Skip to main content

What's New

This page summarizes the major changes in each release of FHIR4DS.

Version 0.0.17​

September 2026

Version 0.0.17 ships the CQL Cleanroom — a fully browser-native CQL workbench built on the operations layer. Every capability runs client-side: Pyodide hosts the fhir4ds wheel for stateless operations (parse, translate, validate), DuckDB-WASM executes the translated SQL, and the TS executor emits schema:1 envelopes that are field-for-field identical to the CLI and MCP adapters. The conformance baseline holds at 2832/2832 (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47).

Operations layer (shared core)

  • New capabilities: validate_resource (loader-identity validation — a resource that passes validation always loads cleanly) and resource_schema (FHIR R4 StructureDefinition-driven field metadata).
  • translate_cql gains an audit-mode enum (none / population / full), output_columns aliasing, and patient_ids pushdown; results carry column_types CQL metadata.
  • parse_cql(include_ast=True) exposes statement-level AST trees.
  • New compare_evidence capability: strict evidence-payload diff with moved/added/removed/flipped classifications per patient and population.
  • CLI verify --evidence / --baseline writes and compares evidence artifacts; MCP gains compare_evidence_tool.

CQL Cleanroom (web/cql-cleanroom)

  • Monaco CQL editor with live diagnostics (markers from structured Diagnostics.location), parameter panel, and multi-library tabs with IndexedDB workspace persistence + zip export/import.
  • Evaluation, test, and evidence panes over DuckDB-WASM: typed result tables, SQL viewer, test harness with failure reasons, patient drill-in with audit evidence, and population Sankey flows.
  • Evidence compare mode (baseline paste vs. current explain) and share links (LZ-compressed URL fragments, ≤100 KB, never contain datasets).
  • Resource Builder: schema-driven FHIR resource forms (8 types + raw JSON fallback) with validate-before-add gating and no-silent-data-loss edit round-trips.
  • Visual algorithm editor v1: graph → CQL emitter with a parse round-trip guard before applying to the library.

Parity & verification

  • Three-way capability matrix: CLI == MCP == browser envelopes on the shared fixture, for both run_tests and compare_evidence.
  • Macro-sync contract test keeps the browser SQL macro surface in lockstep with the Python translator (test_macro_sync.py).

Known limitations

  • 21 deep-comparator SQL macros remain browser-gap allowlisted (BROWSER_MACRO_GAP_ALLOWLIST); the visual editor canvas is a beta seam over the JSON graph editor.

Version 0.0.16​

September 2026

Version 0.0.16 is the CQL type system release. A complete, authoritative CQL type-inference layer (TypeMapBuilder) now runs after translation and attaches typed metadata (cql_type_ref) to every library definition, and the $cql facade serializer flips to a metadata-first contract: declared types drive FHIR Parameters serialization, with runtime shape inference only as a silent fallback for Any or a logged rescue when known metadata would fail. The conformance baseline holds at 2832/2832 (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47) and generated SQL is byte-identical to 0.0.15 across the 47-measure DQM corpus — this is a metadata-only release.

Type system

  • New fhir4ds/cql/types/typeref.py canonical CQLTypeRef representation (List/Interval/Tuple generics) shared by the translator and the FHIR facade; legacy stub modules removed.
  • TypeMapBuilder (fhir4ds/cql/translator/type_map.py) types 70.3% of the 1354 definitions in the DQM-47 corpus (952 non-Any); the remaining Any results are the documented gap backlog (Property/Query-heavy defines).
  • Uncertainty-returning operators (AgeIn*, duration/difference between on imprecise operands) are deliberately typed Any so runtime interval results serialize correctly.

Facade ($cql)

  • CQLResultMetadata fallback chain is now cql_type_ref → cql_type → sql_result_type → Any.
  • Serializer trusts known metadata; rescue is failure-driven, recursive per-level, and logs a WARNING naming both types when it fires. Long values, empty-list/tuple extensions, and interval part shapes are unchanged.
  • Rollback flag: CQLServerConfig.metadata_first_serialization=False restores the 0.0.15 value-first reconcile verbatim.

Consolidation

  • The legacy _infer_cql_type is now a thin wrapper over a verbatim compatibility port (byte-exactness proven by a 1049-define snapshot diff); two duplicated helpers were deleted and their callers now share type_map module functions; a CI drift tripwire pins builder-vs-legacy equivalence on a shared corpus.

Version 0.0.15​

September 2026

Version 0.0.15 is the DuckDB 1.5.5 reliability release. The embedded engine is upgraded from DuckDB 1.5.2 to 1.5.5 (native and WebAssembly binaries rebuilt and redeployed in lockstep), and a full 36-iteration QA campaign across all twelve product domains hardened the FHIRPath engine, the CQL translator, ingestion, and the terminology surface. The conformance baseline holds at 2832/2832 (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47).

Platform

  • DuckDB upgraded 1.5.2 → 1.5.5 (wheel pin, native extensions, duckdb-wasm 1.33.1-dev64.0, submodule trees). Cloud credential options realigned to the DuckDB 1.5.x secret surface: S3/Azure/GCS snake_case aliases map to native option names, Azure service principals emit PROVIDER service_principal, and options removed upstream (account_key, service_account_json) now raise actionable errors.

