Scientific Intent Translation System for Healthcare Analytics Using Amazon Bedrock

aetion · Healthcare · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: healthcare analytics scientific-intent translation

Aetion needed to let healthcare and life-science users express scientific intent in natural language and receive valid AEP Measures instructions without requiring data science expertise, support staff, or lengthy specialized training.

Constraints:

• Scientific variables can involve sequences of clinical events, non-occurrences, numeric calculations, and categorizations over patient journeys.
• The assistant handles sensitive healthcare analytics workflows and must keep data encrypted in transit.
• Generated instructions must be valid for AEP Measures, not just plausible natural language.
• The system must compensate for LLM logical-reasoning errors through local knowledge and guardrails.

Architecture

Aetion built Measures Assistant as a Kubernetes-on-AWS microservice behind a REST API. Substantiate sends user questions and prior message history to the service, which fills a prompt template, calls Amazon Bedrock to invoke Claude 3 Haiku, and augments responses with a local AEP Measures knowledge base, dynamic semantically similar Q&A examples, static instructions, rules, and SME-maintained feedback loops.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  A["Assistant UI"] --> B["Substantiate request layer"]
  B --> C["Measures Assistant microservice"]
  C --> D["Prompt template"]
  D --> G["Local AEP Measures knowledge base"]
  C --> E["Amazon Bedrock API"]
  E --> F["Claude 3 Haiku"]
  G --> H["Mixedbread embedding model"]
  I["Subject matter experts"] --> G

Components

Data flow

A user asks a scientific question through the assistant UI; Substantiate sends the question and optional history over TLS to the Measures Assistant microservice; the service inserts the question into a prompt template with AEP documentation, static rules, and dynamic semantically similar examples from a local knowledge base; Amazon Bedrock invokes Claude 3 Haiku; the assistant returns valid Measure instructions for Aetion's analytics workflow.

Why this architecture

Aetion chose Bedrock and Claude 3 Haiku for managed model access, security posture, runtime, and cost; wrapped the model with a local AEP knowledge base and rules so outputs stayed valid; and used human-maintained examples to keep the assistant aligned with expert scientific workflows.

Models used: Anthropic Claude 3 Haiku · Mixedbread mxbai-embed-large-v1 Sentence Transformer

Scale

• Aetion serves most top-20 biopharma companies plus payors and regulators.
  “Aetion has partnered with the majority of top 20 biopharma, leading payors, and regulatory agencies.”article, stated
• Aetion operates across the US, Canada, Europe, and Japan.
  “customers in the US, Canada, Europe, and Japan with analytics”article, stated
• The assistant reduces measure creation from days to minutes.
  “Users now can turn questions expressed in natural language into measures in a matter minutes as opposed to days”article, stated
• The microservice encrypts data in transit with TLS 1.2.
  “The data transmitted to the service is encrypted using Transport Layer Security 1.2 (TLS).”article, stated

Alternatives considered (roads not taken)

Use a model without a local AEP knowledge base

Rejected because: Aetion needed valid Measure instructions and compensation for logical reasoning errors.

Use only static prompt instructions

Rejected because: Dynamic semantically similar Q&A examples improved answer quality for the user's specific question.

Select another available LLM

Rejected because: Claude 3 Haiku was more efficient in runtime and cost than alternatives.

Patterns

rag-retrieval (retrieval)

Measures Assistant dynamically selects semantically similar solved Q&A examples from a local knowledge base and inserts them into the prompt.

safety-policy-enforcement (safety)

Prompt rules and knowledge-base guardrails restrict creativity, keep sensitive data secure, and force valid AEP Measure instructions.

human-in-the-loop-gate (quality)

Subject matter experts continuously test the assistant and use Q&A pairs to refine its behavior.

Tradeoffs

Use Bedrock and Claude 3 Haiku.

Chose managed Bedrock invocation of Claude 3 Haiku over other available LLM alternatives.

Rationale: Aetion valued provider selection, security posture, extensibility, ease of use, and Haiku's runtime/cost efficiency.

Combine static prompt structure with dynamic examples.

Chose static task/rule instructions plus semantically selected Q&A examples over a fixed prompt without retrieval.

Rationale: Static instructions define behavior while dynamic examples tailor the prompt to the user's question.

Keep SMEs in the loop.

Chose continuous SME testing and Q&A pool maintenance over fully automatic prompt/knowledge maintenance.

Rationale: Scientific correctness and user experience depended on expert-maintained question-and-answer examples.

Gotchas (what bit them)

Natural-language scientific intent can be clinically complex.

Root cause: Variables may require event sequences, combinations of occurrences and non-occurrences, numeric calculations, and categorizations over patient timelines.

Resolution: Use a specialized Measures Assistant that translates user intent into AEP Measures instructions.

The base model can make logical reasoning errors for valid Measure construction.

Root cause: General-purpose LLMs do not inherently know all AEP Measure constraints.

Resolution: Add local expert knowledge, prompt rules, and dynamic solved examples as guardrails.

Healthcare analytics prompts and responses require security controls.

Root cause: The assistant handles sensitive scientific and healthcare workflow data.

Resolution: Deploy as an AWS Kubernetes microservice and encrypt user prompts and Bedrock requests with TLS 1.2.

Evaluation

The article does not report a formal benchmark, but it describes SME continuous testing, Q&A-pool refinement, runtime/cost comparison for model choice, and user-facing time-to-measure improvement.

Metrics: runtime and cost efficiency · time from natural-language question to measures · SME continuous testing · valid Measure instructions

Outcomes

• Users can translate natural-language scientific intent into AEP Measures in minutes rather than days.
• The assistant supports users without data science expertise, support staff, or specialized training.
• Measures Assistant enables scientists to iterate on study designs through quick natural-language queries.

Aetion's customer footprint includes most top-20 biopharma companies and regulated healthcare analytics markets across the US, Canada, Europe, and Japan.