OpenMed 3.3.0 — Local multimodal and ambient workflows¶

OpenMed 3.3.0 brings documents, images, speech and physiological waveforms into a common local workflow. Start with bounded media intake, carry source evidence through extraction and redaction, and send uncertain results to a reviewer before export.

This page describes the candidate for the stable 3.3.0 release. The previous stable release is v3.2.0.

Bounded media intake¶

The offline Python preflight CLI checks media headers, manifests and resource limits before decoding. It covers inputs such as images, PDFs, DICOM, BigTIFF, audio and waveform records. Batch preflight adds aggregate limits, duplicate handling and explicit stop policies. Versioned JSON Schemas let integrations validate manifests, findings, abstentions and provider results.

Document and imaging workflows add reviewed form extraction, OCR and VLM disagreement handling, whole-slide tile provenance, and checks for hidden content in PDFs and Office files. The redaction paths can inventory barcode regions without retaining payloads. They withhold exports when identifier coverage or review evidence is incomplete. Start with the multimodal overview.

Offline speech and ambient drafts¶

The local streaming speech path keeps provisional hypotheses private, screens finalized transcript evidence for identifiers, and holds uncertain clinical tokens for confirmation. It tracks overlapping windows, timestamp revisions, language confidence and session-scoped speaker aliases. The ambient workflow binds capture to consent, handles interruptions and retention deadlines, and carries reviewed corrections through to an evidence-linked FHIR R4 draft. See the local speech provider and reviewed ambient FHIR export.

OpenMedKit adds matching evidence contracts, local review controls, consent-bound capture, and optional Core ML speech. The Android preflight mirror, ONNX streaming ASR, and the opt-in REST preflight route are experimental. See the feature map for platform coverage.

Non-diagnostic waveforms and local decisions¶

Bounded readers cover DICOM ECG, EDF/EDF+, selected WFDB formats and HL7 aECG. ECG interval candidates carry lead, timing and acquisition provenance for human review. The PPG, respiration, pressure and capnography contracts provide acquisition and quality evidence. Use these waveform contracts for engineering and review workflows, with clinical interpretation handled by qualified professionals.

Local decision support adds pinned CPU adapters for d3-edge and d1-omni and a protected System One route. Inputs and modalities depend on the selected backend. Acquire weights separately and review the terms for the revision you deploy. OpenMed records model and license provenance separately from its Apache-2.0 adapter code. See the local decision model guide.

Privacy and reliability fixes¶

• Reject unsafe ZIP expansion and uncovered hidden content before emitting Office exports.
• Check decoder resource limits and preserve evidence lineage through DICOM, PDF and whole-slide processing.
• Validate ONNX external-data references before browser session loading and reject unsupported or ambiguous request media types at the privacy proxy.
• Add OpenMedKit privacy manifests and improve share-extension localization and Dynamic Type.
• Save rollout state atomically, expose value-free CLI configuration provenance, and add TypeScript REST and Node.js examples.

The changelog lists the full changes, including the optional yasbd-lib 1.x update and the OAuthLib security floor for optional integrations.

Upgrade and installation¶

OpenMed requires Python 3.11 or newer. The 3.2-to-3.3 migration guide covers SDK coordinates, optional providers, wire versions, experimental surfaces and validation before deployment.

After publication, use the matching coordinates for each surface.

Validation scope¶

The reference workflow and OpenPHI-Vision fixtures use synthetic data. The documented real-model engineering results for the edge VLM and offline ASR apply to the tested assets and host. Production use requires review of the specific model, device, data and workflow. Keep draft notes, waveform measurements and decision outputs under human review.