FHIRPath engine (24 fixes, QA-001..QA-008)

  • Native/fallback parity: Decimal-typed 0/1 booleans, empty-expression row resilience, metadata-gated temporal + concatenation, JSON cross-kind equality (object/array vs scalar → empty), minute-precision DateTime ± minutes arithmetic, temporal typing of arithmetic results ((@2016-02-29 - 1 month) > 'abc' is now empty), §4.4.1 empty propagation precedence in ('a'|'b') / {}, and singleton-violation ordering for in/contains.

CQL translator (QA-009..QA-021)

  • Dynamic FHIR interval union, bare expand X.period per day parsing, stored-list expand/collapse defines (Count/First consumers), query aggregates over retrieves (patient-correlated folds), braced { <query> } list selectors in every consumer, chained with clauses (CMS71-style), { <query> } as query source, audit-mode contains over multi-valued properties, and audit Count(...) >= N evidence fan-out.
  • evaluate_measure output-columns and retrieve type validation now fail typed and actionable (ValueError / TranslationError) instead of leaking DuckDB CatalogExceptions.

Terminology (medterm4ds)

  • AutoCoderConfig.top_k validated at construction; autocoding extension roundtrip and closure-table surfaces verified end-to-end.

Version 0.0.14​

September 2026

Version 0.0.14 is the WASM C++ parity release. Ten translator-reachable CQL functions that previously existed only as Python UDFs now ship in the native C++ DuckDB extension, closing the last gaps between the desktop runtime and the no-Python browser (WebAssembly) runtime. The conformance baseline holds at 2832/2832 (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47).

WASM C++ parity

  • Ported to the C++ extension: cqlDivide (exact scale-8 half-up decimal division), CQLMessage, coding_matches/coding_matches_exact, fhirpath_in_valueset, cqlChildren/cqlDescendants (with CQL-05 transport markers), cqlDateTimeAdd, ratioCompare, and ConceptToListCode. Every port is differential-tested against the Python authority (3,000+ randomized cqlDivide cases, 1,200 coding-match cases, 800 structural-traversal cases — zero divergences on the conformant input classes).
  • Three dead Python registrations removed (ConvertQuantity, cqlDateTimeSubtract, cqlNormalizeTZ) and the between-family casing reconciled with the translator's emit spellings.
  • cqlDateTimeAdd and ratioCompare remain Python-authoritative on desktop (documented divergent input classes); the C++ scalars serve the no-Python/WASM runtimes.
  • Both DuckDB extensions rebuilt for native and WASM; the browser demos run identical binaries. The CQL playground gained an exact-decimal-division end-to-end case (10.0 / 3.0 → 3.33 display), verified against the rebuilt WebAssembly — including the previously failing decimal division case.

Fixes

  • resolve() macro bind failure (latent since May): registering CQL macros shadowed DuckDB's built-in trim(x, chars), breaking any translated CQL using resolve() in the standard registration flow. The macro body now uses regexp_replace, with regression coverage in both load orders.
  • Browser runtime: Arrow DECIMAL columns now render with their fractional part (unscaled-BigInt rescale), and the four type-disambiguation macros required by quantity comparisons are registered in both web apps.

Integration

  • medterm4ds dependency bumped to 0.0.4 (>=0.0.4,<0.0.5) — CDC CPT↔CVX crosswalk, annotation fields on all surfaces, and whitespace-code 400s. Verified against the 0.0.4 candidate (runtime frozen at the release commit).

Version 0.0.13​

August 2026

Version 0.0.13 is a reliability and integration release. It ships the post-0.0.12 reliability fixes, upgrades the medterm4ds integration to 0.0.3, and lands a broad QA hardening pass over the ingestion, error-handling, audit, installation, and API surfaces. The conformance baseline holds at 2832/2832 (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47), now verified on the full CI invocation across Python 3.10 and 3.11 with fresh orjson.

Reliability fixes

  • orjson 3.12 compatibility: deep-nesting decode recovery now matches both orjson depth-limit wordings (≤3.11 "recursion depth", 3.12+ "depth limit") instead of pinning a message — the published 0.0.12 wheel broke on orjson 3.12; 0.0.13 is the patch.
  • DQM evaluator: population truthiness is numpy-safe (sequence semantics instead of bool() on array-likes — warns on old numpy, raises on new).
  • DQM CLI: runtime parameters are filtered to library-declared names (typos can no longer fail measures that never declared them).
  • Recursion-limit import leak: the raised limit is scoped to script execution instead of leaking into every importer's process.

medterm4ds 0.0.3

  • Bounded dependency bumped on both extras: medterm4ds>=0.0.3,<0.0.4 (terminology) and medterm4ds[extraction]>=0.0.3,<0.0.4 (ner).
  • The full integration surface (adapters, factory, ValueSet expansion with the active-only/retired policy split, code search, annotation, the NER notes pipeline) was diff-audited and re-verified against 0.0.3. Note: medterm4ds 0.0.3 defaults its model revision to v0.0.2 (MEDTERM4DS_HF_REVISION) and adds device auto-detection (MEDTERM4DS_DEVICE).

Hardening highlights

  • Native engine restorations: exact digit-string div/mod (never binary64 mediators — values beyond 2^53 are exact), 2^63 literal contracts, named-group regexes ((?<name>…) + ${name} substitution), leading inline regex flags, and PCRE $ trailing-newline semantics.
  • FHIRPath validity contract: §6.7 temporal unit/precision violations now classify as invalid in fhirpath_is_valid on both engines.
  • Loader error contracts: duplicate JSON members rejected per FHIR json.html, lone-surrogate and non-JSON-native value guards, wrapped closed-connection errors, BOM/URL parity, and a shared patient-reference doctrine across the loader and FileSystemSource (including versioned absolute references and list-valued references).
  • FileSystemSource directory paths work everywhere (scans and incremental delta scans), and bundle files are documented as inert rows.
  • API surface: README examples verified verbatim (including the evaluate_measure DataFrame contract) and the full documented entry point set exercised on both engines.

