Clinical Safeguards API

Important

This feature is currently in Private Preview

The Clinical safeguards API offers a suite of healthcare specific safeguards that are available as an API offering through the Azure Health Agent Service.

Clinical Safeguards

The following safeguards are currently available via the API in the private preview:

  1. Detect Hallucinations And Omissions - Clinical hallucinations and omissions detection Given two texts (usually a Reference Text and a Generated Text), we detect: A. Clinical Omissions - clinical concepts appearing in reference and not in generated text. B. Clinical Hallucinations - clinical concepts appearing in generated text and not in reference

  2. Clinical Provenance - Text source identification Given Source and Generated Documents for each text unit in the generated document find a link to original Reference in the source document.

  3. Clinical Codes Validation - Verify clinical codes exist and relevant to context The safeguard verifies that generated clinical codes are valid clinical codes and that the entities they refer exist in the generated text. There are two types of errors: - Invalid code - A valid code, which exists in the ontology, but is unrelated to the given text

  4. Clinical Anchoring - Provide clinical context and concept identification to prompts Pre-processing safeguard which helps to improve the prompt. Encapsulate contextual information and identifiers (anchors) directly within the input text. This enables Large Language Models to effectively utilize this information during its decision-making process.

  5. Clinical Semantic Validation - Verify LLM text conforms to known valid semantic structures This safeguard validates Generative AI text output by verifying the presence of required entity types and relations for it to be considered valid.

  6. Clinical Conflict Detection - Detect clinical findings that are clinically incompatible This safeguard identifies potential discrepancies or conflicts within clinical documentation, categorizes them by conflict type, and provides references to the relevant text. These conflicts pertain to findings that are clinically incompatible with one another. Clinical Conflict Types: Anatomical, Opposition, Metadata, Value, Contraindication, Comparison and Descriptive.

Customers can apply for the private preview of the Clinical Safeguards API here.