Natural Language Interface for Healthcare Data Analytics using LLMs

aachen_uniklinik_/_aurea_software · Healthcare · year n/a · 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 data analytics

Aachen University Clinic needed ICU data analytics accessible to physicians and department heads without requiring SQL expertise.

Constraints:

• Healthcare users need patient analytics and complication calculations.
• The interface must transform natural language into structured SQL queries.
• Healthcare data access requires intent detection and slot filling rather than free-form generation alone.
• Outputs include visualizations and analytics for clinical decision-making.

Architecture

NLQ built a hybrid system using vector databases with metadata embeddings, fine-tuned LLMs for intent detection and slot filling, and structured query generation over ICU databases.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  natural-language-interface["Natural Language Interface"]
  metadata-vector-database["Metadata Vector Database"]
  intent-and-slot-llms["Intent and Slot LLMs"]
  sql-generator["SQL Generator"]
  icu-databases["ICU Databases"]
  visualization-and-analytics-output["Visualization and Analytics Output"]
  natural-language-interface --> metadata-vector-database
  natural-language-interface --> intent-and-slot-llms
  metadata-vector-database --> intent-and-slot-llms
  intent-and-slot-llms --> sql-generator
  sql-generator --> icu-databases
  icu-databases --> visualization-and-analytics-output

Components

Data flow

Clinical users ask natural-language questions, the system retrieves relevant metadata context, fine-tuned LLM components detect intent and slots, structured query generation creates SQL, and the interface returns analytics or visualizations.

Why this architecture

The hybrid architecture constrains natural-language analytics with metadata retrieval and structured slots so clinicians can query ICU data without writing SQL.

Models used: Fine-tuned large language models · Vector database metadata embeddings

Alternatives considered (roads not taken)

Require SQL expertise from healthcare staff

Rejected because: Physicians and department heads needed direct analytics access.

Use unconstrained text-to-SQL alone

Rejected because: The system needed intent detection, slot filling, and metadata retrieval to ground query generation.

Patterns

query-rewriting (query-generation)

The system rewrites natural-language healthcare questions into SQL queries.

rag-retrieval (retrieval)

Metadata embeddings and vector databases provide schema context for query generation.

fine-tuning-adaptation (model-adaptation)

Fine-tuned LLMs perform domain-specific intent detection and slot filling.

Tradeoffs

Use a hybrid architecture

Chose Vector metadata, fine-tuned LLMs, and structured SQL generation over A generic natural-language chatbot over healthcare data.

Rationale: The healthcare analytics task needed constrained schema grounding and slot filling.

Expose analytics to non-technical staff

Chose Natural-language interface over Analyst- or SQL-mediated access.

Rationale: The system aimed to accelerate clinical decision-making for physicians and department heads.

Gotchas (what bit them)

Clinical analytics questions need structured parameter extraction.

Root cause: ICU analytics questions include intents, slots, and database-specific constraints.

Resolution: Use fine-tuned LLMs for intent detection and slot filling before SQL generation.

Healthcare users need results in decision-ready forms.

Root cause: Physicians and department heads often need visualizations and complication calculations, not just rows.

Resolution: Return visualizations and downstream complication calculations through the interface.

Evaluation

The summary describes enabled workflows but does not provide quantitative accuracy, latency, or adoption metrics.

Metrics: No quantitative metrics reported

Outcomes

• Non-technical healthcare staff can query ICU databases with natural language.
• Physicians and department heads can retrieve analytics, visualizations, and complication calculations.
• The interface supports faster clinical decision-making.

No concrete scale numbers are reported in the summary.