Known limitation: five pre-release parity tests from a 0.0.12 release-commit source-loss incident remain carried (structural-type trio, ratio tuple comparison, WASM no-Python gate); they sit outside the 2832 conformance gate and the CI suite selection and are tracked for upstream tree recovery.

Version 0.0.12​

August 2026

Version 0.0.12 ships the completed SQL-on-FHIR v2 spec-compliance campaign (58 chunks across ViewDefinition, the Analytics Layer, CQL, and FHIRPath), hardens the CQL translator (exact Decimal arithmetic, age-uncertainty semantics, list/aggregate correctness, and query-level Min/Max audit attribution), closes FHIRPath dual-engine parity work, and upgrades the medterm4ds integration to 0.0.2 (bounded dependency >=0.0.2,<0.0.3, retired-code policy split, in-process intensional ValueSet expansion). Conformance rises to 2832/2832 across all four suites (ViewDefinition 144, FHIRPath 935, CQL 1706, DQM 47).

New: SQLQuery / SQLView (the SQL-on-FHIR v2 Analytics Layer)​

  • FHIR Library resources conforming to the SQLQuery and SQLView profiles: typed, FHIR-declared parameters, versioned SQL content, and canonical-URL dependencies that materialize as virtual tables.
  • fhir4ds.sqlquery ships parse_library / parse_sqlquery / parse_sqlview plus SQLQueryRunner, with typed errors for missing dependencies, dependency cycles, and parameter type mismatches.
  • See the sqlquery API reference for the profile model and a runnable example.

Data ingestion correctness​

  • patient_ref attribution now follows a single Patient-typed doctrine across the loader and file sources: only Patient-typed references (and bundle-local urn:uuid:) populate the patient identity — Group or Location subject references and bare ids no longer fabricate phantom patients in measure results.
  • Version-specific references (Patient/123/_history/2) resolve to the current resource instead of silently dropping.
  • UTF-8 BOM-prefixed JSON/NDJSON files (common from Windows tooling) load cleanly; case-variant extensions (.JSON) and unreadable files no longer break or silently skip during directory loads.
  • Bundle entries are validated per-entry with index attribution; empty entry.resource objects are rejected with a clear error.
  • Batch ingestion is ~70-150x faster (Arrow bulk path with automatic fallback); a 25,000-resource batch loads in ~0.3s.

Audit evidence and measure evaluation​

  • Min/Max aggregates over quantity queries now attribute the winning resource in audit evidence (Observation/xyz) with value and threshold — previously the evidence cited the aggregate value itself.
  • Quantity-typed aggregate references compare correctly (MinBP < 140 'mmHg') instead of evaluating to false.
  • audit_mode="full" output is one row per patient with the union of per-resource evidence (previously rows multiplied per evidence resource).
  • evaluate_measure validates output_columns definition names and parameter names against the library — typos raise actionable errors instead of producing silently wrong results.

medterm4ds 0.0.2​

  • Dependency bound medterm4ds>=0.0.2,<0.0.3 on both the terminology and ner extras; expansion behavior verified against the installed wheel.
  • Membership expansion includes retired codes for historical data while discovery/search stays active-only.
  • In-process intensional ValueSet expansion (?fhir_vs=isa/…) with graded code search (lexical/hybrid/semantic).

Version 0.0.11​

August 2026

Version 0.0.11 introduces the medterm4ds integration — a three-phase pipeline that adds terminology services, auto-coding, and clinical NER to fhir4ds. The release also ships a spec-compliance sweep across the translator and the data-ingestion surfaces. Conformance holds at 2822/2822 across all four suites.

medterm4ds integration​

The integration lands in three opt-in phases, each gated behind an install extra so the core import fhir4ds path stays zero-dependency.

  • Phase 1 — Terminology Service (pip install 'fhir4ds-v2[terminology]'): an env-driven factory (fhir4ds.cql.terminology.factory.get_terminology_endpoint) builds a TerminologyEndpoint from FHIR4DS_TERMINOLOGY_* env vars or an explicit TerminologyConfig. Two modes: http (any FHIR R4 terminology server, including the medterm4ds sidecar) and in_process (medterm4ds embedded directly — no HTTP hop). The endpoint exposes expand(), search_text(), search_batch(), plus a per-instance circuit breaker for transient failures and a system-URL normalizer that expands UMLS mnemonics (SNOMEDCT_US) to FHIR canonical URLs (http://snomed.info/sct). See the Terminology Service docs.

  • Phase 2 — Auto-Coding (same extra): fhir4ds.cql.loader.auto_coder.AutoCoder augments text-only CodeableConcepts (e.g. Condition.code.text) with coded Codings drawn from the terminology endpoint's top-k matches. The original text is preserved, existing coding[] is never overwritten, every derived Coding carries userSelected=False and a structured autocoding extension (engine="medterm4ds", index version pinned). Cache keys are deterministic — re-running on the same input produces byte-identical output. See the Auto-Coding docs.

