Unlocking Patient Population Insights Using Smart Subgroups and LLMs

aetion · Healthcare · 2025 · source Agentic: no · promotional bias: high · 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 real-world evidence analytics

Aetion needed to make patient-population subgroup analysis accessible through natural language for researchers who lacked deep familiarity with the underlying real-world datasets and structured query formats.

Constraints:

• The system operated in a healthcare analytics environment with stringent security requirements.
• Patient subgroup interpretation depended on over 1,000 computed scientific features.
• Outputs needed to explain clusters using medical feature descriptions and prevalence statistics.
• Interactive use required a latency-conscious model choice.
• The public case study did not disclose robust evaluation, guardrail, or monitoring details.

Architecture

Aetion added a Smart Subgroups Interpreter to its Discover product by combining traditional unsupervised ML clustering with a Bedrock-hosted LLM interface. AEP computes more than 1,000 AML features per patient, stores feature data in S3 and Aurora, runs on-demand Kubernetes clustering jobs, identifies distinctive cluster features and prevalences, and passes expanded feature descriptions plus statistics into Claude 3 Haiku prompts to name clusters and answer subgroup questions.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  aetion-discover-user-interface["Aetion Discover user interface"] --> aetion-evidence-platform["Aetion Evidence Platform"]
  aetion-discover-user-interface --> smart-subgroups-interpreter["Smart Subgroups Interpreter"]
  aetion-evidence-platform --> amazon-s3-feature-store["Amazon S3 feature store"]
  aetion-evidence-platform --> smart-subgroups-clustering-job["Smart Subgroups clustering job"]
  amazon-s3-feature-store --> smart-subgroups-clustering-job
  smart-subgroups-clustering-job --> smart-subgroups-interpreter
  smart-subgroups-interpreter --> amazon-aurora["Amazon Aurora"]
  smart-subgroups-interpreter --> prompt-construction-layer["Prompt construction layer"]
  smart-subgroups-interpreter --> amazon-bedrock["Amazon Bedrock"]
  prompt-construction-layer --> claude-3-haiku["Claude 3 Haiku"]
  claude-3-haiku --> amazon-bedrock

Components

Data flow

A user defines a patient population with the Aetion Measure Library. AEP computes standardized features, stores them in S3, and Smart Subgroups trains a topic model and classifier to find clusters and distinctive features. The interpreter expands coded features into human-readable clinical descriptions, injects prevalence statistics into versioned prompt templates, calls Claude 3 Haiku through Bedrock, and returns natural-language cluster names or subgroup answers.

Why this architecture

Aetion used traditional ML to perform the patient clustering and distinctive-feature identification, then used an LLM as an interpretation layer. That division kept quantitative subgroup discovery grounded in computed features while letting non-specialists query and understand the resulting clusters in natural language.

Models used: Anthropic Claude 3 Haiku · Topic model · Classification model

Scale

• AEP computes over 1,000 AML features for each patient.
  “The Aetion Evidence Platform (AEP) then computes over 1,000 AML features for each patient across categories including diagnoses, therapies, and procedures.”summary, stated
• Aetion serves regulatory agencies including the FDA and EMA.
  “The company serves top pharmaceutical companies, regulatory agencies including the FDA and EMA, and operates across the US, Canada, Europe, and Japan.”summary, stated
• The system reports time-to-evidence improvements from days to minutes.
  “The claimed benefit is reducing the time to generate decision-grade evidence from days to minutes, without requiring support staff.”summary, stated
• Transport encryption used TLS 1.2.
  “Transport Layer Security 1.2 for data encryption in transit”summary, stated

Alternatives considered (roads not taken)

Ask the LLM to perform clustering analysis directly

Rejected because: Aetion used unsupervised ML for clustering and reserved the LLM for interpretation, which may be more reliable.

Use a more capable but slower model

Rejected because: Aetion selected Claude 3 Haiku to balance speed with sufficient intelligence for interactive subgroup interpretation.

Deploy generative AI as a standalone product

Rejected because: Aetion embedded the feature into its existing Evidence Platform workflow so users could move from exploration to causal inference.

Patterns

context-engineering (prompting)

Aetion constructed prompts from scientific feature expansions and prevalence statistics so the model could explain coded healthcare clusters.

prompt-versioning (operations)

The feature used versioned prompt templates as managed artifacts for iteration and rollback.

model-routing (model operations)

The system assigned analytical clustering to traditional ML and natural-language interpretation to Claude 3 Haiku.

batch-offline-processing (infrastructure)

The clustering step runs as an on-demand Kubernetes job rather than permanently provisioned compute.

privacy-preserving-processing (security)

Healthcare feature data used KMS encryption at rest, TLS in transit, and Bedrock's managed security posture.

Tradeoffs

Hybrid ML plus LLM interpretation

Chose Use traditional ML for clustering and Claude for interpretation over Ask the LLM to discover patient clusters directly.

Rationale: Quantitative clustering remains grounded in computed features while the LLM makes results accessible.

Claude 3 Haiku model choice

Chose A faster, satisfactory-intelligence model over A more capable but potentially slower model in the Claude family.

Rationale: Interactive subgroup exploration needed acceptable latency as well as sufficient clinical-language capability.

Embedded platform integration

Chose Add Smart Subgroups Interpreter inside Aetion's Evidence Platform over Ship a standalone generative AI tool.

Rationale: Embedding AI in the existing workflow let users move from exploratory subgroup discovery to rigorous causal inference.

Gotchas (what bit them)

Healthcare cluster interpretation needs domain context before the LLM can produce meaningful labels.

Root cause: Raw AML feature codes are too terse for reliable natural-language interpretation.

Resolution: Aetion expanded features into human-readable scientific descriptions and injected prevalence statistics into prompts.

The public case study does not describe evaluation or guardrails for medical accuracy.

Root cause: The blog focuses on architecture and reported benefits rather than LLM quality governance.

Resolution: The extraction should treat reported healthcare outcomes as promotional and note missing validation details.

Latency mattered enough to influence model choice.

Root cause: The feature was an interactive user-facing healthcare analytics experience.

Resolution: Aetion chose Claude 3 Haiku for speed while accepting a satisfactory intelligence level for the task.

Evaluation

The article reports product outcomes and architectural choices but does not disclose a detailed evaluation methodology for LLM response quality, medical accuracy, guardrails, monitoring, or user feedback.

Metrics: Over 1,000 AML features per patient · Days-to-minutes reported time savings · No disclosed LLM response-quality metric · No disclosed medical accuracy guardrail metric

Outcomes

• Researchers can query patient subgroups in natural language.
• The system generates descriptive cluster names and answers subgroup questions from feature statistics.
• Aetion claims time to decision-grade evidence fell from days to minutes without support staff.
• Reported outcomes are self-reported in an AWS/Aetion promotional case study.

The system handles patient populations described by more than 1,000 AML features per patient and is part of a platform used by biopharma, payor, and regulatory organizations across the US, Canada, Europe, and Japan.