protocol to manuscript, with each transition gated on evidence:

  • Provenance spine — content-addressable hashes across analysis outputs and draft artifacts, version pinning, and reproducible study packages, so any result can be traced to the exact inputs and code that produced it
  • Empirical calibration — calibrated confidence intervals, EASE, and multiplicity correction, with negative-control calibration wired into the CohortMethod sidecar; uncalibrated estimates never reach the manuscript
  • Scientific gate ledger + estimate blinding — blocking gates with estimate blinding, behind the studies.gating_enabled flag
  • Missing-cohort diagnostics — index-event breakdown, inclusion attrition, and orphan-concept diagnostics with user-visible explanations
  • Manuscript composer — a gate-aware, protocol-ordered STROBE/RECORD synthesis with a provenance appendix, exportable to DOCX and XLSX

Driving it is the Abby study orchestrator, a Claude Agent SDK copilot with an orchestrator launch surface and copilot panel that carries study design → execution → publication. (This orchestrator shipped under the working name "Clio" and was unified under the single Abby brand before release — one universal AI brand, zero behavior change.)

FHIR interoperability — Medgnosis parity and a live Epic handshake​

Parthenon's FHIR ingestion now matches the Medgnosis mapper set and is proven on real Epic data:

  • Six new resource types — CarePlan, Goal, CareTeam, DocumentReference, Coverage, and ServiceRequest map into the OMOP CDM (and its extension tables) through a new ResourceMapper interface with registry dispatch
  • Soft-delete semantics — entered-in-error status plus a Bulk deleted manifest, with soft-delete crosswalk columns on the care-extension tables
  • Epic SMART Backend Services auth — a public JWKS endpoint with a kid in the Backend Services JWT, validated end-to-end against the live Epic FHIR sandbox: the auth handshake plus six mappers exercised on real data
  • OMOP-to-FHIR bulk $export with per-user rate limiting, surfaced through an admin bulk-export page that creates, polls, and downloads NDJSON

Copilots everywhere — and a local-model backend for CE​

  • Omnipresent Abby — an action-taking copilot on the Claude Agent SDK, grounded on all 38 page contexts, with help and Abby drawer coverage on every page (previously only 21 contexts were grounded)
  • Provider flexibility — an admin-switchable copilot provider (Anthropic cloud ↔ local Ollama) with graceful fallback, plus a config-driven local-model backend so Community Edition can run the copilots fully on-prem
  • Cloud safety — Abby provider entitlements with cloud-safety gates and HIGHSEC API key masking

Production-readiness hardening​

Most of the release is the unglamorous work that turns a feature-complete beta into something you can run on real PHI:

  • Security — CORS scoped to app origins (was wildcard with credentials); RBAC permission middleware added to Achilles, DQD, risk-score, and clinical-coherence routes; profiles.view enforced on Morpheus and Patient Profile PHI routes
  • R / HADES proven end-to-end — estimation, prediction, SCCS, and evidence-synthesis runs now fail with actionable diagnostics when the R sidecar is unavailable instead of silently completing; the response contract is hardened and HADES is verified end-to-end; a per-contrast diagnostics gate means one bad contrast no longer fails an entire study
  • sidecars:readiness — a single artisan command that probes all nine sidecars (darkstar/R, python-ai, redis, orthanc, hecate, fhir-to-cdm, templates, anonymizer, scispacy) with --json and a non-zero exit, for environment promotion gates
  • No "coming soon" — the FHIR bulk-export and source-profiler comparison surfaces are wired to real backends, and non-functional stubs were removed outright (the dead Abby plan-execution UI, Chroma Studio seed-FAQ, imaging AI measurement extraction) rather than left as placeholders
  • Quality gates — coverage floor ratcheted 25% → 32% (measured 33.59%), real PHPStan level 8 restored via a dated baseline, and new contract tests (R↔PHP calibration, Abby data-interrogation, a manuscript numeric-fabrication guard)

Research surface and Publish​

  • OHDSI-lifecycle study tabs — Study Protocol → Analysis Results → Draft Manuscript, populated from analysis results with a composed manuscript and merged progress tracking
  • Publish exports — XLSX manuscript export, PNG/SVG figure export, OHDSI report-bundle import/export, a rich study-first picker with typed result summaries, and ### subheadings in the paper renderer
  • Multi-reviewer phenotype adjudication, patient-similarity PSM cohort export to the results table, GIS .xlsx/.xls upload, and a wired self-controlled cohort detail route
  • Hypertension v4 designed and run end-to-end on the Acumenus OMOP CDM — BP-indexed target, delay strata, a negative-control panel, and a recording-comparable normotensive comparator

Easier to install​

Fresh-machine install hardening closes the gaps that made a clean clone stumble:

  • docker compose now starts from a clean .env.example — the previously undocumented, required PARTHENON_INTERNAL_TOKEN is present, along with default-service credentials for Reverb, Orthanc, Hecate, and JupyterHub, so the first docker compose up no longer aborts or comes up blank-credentialed
  • the installer creates the acropolis-backend network and the data/wiki bind directory before starting services
  • the manual-install docs now match reality (the node frontend builder lives in the dev profile, the external networks and host directory are created up front)

The recommended path remains the one-command installer:

git clone https://github.com/Acumenus-Data-Sciences/Parthenon.git
cd Parthenon
python3 install.py

Upgrade notes​

  • git pull && ./deploy.sh is sufficient for most environments. New columns (FHIR care-extension tables, soft-delete crosswalks, provenance tables) are added by idempotent migrations — run ./deploy.sh --db to apply them.
  • The protocol-to-publication gates are off by default — enable them with studies.gating_enabled once you want blocking gates and estimate blinding.
  • The AI Agents copilots remain off by default; enable them from the admin AI Agents toggle. To run them locally in Community Edition, set AGENT_PROVIDER=local, pull AGENT_LOCAL_MODEL in Ollama, and start the sidecar (docker compose --profile ce up -d claude-router python-ai).
  • Existing installs already have PARTHENON_INTERNAL_TOKEN set; the install changes only affect fresh clones.

By the numbers​

  • 161 commits since v1.0.8 over 29 days (49 feat, 32 fix)
  • 8 feat(fhir) commits bringing the mapper set to Medgnosis parity
  • 26 commits across the studies / publish / orchestration / estimation / agents feature lines
  • 3 install blockers found and fixed so a clean clone reaches login

Contributors​

Claude Code + @sudoshi