  • Phase 4 — Clinical Notes Pipeline (pip install 'fhir4ds-v2[ner]'): fhir4ds.cql.loader.notes_pipeline.NotesPipeline.extract_conditions() derives FHIR Condition resources from free-text note paths on ClinicalImpression, DocumentReference, Encounter, and friends via medterm4ds.extract() (medspaCy + GLiNER + SapBERT cascade). Each derived Condition carries a derived-from-text extension with source reference, note path, and span offsets for full audit. Deterministic IDs (sha256 of (source_ref, span_start, span_end, system, code)) make re-runs byte-identical. The heavy ML deps (medspaCy, transformers, torch) are owned transitively by medterm4ds[extraction] — fhir4ds never imports them directly. See the Notes Pipeline docs.

medterm4ds is now declared as a normal PyPI dependency in the [terminology] and [ner] extras — no sibling-repo install, no --find-links pinning. The published wheel (medterm4ds 0.0.1 on PyPI) exposes the canonical surface this release targets.

Additional hardening​

A spec-compliance sweep across the translator and data-ingestion surfaces closed 19 release-gate findings (QA-001 through QA-019) without altering the public API: fluent-form CQL .distinct() now matches the function form's NULL-preserving semantics; SQL-on-FHIR v2 forEachOrNull preserves parent rows through nested iteration; the FHIR data loader and notes pipeline gained strict input validation (Bundle.type attribution, NDJSON strict-line attribution, malformed-URL/JSON guards); closed-connection / malformed-URL / malformed-JSON errors now raise typed package exceptions instead of leaking decoder/attribute/connection errors; audit_mode='full' SQL emission now runs on 4 of 5 sampled CMS measures (CMS135, CMS165, CMS71, CMS996) where it previously broke. Public __version__ is aligned across all 7 subpackage __init__.py files.

Conformance​

SuitePassedTotalRate
ViewDefinition v2134134100.0%
FHIRPath R4935935100.0%
CQL1,7061,706100.0%
DQM QI-Core 20254747100.0%
Overall2,8222,822100.0%

Upgrade​

pip install fhir4ds-v2==0.0.11

For the terminology tier: pip install 'fhir4ds-v2[terminology]'. For clinical-notes NER: pip install 'fhir4ds-v2[ner]' (pulls medterm4ds[extraction] transitively).

Version 0.0.10​

June 2026

Version 0.0.10 is a release-preparation cycle focused on translator correctness for production CDS workflows and full release-readiness validation across package metadata, public docs, conformance baselines, benchmarks, and the release artifact surface.

Highlights​

  • Cross-Library Define References: A."PatientAge" in a CQL library that uses include to reference another library now translates correctly in every usage context. The translator emits a correlated scalar subquery with LIMIT 1 for SCALAR contexts (closing the binder error that previously affected case/if conditions), EXISTS for BOOLEAN contexts, and identity passthrough for LIST/query sources. Covered by test_cross_library_reference.py (5 cases) and test_definition_ref_parity.py (2 cases). The usage-aware translation strategy avoids the CTE-creation interception that caused an earlier draft of this fix to be reverted during release-prep.
  • Translator Correctness: Fixes a binder error when a CQL case ... when expression references a previously-defined boolean define; the translator now emits EXISTS against the boolean CTE instead of selecting a non-existent value column. if ... then ... else conditions benefit from the same fix.
  • Definition-Reference Consolidation: The two parallel code paths that resolved promoted vs non-promoted definition references are now unified through _classify_definition_ref, closing a class of parity bugs and making the EXISTS-vs-correlated-subquery decision a single source of truth.
  • MedicationStatement Retrieve Fix: [MedicationStatement: "X"] now resolves against medicationCodeableConcept (and the medicationReference choice-type alternative) instead of silently falling back to the non-existent code field. A full audit added 16 more FHIR R4 resource types whose primary code path is not code.
  • FHIRPath is() Empty-Collection Parity: is() now propagates the empty collection per FHIRPath §5.1 on both the native C++ engine (fn_isType) and the Python fallback (invocations/types.py), resolving the long-standing TestIsLowercaseTypes::test_empty_collection_returns_empty case. Both paths now return the empty collection instead of [false] for non-FHIR-primitive type arguments.
  • FHIRPath Engine Hardening: Choice-type as() and is() parity for value[x] on Parameters resources; bundled native extension refreshed (11.1 MB → 11.5 MB).
  • Release-Surface Alignment: Package metadata, public subpackage versions, landing-page version, WASM wheel reference, install snippets, and release notes are aligned with 0.0.10.
  • Validation Gates: The release-prep pipeline includes code review, release validation, documentation audit, benchmark validation, and final scribe handoff gates before completion. All four conformance suites (ViewDefinition, FHIRPath, CQL, DQM) remain at 100%.

Conformance​

SuitePassedTotalRate
ViewDefinition v2134134100.0%
FHIRPath R4935935100.0%
CQL1,7061,706100.0%
DQM QI-Core 20254747100.0%
Overall2,8222,822100.0%

Upgrade​

pip install fhir4ds-v2==0.0.10

Version 0.0.9​

June 2026

Version 0.0.9 is a release-preparation hardening cycle focused on preserving the full conformance baseline while revalidating infrastructure domains, release metadata, public documentation, browser assets, and benchmark evidence.

