healthcare-ai-guardrails

Healthcare AI Guardrails

Lightweight validation guardrails for AI model inputs/outputs in healthcare workflows, with first-class DICOM support.

Features

• Declarative YAML spec for checks on input and output data
• Built-in validators: numeric ranges, choices, required fields
• Specific DICOM validators: patient age, modality, patient sex, patient position, slice thickness, pixel spacing, image orientation, SOP Class UID, BodyPartExamined, PhotometricInterpretation, pixel intensity range, KVP, X-Ray Tube Current, Exposure Time, Protocol Name, and RT Structure Set ROI presence.
• Generic DICOM validators: check if a tag's value is in a list, check a tag's value representation (VR), and check if a tag's numeric value is within a range.
• Output structure validation via JSON Schema
• Simple Python API and CLI (hc-guardrails)
• HL7 v2 support: basic field, value-in-list, regex, and numeric range checks via simple path syntax (e.g., PID-5.1)
• HL7 v3 (XML) support: XPath-based validators for exists, value-in-list, regex, and numeric range with namespace support
• MCP server (hc-guardrails-mcp): expose the full QA pipeline as tools for AI agents (Claude Desktop, Claude Code, etc.)

Install (users)

Install from PyPI - https://pypi.org/project/healthcare-ai-guardrails:

pip install healthcare-ai-guardrails

This installs the Python API and a CLI named hc-guardrails.

Quick CLI check:

hc-guardrails examples/spec.example.yaml examples/output.sample.json --mode output

If you’re validating DICOM, pydicom and numpy are already included as dependencies.

Install (contributors)

Dev install (includes test, lint tools, and pre-commit):

python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
pre-commit install

With uv (fast Python package manager):

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv
source .venv/bin/activate
uv pip install -e .[dev]
pre-commit install

See CONTRIBUTING.md for full contributor guidelines.

Quick start

Spec (examples/spec.example.yaml):

• Input: verify DICOM patient age in [18, 90], modality in {CT, MR}, patient sex in {M,F,O}, slice thickness/pixel spacing ranges, and sane image orientation
• Output: ensure probability ∈ [0, 1] and match a JSON Schema

Run on a DICOM file:

hc-guardrails examples/spec.example.yaml path/to/file.dcm --mode input

Run on a JSON output:

hc-guardrails examples/spec.example.yaml path/to/output.json --mode output

Autocontouring tutorial (CT + RTSTRUCT)

See examples/tutorials/ for a small end-to-end example that validates a CT input and an RT Structure Set output.

Run the Python walkthrough:

python examples/tutorials/autocontouring_tutorial.py

HL7 v2 (ADT/ORM/ORU etc.)

You can validate HL7 v2 messages using a lightweight path syntax: SEG-Field[rep].Comp.Sub (1-based indices). Examples:

• MSH-9.1 → Message type (e.g., ADT)
• PID-5.1 → Family name
• PID-3[2].1 → Second repetition of PID-3, first component

Example spec: examples/hl7v2.example.yaml (preferred naming; examples/hl7.example.yaml retained for compatibility). Run against the provided sample message (or any .hl7 file starting with MSH):

hc-guardrails examples/hl7v2.example.yaml examples/hl7v2.sample.hl7 --mode input

HL7 v3 (XML/CDA/CCDA)

Validate HL7 v3 XML using XPath with namespaces.

Example spec: examples/hl7v3.example.yaml. Run against the provided sample XML (or any HL7 v3 XML document):

hc-guardrails examples/hl7v3.example.yaml examples/hl7v3.sample.xml --mode input

HL7 v2 vs FHIR

• HL7 v2: Pipe-delimited messages (MSH/PID/OBR/OBX…). Use the HL7 v2 validators and path syntax above (SEG-Field[rep].Comp.Sub). The CLI auto-detects HL7 v2 when the file starts with MSH.
• FHIR: JSON or NDJSON resources (Patient, Observation, Bundle, etc.). Treat these as JSON and validate using json_schema plus the generic validators (range, choice, required_fields). You can author a JSON Schema for your resource(s) and reference it directly in the YAML spec.

