Introduction
Purpose and Scope
This Responsible AI content for the EHR Information Retrieval use case aligns to the CHAI Responsible AI Guide (RAIG) by establishing a Testing and Evaluation (T&E) Framework: a consensus-defined set of methods, metrics, and/or benchmarks for developers and implementers to more concretely evaluate the responsible use of AI-enabled EHR information retrieval solutions.
Teams developing, deploying, or monitoring AI-enabled EHR information retrieval solutions can use CHAI’s consensus-defined T&E Framework to guide evaluation. Additionally, organizations should review use case-specific T&E Frameworks for recommended CHAI-endorsed methods/metrics when completing the CHAI Applied Model Card.
Audience
This document is intended for stakeholders involved in the development, implementation, and governance of AI-enabled EHR information retrieval solutions. As such, these methods/metrics should be tailored to developers and implementers.
Developers: individuals involved in the software development process, including requirements gathering, design, coding, testing, and maintenance of software applications (derived from IEEE, 12207:2017)
Implementers: individuals responsible for the procurement, deployment, and/or overall realization of a system or component in accordance with a specified design (derived from IEEE 829 and IEEE 730)
Use Case Description
This use case outlines a workflow for retrieving, standardizing, and processing diverse data formats from Electronic Health Records (EHRs) and related databases. The goal is to transform structured and unstructured healthcare data into a standardized format for further analysis and decision-making. Effective downstream tasks (e.g., summarization, inference) require clean, well-structured data, making early data processing and quality checks essential for meaningful analysis.
Primary End Users
Healthcare Providers: access real-time patient insights for improved decision-making
Secondary End Users
Quality & Patient Safety Teams: track and trend safety events and quality misses
Health Researchers: create high-fidelity cohorts and registries for hypothesis testing
Patients: personalized search of medical events using natural language
Workflow
The below is an example workflow for an AI-enabled EHR information retrieval solution:
Data Retrieval
Extract information from EHRs and related databases, including:
FHIR resources (structured and unstructured - binary FHIR resources contain unstructured data)
CDA documents (semi-structured)
PDFs, TIFs, and images (unstructured)
HL7v2 messages (structured)
Data Standardization
CDA Processing (e.g., utilize the Microsoft GitHub repository for CDA-to-FHIR transformation or parse the XML)
OCR Processing for PDFs/Images:
Apply Optical Character Recognition (OCR) to extract text.
Capture and store metadata of the original documents.
FHIR Structured Data: JSON parsing to grab specific elements.
Data Consolidation
Convert all structured data into appropriate data format(s) for standardization. Examples may include: JSON, a hash table, docDB, or RDS depending on the application.
Feed data (structured/unstructured) directly into the processing pipeline.
Examples include: similarity searches, chunking, filtering, etc.
Parsing: Extract key information from unstructured text.
Rule-Based Processing: Integrate summarized content with structured data for final output determination.
Tokenization: e.g., entity extraction
Final Decision-Making
Combine structured and summarized unstructured data.
Generate the final output/decision, via rule-based logic or AI driven (e.g., EHR information extraction can create a vector of symptoms/findings which then goes to the sepsis prediction module).
This workflow ensures that all relevant healthcare data—structured and unstructured—is standardized, summarized, and processed efficiently, enabling better clinical and administrative decision-making.