Highlights​

  • Release Readiness Campaign: Revalidates FHIRPath, CQL, SQL-on-FHIR ViewDefinition, ingestion/source adapters, DQM audit evidence, installation, and environment behavior across the scheduled 0.0.9 evolution loop.
  • Version and Artifact Alignment: Package metadata, public subpackage versions, notebook snippets, website install references, and WASM translator wheel references are aligned with 0.0.9.
  • Validation Gates: The release-prep pipeline includes conformance, package pytest, documentation audit, benchmark validation, and final scribe handoff gates before completion.

Upgrade​

pip install fhir4ds-v2==0.0.9

Version 0.0.8​

June 2026

Version 0.0.8 focuses on Mongo-backed FHIR server integration, FHIR $cql conformance-runner compatibility, release hardening, public API robustness, audit evidence fidelity, and refreshed browser/package artifacts while preserving the full conformance baseline.

Highlights​

  • Mongo FHIR Server Integration: Added MongoFhirServerSource for read-only analytics over Mongo-backed FHIR servers through DuckDB's community mongo_scan extension, with support for per-resource, explicit, and shared collection layouts.
  • Mongo DQM Materialization: Added fhir4ds dqm mongo ... commands and worker support for durable patient-change queues, change-stream processing, result/audit storage, optional generated MeasureReport publishing, and patient-scoped source pushdown.
  • FHIR $cql Facade: Added a narrow local FHIR R4 $cql operation facade and fhir4ds cql-server CLI for running cqframework/cql-tests-runner against the FHIR4DS CQL engine.
  • CQL Spec Parity Sweep: Tightened CQL primitive, clinical, temporal, interval, list, quantity/ratio, conversion, rounding, and no-Python/native DuckDB parity behavior across the conformance surface.
  • FHIRPath Native Parity: Native DuckDB FHIRPath now rejects malformed Section 5.1 existence-helper arities, including specialized exists() and FHIR-specific hasValue() dispatch paths, in parity with the Python fallback.
  • Measurement Period Fidelity: CQL-authored Interval<DateTime> parameter defaults retain DateTime precision and authored open/closed boundary flags through DQM population SQL generation.
  • Public API Error Contracts: DQM config loaders, HAPI/Mongo materialization config parsing, and filesystem sources now raise actionable package errors for malformed public inputs.
  • Audit Evidence Accuracy: Multi-group DQM audit pruning preserves group-local causal resource targets, and Mongo compact materialization now mirrors the HAPI compact result/full audit split.
  • Release Artifact Consistency: Package metadata, public subpackage versions, notebook snippets, CQL runner metadata, wheel contents, and WASM translator assets are aligned with 0.0.8.

Bug Fixes​

  • CQL $cql Runner Compatibility: FHIR Parameters serialization now handles intervals, empty/null values, temporal values, quantities, ratios, Codes, Concepts, Lists, and Tuples in runner-compatible shapes, and runner report status now reflects JSON-level failures even when the Node process exits successfully.
  • CQL Semantic Edge Cases: Fixed ratio literals, Long metadata preservation, DateTime null precision, decimal equivalence, interval null/open-bound behavior, interval uncertainty propagation, ToString formatting, and public Round/RoundTo half-away-from-zero behavior across Python fallback, native-loaded DuckDB, and no-Python/browser-style surfaces.
  • Measurement Period Fidelity: CQL-authored Interval<DateTime> parameter defaults retain DateTime precision and authored open/closed boundary flags through DQM population SQL generation.
  • Public API Error Contracts: DQM config loaders, HAPI/Mongo materialization config parsing, and filesystem sources now raise actionable package errors for malformed public inputs instead of leaking decoder or attribute errors.
  • Audit Evidence Accuracy: Multi-group DQM audit pruning preserves group-local causal resource targets, and Mongo compact materialization now mirrors the HAPI compact result/full audit split.
  • Release Artifact Consistency: Package metadata, public subpackage versions, notebook snippets, CQL runner metadata, wheel contents, website release surfaces, and browser demo assets are aligned with 0.0.8.

Conformance​

SuitePassedTotalRate
ViewDefinition v2134134100.0%
FHIRPath R4935935100.0%
CQL1,7061,706100.0%
DQM QI-Core 20254747100.0%
Overall2,8222,822100.0%

Upgrade​

pip install fhir4ds-v2==0.0.8

Version 0.0.7​

May 2026

Version 0.0.7 focuses on release readiness, SQL safety, audit evidence accuracy, and packaging consistency. It preserves the 100% conformance posture while tightening loader boundaries, DQM population attribution evidence, and native/Python fallback release gates.

Highlights​

  • Compiled Measures API: Introduced compile_measure() and execute_compiled_measure() in the DQM API. This allows parsing and compiling CQL to SQL once, enabling highly efficient batched execution across patient cohorts.
  • Artifact Resolvers: Replaced static file-path dependencies with a pluggable ArtifactResolver interface (FileArtifactResolver and HapiArtifactResolver), enabling dynamic resolution of Measures, CQL libraries, and ValueSets from file systems or remote FHIR servers.
  • HAPI FHIR Integration: Added HapiPostgresSource for querying HAPI FHIR JPA Server PostgreSQL backends in place, along with comprehensive guides for event-driven DQM materialization.