Example (FHIR Patient minimal schema):

output:
  - type: json_schema
    name: fhir_patient_minimal
    schema:
      $schema: https://json-schema.org/draft/2020-12/schema
      type: object
      required: ["resourceType", "id"]
      properties:
        resourceType:
          const: "Patient"

Or run the same checks via CLI using the YAML spec:

# Input (CT)
hc-guardrails examples/tutorials/autocontouring_tutorial.yaml path/to/ct.dcm --mode input

# Output (RTSTRUCT)
hc-guardrails examples/tutorials/autocontouring_tutorial.yaml path/to/rs.dcm --mode output

MCP server (AI agent integration)

The package ships a Model Context Protocol server so AI agents (Claude Desktop, Claude Code, any MCP-compatible client) can run the guardrails QA pipeline automatically.

Setup

The hc-guardrails-mcp command is installed alongside the package:

pip install healthcare-ai-guardrails

Configure it as a stdio MCP server in your client. For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "healthcare-guardrails": {
      "command": "hc-guardrails-mcp"
    }
  }
}

For Claude Code:

claude mcp add healthcare-guardrails hc-guardrails-mcp

Available tools

Tool	Description
run_guardrails	Run all validators from a spec against a data file or inline content. Returns per-validator pass/fail results and a summary.
list_validator_types	Return the full catalog of available validator types, grouped by category, with their parameters. Useful for building new specs.
validate_spec	Parse a YAML spec and report any configuration errors (unknown types, YAML syntax problems) before running it on real data.
describe_spec	Load a spec and return a human-readable description of what each validator checks (name, description, severity).

Example: run guardrails from a spec file

Tool: run_guardrails
  spec_path: /data/ct_input_spec.yaml
  data_path:  /data/patient001.dcm
  mode:       input

Response:

{
  "summary": { "total": 5, "passed": 4, "failed": 1, "all_passed": false },
  "results": [
    { "name": "modality_check", "passed": true,  "message": "", "severity": "warning" },
    { "name": "age_range",      "passed": true,  "message": "", "severity": "warning" },
    { "name": "slice_thickness","passed": false, "message": "2.5 > max 2.0", "severity": "error" },
    ...
  ]
}

Example: run guardrails with an inline spec and JSON data

Tool: run_guardrails
  spec_yaml: |
    input:
      - type: range
        name: probability_range
        path: ["probability"]
        min: 0.0
        max: 1.0
  data_content: '{"probability": 0.7}'
  mode: input

Data formats supported

Format	Via data_path	Via data_content
DICOM (.dcm)	Yes	No (binary)
JSON	Yes	Yes
HL7 v2	Yes	Yes (auto-detected by MSH prefix)
HL7 v3 / XML	Yes	Yes (auto-detected by < prefix)

Python API

from healthcare_ai_guardrails import GuardrailRunner
from healthcare_ai_guardrails.validators.dicom import (
    DICOMPatientAgeCheck,
    DICOMModalityCheck,
    DICOMPatientSexCheck,
    DICOMPatientPositionCheck,
    DICOMSliceThicknessCheck,
    DICOMPixelSpacingCheck,
    DICOMImageOrientationCheck,
    DICOMKVPCheck,
    DICOMTubeCurrentCheck,
    DICOMExposureTimeCheck,
    DICOMProtocolNameCheck,
    DICOMRTStructureCheck,
)
from healthcare_ai_guardrails.validators.generic_dicom import (
    DICOMGenericNumericRangeCheck,
    DICOMGenericValueInListCheck,
    DICOMGenericTagTypeCheck,
)
import pydicom

