How Carta Healthcare gets AI to reason like a clinical abstractor

2026-04-08 · Source: Claude Blog · Field: Health & Wellbeing — Clinical Care & Medical Practice, Medical Devices & Health Technology · Depth: Intermediate, short

Summary

Carta Healthcare's Lighthouse platform utilizes Claude in Amazon Bedrock to automate clinical data abstraction, processing over 22,000 surgical cases annually across 14 hospitals with 98-99% accuracy. This system addresses the challenge of converting complex, inconsistent patient records into standardized data for clinical registries, a task that traditionally requires extensive skilled labor (over 11,000 hours annually for a single registry program). Unlike previous rules-based or NLP systems that struggled with the variability of clinical language and the need for contextual reasoning, Lighthouse focuses on "context engineering." This involves meticulously assembling the correct patient-specific documentation, time windows, and priority order at runtime to enable Claude to reason like a trained abstractor, weighing conflicting evidence and applying temporal logic.

Key takeaway

For AI Architects developing solutions in complex, context-dependent domains like healthcare, you should prioritize "context engineering" over solely optimizing model prompts. Your focus must be on building robust data pipelines that dynamically assemble and present the most relevant, time-bound information to the LLM. This approach ensures the AI can perform nuanced reasoning, leading to higher accuracy and faster iteration cycles for integrating expert feedback into production systems.

Key insights

Effective AI reasoning in complex domains hinges on precise context engineering, not just model capability.

Principles

• Clinical documentation is too inconsistent for rules-based automation.
• AI performance is determined by context provided, not just the model.
• Granular evaluation frameworks isolate performance variables.

Method

Build a pipeline to assemble the right documentation, timeframe, and priority order for each query at runtime, ensuring the AI model receives precise, relevant context.

In practice

• Specify exact time boundaries for data extraction queries.
• Integrate clinical expertise directly into prompt refinement.
• Design evaluation to isolate prompt, context, or retrieval issues.

Topics

• Clinical Data Abstraction
• Carta Healthcare
• Lighthouse Platform
• Context Engineering
• Claude LLM

Best for: AI Architect, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• AIPatient Arena: EHR-grounded evaluation of large language models in end-to-end clinical consultation workflows — AIPatient Arena reveals LLMs need process-based evaluation beyond final answers for clinical readiness, highlighting interaction failures.
• Enabling a new model for healthcare with AI co-clinician — AI co-clinician research explores triadic care, augmenting physicians and assisting patients through multimodal AI under clinical supervision.
• Vivek Natarajan of Google DeepMind at RAAIS 2026 — Building useful AI in expert domains requires moving beyond benchmarks to real-world clinical interaction and new knowledge generation.
• Grounding Clinical AI Competency in Human Cognition Through the Clinical World Model and Skill-Mix Framework — Clinical AI competency requires a formalized world model and a multi-dimensional skill-mix framework for robust evaluation.
• From Radiology to Drug Discovery, Survey Reveals AI Is Delivering Clear Return on Investment in Healthcare — Healthcare and life sciences are rapidly moving from AI experimentation to execution, driven by strong ROI and open-source adoption.
• Benchmarking GPT Models for Conversational AI Systems: Can AI Read a Doctor’s Notes? — LLMs struggle with medical abbreviation expansion, especially context-dependent ones, despite larger models performing better.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Claude Blog.