Bug Fixes​

  • FHIRPath Native Parity: Massively tightened the behavior of the C++ DuckDB extension to exactly match the Python fallback across arithmetic overflows, collection equality, singleton evaluation, string reflection, and edge-case Boolean precedence.
  • CQL Type Inference: Fixed row shapes and nested tuples generated during compiled measure materialization so Query and Return clauses preserve complex type structure correctly across batch runs.
  • CQL Optimization: Ensure degenerate intervals (e.g. [x, x]) safely propagate through StartsSame and EndsSame temporal operators without triggering DuckDB null-handling faults.
  • CQL Temporal Intervals: Optimized interval overlap SQL handles null low bounds using CQL interval semantics instead of propagating SQL NULL.
  • DQM Audit Evidence: Exclusion-style audit narratives use effective population masks so denominator exceptions and numerator exclusions match final population attribution rules.
  • Loader SQL Safety: Custom resource and ValueSet table names are quoted consistently, including SQL keywords such as select and where.
  • FHIR JSON Validation: File, URL, directory, and DQM ValueSet loaders now reject non-object JSON cleanly instead of leaking implementation errors.

Conformance​

SuitePassedTotalRate
ViewDefinition v2134134100.0%
FHIRPath R4935935100.0%
CQL1,7061,706100.0%
DQM QI-Core 20254747100.0%
Overall2,8222,822100.0%

Upgrade​

pip install fhir4ds-v2==0.0.7

Version 0.0.6​

May 2026

Version 0.0.6 focuses on release hardening, browser runtime parity, and clinical correctness for HEDIS/DQM workflows. The release preserves 100% conformance across ViewDefinition, FHIRPath, CQL, and DQM while tightening the packaged DuckDB extension boundary.

Highlights​

  • C++ DuckDB WASM Parity: Browser-required CQL interval, boundary, date/quantity, interval set, logical, list, valueset, and FHIRPath repeat functions are available through C++ extensions.
  • No-Python Runtime Gate: Added direct-extension tests that load packaged C++ extensions without registering Python fallback UDFs.
  • 100% Conformance Gate: ViewDefinition (134/134), FHIRPath R4 (935/935), CQL (1,706/1,706), and DQM QI-Core 2025 (47/47) all pass.
  • HEDIS Clinical Logic: Calendar age helpers, episode aggregate folds, repeated-extension predicates, interval boundaries, stratifiers, and reference resolution were hardened for native DuckDB, Python fallback, and browser-style execution.
  • DQM Reporting: Measure group stratifiers are parsed, evaluated, summarized, and exported in MeasureReport output.
  • Packaging Safety: The wheel pins DuckDB to 1.5.2 so installed Python dependencies match the bundled native extension ABI.
  • WASM Assets: Updated translator wheel to fhir4ds_v2-0.0.6-py3-none-any.whl and rebuilt DuckDB extension side modules.

Bug Fixes​

  • CQL: Patient-context AgeIn*At(asOf) now routes through CalculateAgeIn*At(birthDate, asOf) helpers for leap-day and calendar-period parity.
  • CQL: Query aggregate list folds preserve typed empty-list semantics so HEDIS episode deduplication counts episode collections correctly.
  • CQL: resolve() accepts ResourceType/id, bare ids, full URLs ending in ResourceType/id, and JSON Reference objects.
  • DQM: Proportion summaries now assign patient/case population labels before aggregation so exclusions and exceptions are counted in the correct order.
  • Sources: Strict NDJSON loading is all-or-nothing for both malformed JSON and valid-JSON invalid-FHIR records.
  • CSV Source: Constructor inputs are validated at the public API boundary before SQL projection registration.

Upgrade​

pip install fhir4ds-v2==0.0.6

Version 0.0.5​

May 2026

Version 0.0.5 continues the 100% Compliance Milestone — every conformance suite passes at 100%, totaling 2,822 tests across CQL, FHIRPath, ViewDefinition, and DQM.

Highlights​

  • 100% Spec Compliance: All test suites now pass at 100% — CQL (1,706/1,706), FHIRPath (935/935), ViewDefinition (134/134), and DQM (47/47 measures).
  • CQL Gap Closure: Resolved remaining 2 CQL spec tests — RolledOutIntervals (DuckDB correlated UNNEST workaround) and IntegerIntervalProperlyIncludedInNullBoundaries (spec ambiguity resolved).
  • DQM Full Pass Rate: All 47 QI-Core 2025 CMS eCQMs now pass. 4 measures have documented upstream test data accuracy gaps (CMS135, CMS145, CMS157, CMS1017) that affect all conformant engines equally.
  • ReactiveEvaluator API Update: ReactiveEvaluator constructor now accepts measure_bundle and cql_library_path parameters, aligning with the MeasureEvaluator API.
  • Interval JSON Handling: Fixed precision comparisons to correctly handle interval JSON in CQL temporal operations.

Bug Fixes​

  • CQL: Fixed interval JSON parsing in precision comparisons (affected CMS157 and related temporal tests).
  • CQL: Prevented parser loop and enforced inline recursion limit for deeply nested expressions.
  • CQL: Fixed RolledOutIntervals — implemented alternative approach avoiding DuckDB correlated UNNEST limitation.
  • DQM: Fixed CMS157 test data — corrected measurement period alignment with encounter dates.

API Changes​

  • ReactiveEvaluator.__init__: Parameters changed from (con, measure, adapter) to (con, measure_bundle, cql_library_path, adapter).
  • Version Bump: All fhir4ds packages bumped to version 0.0.5.
  • WASM Assets: Updated translator wheel for the 0.0.5 release series.

