Skill UI

This skill deploys the OpenMed Model Context Protocol (MCP) server, exposing critical clinical NLP tools such as NER, PII extraction, and de-identification. It enables AI coding agents and chat clients to interact with OpenMed's capabilities via simple tool calls, running locally on-device to ensure user data privacy and security. Supports both stdio and Streamable HTTP deployment modes.

This skill evaluates de-identification and clinical NER models against rigorous leakage-first release gates (G1a-G8). Instead of relying on aggregate F1 scores, it focuses on detecting residual Protected Health Information (PHI) leakage and enforcing per-label recall floors, which is crucial for HIPAA compliance. Determines if a model is RELEASABLE or QUARANTINED.

This skill facilitates the extraction of population-scale (cohort-level) clinical text from Electronic Health Records (EHR) or FHIR data warehouses using the FHIR Bulk Data Access ($export) operation. It manages the asynchronous job kickoff, polling, and downloading of NDJSON files. The resulting data stream is designed to be fed in batches into OpenMed for mandatory de-identification and Named Entity Recognition (NER) analysis.

This skill extracts critical patient metadata (PHI) from DICOM headers using pydicom, and recursively pulls structured report narrative text from DICOM-SR content trees. It is essential for ingesting medical images into NLP pipelines, enabling subsequent de-identification and clinical text analysis.

Use openmed.extract_pii to detect and retrieve specific types of Protected Health Information (PHI) or Personally Identifiable Information (PII) spans (names, dates, MRNs, etc.) from clinical text. Unlike deidentification, this function returns the original text and the detected spans, allowing users to audit, preview redactions, normalize labels, or route spans to a custom anonymizer without altering the source data. It is ideal for pre-processing and policy enforcement.

This skill facilitates the ingestion of unstructured clinical notes and reports from FHIR R4 EHR APIs (e.g., Epic, Cerner). It handles complex tasks including paginating large search results, decoding base64 attachments (PDF, text), and extracting narrative content. The extracted text is then passed to the OpenMed pipeline for HIPAA-compliant de-identification and subsequent NLP analysis, enabling structured data insights from raw clinical records.

This skill implements a critical CI gate to ensure the privacy and performance integrity of de-identification models. It automatically fails the build if the model's recall on held-out PHI data drops below the defined threshold, or if any critical Personally Identifiable Information (PHI) is leaked (leakage must be zero). It enforces strict privacy standards in continuous integration pipelines, blocking unsafe merges.

Synthea generates statistically realistic, fully synthetic patient records in multiple standards (FHIR R4, C-CDA, CSV). This tool is essential for developers and researchers needing safe, shareable test data for OpenMed pipelines, CI fixtures, or specialized de-identification leakage gate test sets. It ensures zero risk of exposing real Protected Health Information (PHI).

This skill replaces detected Protected Health Information (PHI) in clinical notes with realistic, type-matched synthetic surrogates. Unlike simple redaction, surrogates maintain the natural readability and structure of the text, ensuring the data remains parseable for downstream NLP tasks, QA, or human review. It supports consistent, reproducible fake identities and allows custom identifier generation for proprietary formats.

This module converts diverse clinical inputs—including scanned images, faxes, CSV/TSV exports, and C-CDA XML—into a normalized, clean text format (`ExtractedDocument`). It supports on-device OCR (Tesseract/PaddleOCR) and structured redaction for tables. It is the crucial first stage for feeding de-identification and Named Entity Recognition (NER) pipelines.

This skill parses standardized C-CDA/CCD XML clinical documents exported from EHRs. It extracts both human-readable section narrative text and coded entries, mapping them by LOINC codes and template IDs. It is essential for preparing clinical free-text data for subsequent NLP analysis and supports XML-aware de-identification while preserving document structure.

This skill decodes complex pipe-delimited HL7 v2.x messages (including ADT, ORU, MDM) into structured segments and fields. It is essential for ingesting HL7 feeds from interface engines, allowing users to surface embedded clinical narratives (OBX-5 and NTE-3) for subsequent Natural Language Processing (NLP) analysis. It also provides segment-aware de-identification capabilities to redact PHI while preserving the native HL7 message structure.