@flexpa/llm-fhir-eval

[!NOTE] Follow the development progress on FHIR Chat.

Overview

@flexpa/llm-fhir-eval is an evaluation framework designed to benchmark the performance of LLMs on FHIR-specific tasks including generation, validation, and extraction. This framework systematically tests and validates the capabilities of LLMs in handling various healthcare-interoperability related tasks, ensuring they meet the standards required for effective FHIR implementations. It implements evaluations from prior art such as FHIR-GPT.

Benchmark

@flexpa/llm-fhir-eval benchmarks FHIR-specific tasks including:

1. FHIR Resource Generation:

   • Generate accurate FHIR resources such as Patient, Observation, MedicationStatement, etc.
   • Test the ability to create complex resource relationships and validate terminology bindings.

2. FHIR Resource Validation:

   • Validate FHIR resources using operations like $validate.
   • Check for schema compliance, required field presence, and value set binding verification.

3. Data Extraction:

   • Extract structured FHIR-compliant data from clinical notes and other unstructured data.
   • Evaluate the proficiency of LLMs in extracting specific healthcare data elements.

4. Tool Use:

   • Test models' ability to use FHIR validation tools and other healthcare-specific functions.
   • Validate proper tool calling for FHIR operations.

Available Evaluations

1. Data Extraction (evals/extraction/)

   • Description: Comprehensive evaluation of extracting structured FHIR data from unstructured clinical text.
   • Configurations: Both minimalist and specialist approaches available.
   • Test categories: Basic demographics, conditions, explanations of benefit, medication requests, observations.

2. FHIR Resource Generation (evals/generation/)

   • Description: Tests the ability to generate valid FHIR resources and bundles.
   • Configurations: Zero-shot bundle generation and multi-turn tool use scenarios.
   • Models supported: GPT-3.5-turbo, GPT-4.1, O3 (low/high reasoning), Claude 3.5 Haiku, Claude 3.5 Sonnet, Claude Sonnet 4, Claude Opus 4

Custom Assertions

The framework includes custom assertion functions:

• fhirPathEquals.mjs: Validates FHIR Path expressions
• isBundle.mjs: Checks if output is a valid FHIR Bundle
• metaElementMissing.mjs: Validates required metadata elements
• validateOperation.mjs: Validates FHIR operation results

Tools

• validateFhirBundle.mjs: Tool for validating FHIR Bundle resources

Custom Providers

• AnthropicMessagesWithRecursiveToolCallsProvider.ts: Enhanced Anthropic provider with recursive tool calling (up to 10 depth levels)
• OpenAiResponsesWithRecursiveToolCallsProvider.ts: Enhanced OpenAI provider with recursive tool calling

These providers enable multi-turn tool interactions where models can iteratively call validation tools to improve their FHIR resource generation.

Commands to Run Evaluations

Install dependencies and set up environment variables:

bash yarn install

Copy the .env.template file to .env and supply your API keys for the models you plan to test.

Run an evaluation:

```bash

Example: Run the extraction evaluation with minimalist config

promptfoo eval -c evals/extraction/config-minimalist.yaml

Example: Run the FHIR bundle generation evaluation

promptfoo eval -c evals/generation/config-zero-shot-bundle.yaml

Example: Run multi-turn tool use evaluation

promptfoo eval -c evals/generation/config-multi-turn-tool-use.js ```

The evaluation will print its performance metrics to the console and optionally save results to files.