Upgrade​

pip install fhir4ds-v2==0.0.5

Version 0.0.4​

May 2026

Version 0.0.4 marks the 100% Compliance Milestone — every test suite now passes at 100%, totaling 2,821 tests across CQL, FHIRPath, ViewDefinition, and DQM.

Highlights​

  • 100% Spec Compliance: All test suites now pass at 100% — CQL (1,706/1,706), FHIRPath (935/935), ViewDefinition (134/134), and DQM (46/46 measures).
  • CQL Gap Closure: Resolved remaining 2 CQL spec tests — RolledOutIntervals (DuckDB correlated UNNEST workaround) and IntegerIntervalProperlyIncludedInNullBoundaries (spec ambiguity resolved).
  • DQM Full Pass Rate: All 46 QI-Core 2025 CMS eCQMs now pass. 4 measures have documented upstream test data accuracy gaps (CMS135, CMS145, CMS157, CMS1017) that affect all conformant engines equally.
  • ReactiveEvaluator API Update: ReactiveEvaluator constructor now accepts measure_bundle and cql_library_path parameters, aligning with the MeasureEvaluator API.
  • Interval JSON Handling: Fixed precision comparisons to correctly handle interval JSON in CQL temporal operations.

Bug Fixes​

  • CQL: Fixed interval JSON parsing in precision comparisons (affected CMS157 and related temporal tests).
  • CQL: Prevented parser loop and enforced inline recursion limit for deeply nested expressions.
  • CQL: Fixed RolledOutIntervals — implemented alternative approach avoiding DuckDB correlated UNNEST limitation.
  • DQM: Fixed CMS157 test data — corrected measurement period alignment with encounter dates.

API Changes​

  • ReactiveEvaluator.__init__: Parameters changed from (con, measure, adapter) to (con, measure_bundle, cql_library_path, adapter).
  • Version Bump: All fhir4ds packages bumped to version 0.0.4.
  • WASM Assets: Updated translator wheel to fhir4ds_v2-0.0.4-py3-none-any.whl.

Upgrade​

pip install fhir4ds-v2==0.0.4

Version 0.0.3​

April 2026

Version 0.0.3 introduces the Zero-ETL Source Adapter architecture, enabling CQL measures, FHIRPath queries, and ViewDefinitions to run directly against external data without copying it into DuckDB.

Highlights​

  • Zero-ETL Source Adapters: Run clinical logic directly against Parquet data lakes, PostgreSQL databases, and CSV files — no data movement, no PHI duplication.
  • New API: fhir4ds.attach(), fhir4ds.detach(), and create_connection(source=...) provide a uniform lifecycle for all data sources.
  • Schema Validation at Registration: SchemaValidationError is raised immediately at attach() time if the adapter's view doesn't conform — fail fast, not during measure evaluation.
  • Backward Compatibility: ExistingTableSource provides the adapter interface over pre-loaded data. No breaking changes for FHIRDataLoader users.

New: Source Adapters​

AdapterPurpose
FileSystemSourceParquet, NDJSON, Iceberg files — local or cloud (S3, Azure, GCS)
PostgresSourceFHIR JSON stored in PostgreSQL columns
ExistingTableSourceWraps pre-loaded DuckDB tables in the adapter API
CSVSourceFlat CSV files with user-defined SQL projection

FileSystemSource​

  • Supports Parquet (default), NDJSON, JSON, and Iceberg formats.
  • Cloud storage via CloudCredentials (S3, Azure, GCS).
  • Hive partition pruning for large datasets.
  • Incremental delta tracking via file mtime.

PostgresSource​

  • Attaches to Postgres via DuckDB's postgres extension (read-only).
  • PostgresTableMapping defines column-to-schema mappings per table.
  • All identifiers quoted via quote_identifier() — prevents SQL injection from user-supplied names.
  • Scope boundary: Requires FHIR JSON in a column. Relational-to-FHIR column mapping is out of scope.

ExistingTableSource​

  • Wraps any existing DuckDB table/view in the adapter API.
  • Validates schema at registration time.
  • Zero migration cost for current FHIRDataLoader users.

CSVSource​

  • User provides a projection_sql with a {source} placeholder.
  • Full control over how flat CSV columns map to FHIR JSON (via json_object()).

New API​

  • fhir4ds.attach(con, adapter) — Registers a source adapter on an existing connection.
  • fhir4ds.detach(con, adapter) — Unregisters an adapter, dropping the view and cleaning up.
  • fhir4ds.create_connection(source=adapter) — Mounts a source immediately on connection creation.
  • SourceAdapter Protocol — Interface contract for third-party adapter implementations (register(), unregister()).
  • SchemaValidationError — Raised at registration time if the adapter schema doesn't conform.
  • validate_schema() — Validates the resources view against the required schema contract.
  • quote_identifier() — Safely quotes identifiers to prevent SQL injection.
  • CloudCredentials — Encapsulates DuckDB secret configuration for S3, Azure, and GCS.

Security​

  • Identifier Quoting: PostgresSource and ExistingTableSource use quote_identifier() to prevent SQL injection from user-supplied table and column names.
  • Scope Boundary: PostgresSource documentation clearly states it requires FHIR JSON in a column — preventing misuse as a generic relational mapper.

Known Limitations​

  • PostgresSource requires FHIR JSON in a Postgres column — constructing FHIR JSON from arbitrary relational schemas is not supported in this release.
  • CSVSource does not support incremental delta tracking.
  • FileSystemSource incremental tracking is mtime-based and may produce false positives.
  • Iceberg format does not support incremental delta tracking.