runner = GuardrailRunner(
    [
        # Specific Validators
        DICOMPatientAgeCheck(min_years=18, max_years=90),
        DICOMModalityCheck(allowed_modalities=["CT", "MR"]),
        DICOMPatientSexCheck(allowed=["M", "F", "O"]),
        DICOMPatientPositionCheck(allowed=["HFS", "FFP", "FFS"]),
        DICOMSliceThicknessCheck(min_mm=0.5, max_mm=5),
        DICOMPixelSpacingCheck(min_mm=0.2, max_mm=2.0),
        DICOMImageOrientationCheck(tolerance=1e-3),
        DICOMKVPCheck(min_kvp=80, max_kvp=140),
        DICOMTubeCurrentCheck(min_ma=100, max_ma=500),
        DICOMExposureTimeCheck(min_ms=50, max_ms=200),
        DICOMProtocolNameCheck(allowed=["Axial Brain", "Sagittal Spine"]),
        DICOMRTStructureCheck(required_rois=["Heart", "Lungs"]),
        # Generic Validators
        DICOMGenericValueInListCheck(
            tag="Manufacturer", allowed_values=["SIEMENS", "GE"]
        ),
        DICOMGenericTagTypeCheck(tag="PatientName", expected_vr="PN"),
        DICOMGenericNumericRangeCheck(tag="BeamNumber", min_val=1, max_val=10),
    ]
)

ds = pydicom.dcmread("/path/to/file.dcm")
results = runner.run(ds)
for r in results:
    print(r.name, r.passed, r.message)

YAML Spec schema

Specific DICOM validators:

• dicom_patient_age_range – min_years, max_years, inclusive (default: true)
• dicom_modality_allowed – allowed_modalities: ["CT", "MR", ...]
• dicom_patient_sex_allowed – allowed: ["M", "F", "O"]
• dicom_patient_position_allowed – allowed: ["HFS", "FFP", "FFS"]
• dicom_slice_thickness_range – min_mm, max_mm, inclusive
• dicom_pixel_spacing_range – min_mm, max_mm, inclusive
• dicom_image_orientation_sane – tolerance (default: 1e-3)
• dicom_kvp_range – min_kvp, max_kvp, inclusive
• dicom_tube_current_range – min_ma, max_ma, inclusive
• dicom_exposure_time_range – min_ms, max_ms, inclusive
• dicom_protocol_name_allowed – allowed: ["Axial Brain", ...]
• dicom_rt_structure_present – required_rois: ["Heart", ...]

Generic DICOM validators:

• dicom_generic_numeric_range – tag, unit, min_val, max_val, inclusive
• dicom_generic_value_in_list – tag, allowed_values: [...]
• dicom_generic_tag_type_check – tag, expected_vr

Other generic validators:

• range – path: [..], min, max, inclusive
• choice – path: [..], allowed: [...], case_insensitive
• required_fields – paths: [[..], [..]]

Output validators:

• json_schema – schema: {..} (JSON Schema Draft 2020-12 compatible via jsonschema)
• All generic validators above

Example output schema:

output:
  - type: json_schema
    name: output_schema
    schema:
      type: object
      required: ["probability", "label"]
      properties:
        probability:
          type: number
          minimum: 0
          maximum: 1
        label:
          type: string

Development

Supported Python: 3.10–3.13 (tested in CI on Linux; library is pure Python and should work across platforms). Python 3.10 is the minimum due to the mcp dependency.

Run tests locally:

pytest -q

With uv:

uv run pytest -q

Lint/format/type-check (optional suggestions):

pip install ruff mypy
ruff check .
ruff format .
mypy src

With uv:

uv pip install ruff mypy
uv run ruff check .
uv run ruff format .

Notes

• DICOM tags used include: PatientAge, PatientBirthDate, StudyDate, SeriesDate, ContentDate, Modality, PatientSex, PatientPosition, SliceThickness, PixelSpacing, ImageOrientationPatient, KVP, XRayTubeCurrent, ExposureTime, ProtocolName, SOPClassUID, StructureSetROISequence.
• Age parsing supports Y/M/W/D suffixes (per DICOM), falls back to birthdate computation.
• Validators never raise; failures are returned as ValidationResult and can be surfaced as warnings or errors.

Contributing

PRs welcome! See CONTRIBUTING.md for setup instructions, how to add validators, and the PR checklist.

To create DICOMs in tests, use create_test_dicom from healthcare_ai_guardrails.testing.dicom_factory.

Releases and changelog

• PyPI: https://pypi.org/project/healthcare-ai-guardrails/
• Changelog: see CHANGELOG.md

Maintainers (publishing):

• Create a GitHub Release on the main branch. The workflow runs tests across Python 3.9–3.13, builds the sdist and universal wheel, and publishes to PyPI.
• Ensure the repository has PYPI_API_TOKEN set in Secrets.

License

MIT