Migration Guide​

Existing FHIRDataLoader users have no breaking changes. To adopt the adapter pattern:

# Before (still works, no changes needed):
loader = FHIRDataLoader(con)
loader.load_directory('/data/fhir/')

# After (optional — adds uniform adapter API):
from fhir4ds.sources import ExistingTableSource
source = ExistingTableSource()
fhir4ds.attach(con, source)

Upgrade​

pip install fhir4ds-v2==0.0.3

API Changes​

  • Version Bump: All fhir4ds packages bumped to version 0.0.3.
  • WASM Assets: Updated translator wheel to fhir4ds_v2-0.0.3-py3-none-any.whl.
  • New Module: fhir4ds.sources — exported from the top-level fhir4ds namespace.

Version 0.0.2​

April 2026

Version 0.0.2 focuses on reaching full spec compliance, enhancing performance through architectural optimizations and a new C++ extension, and hardening security.

Highlights​

  • Near-Full Spec Conformance: Achieved 99.8% conformance across CQL, DQM, FHIRPath, and ViewDefinition test suites, resolving over 150 identified gaps.
  • 69.5x Performance Boost: Evaluation speed has increased significantly due to a hybrid C++/Python execution model and optimized metadata caching.
  • DuckDB v1.5.2 Integration: Fully stabilized on the latest DuckDB version, including optimized WASM builds for browser-side execution.
  • Standardized Conformance Logging: A new unified logging framework provides detailed pass/fail reporting across all engine components.

Performance Improvements​

  • Registry Caching: MeasureEvaluator now caches FHIR schemas and profile registries, eliminating redundant 1.5MB allocations per evaluation call.
  • Audit Deduplication: Optimized "Full" audit mode to automatically deduplicate Cartesian product results generated by complex LEFT JOINs in retrieve CTEs.
  • C++ Extension Parity: Reached 100% feature parity between the Python UDFs and the high-performance C++ extension, allowing for hybrid execution that combines C++ speed with Python's flexibility.

Security Fixes​

  • JSON Injection Remediation: Fixed critical JSON injection vulnerabilities in the C++ evaluator's type() and width_string() functions.
  • Thread Safety: Added synchronization locks to singleton registries and cache stores (profile_registry, fhir_loader, variable_store) to prevent race conditions in multi-threaded environments.
  • ReDoS Protection: Implemented guards on maximum regex lengths to prevent Regular Expression Denial of Service attacks in string manipulation logic.

Enhancements​

CQL Translator​

  • Point-to-Interval Promotion: Added automatic promotion of point operands to degenerate intervals (e.g., [x, x]) for StartsSame and EndsSame operators to ensure spec-compliant temporal comparisons.
  • Distinct List Aggregates: Added support for standard aggregates (Sum, Min, Max, Avg) on distinct list literals.
  • Clinical UDF Macros: Externalized clinical logic into structured DuckDB macros for better maintainability.

FHIRPath Engine​

  • Cross-Namespace Equivalence: Updated TypeInfo to correctly treat FHIR primitive types and their System equivalents as the same type per FHIRPath §5.1.
  • Conversion Function Flexibility: Support for both functional and member invocation forms for convertsToX functions (e.g., convertsToBoolean(v) and v.convertsToBoolean()).
  • Semantic Temporal Validation: Added strict validation for calendar dates and times, including month day limits and leap year checks.

DQM & ViewDefinition​

  • Enhanced Narratives: Improved narrative generation for population results when detailed evidence capture is disabled.
  • ViewDef String Escaping: Updated the SQL generator to correctly handle backslashes and single quotes within FHIRPath strings in ViewDefinitions.

Infrastructure & Tooling​

  • Centralized Logging: Added a new unified logging framework that tracks pass rates and failure details across all subprojects.
  • Build-Time Auto-Discovery: Improved the build system to auto-discover DuckDB versions and wheel names, ensuring smoother Pyodide package installations.
  • Benchmarking Submodule: Added tests/data/dqm-content-qicore-2026 as a git submodule to provide a standard set of 2026 QI Core measures for benchmarking.

Bug Fixes​

  • CQL: Fixed circular definition handling (QA-019) which previously caused RecursionError in complex libraries.
  • CQL: Fixed Count(distinct(defRef)) producing invalid SQL when referencing non-existent columns.
  • FHIRPath: Resolved a regression in DateTime > Date comparisons.
  • FHIRPath: Fixed type compatibility logic in equality operators to correctly return empty for incompatible types per §6.1.1.
  • ViewDefinition: Fixed a regression in time boundary generation where an unnecessary 'T' prefix was occasionally added.

API Changes​

  • Version Bump: The fhir4ds, fhir4ds.cql, fhir4ds.fhirpath, and fhir4ds.viewdef packages have all been bumped to version 0.0.2.
  • WASM Assets: Updated the required WASM assets; integrations must now use the 0.0.2 translator wheel.

Version 0.0.1​

Initial Release - Early 2026

The initial release established the foundation for the FHIR4DS engine.

  • Unified Interface: Provided a single entry point for CQL, FHIRPath, and ViewDefinition execution.
  • DuckDB Integration: Implemented the first set of Python-based UDFs for FHIR logic.
  • Initial WASM Support: Demonstrated browser-side clinical reasoning using Pyodide and DuckDB-WASM.