Color Health uses GPT-4o to accelerate cancer care and diagnostic workups
Color Health is partnering with OpenAI to deploy a copilot application powered by GPT-4o that accelerates cancer patients' access to treatment by automating the identification of missing diagnostics and the creation of tailored workup plans. This tool aims to reduce the critical delays in cancer care where a four-week delay in treatment can increase mortality risk by 6–13%.
AI-Driven Personalized Treatment Planning
Color Health's copilot integrates patient medical data with clinical knowledge to generate customized treatment plans for clinician review. The system operates through a four-step workflow designed to maintain safety and accuracy:
- Data Extraction and Normalization: The system extracts and normalizes patient information (such as individual risk factors and family history) and clinical guidelines. GPT-4o is specifically used to handle inconsistently structured data found in PDFs and clinical notes.
- Plan Generation: The AI identifies missing diagnostics and generates personalized screening plans, as well as necessary documentation for insurance pre-authorizations and medical necessity.
- Clinician Review: A clinician-in-the-loop evaluates the output and the provided source information, editing the results as necessary.
- Integration: Once approved, the clinician adds the information to the patient's existing treatment plan.
Technical Implementation and Safety
Color developed the proof of concept using OpenAI's APIs and HIPAA-compliant data protection standards. To ensure high accuracy and safety, the team employed several technical strategies:
- RAG over Fine-Tuning: OpenAI engineers recommended using retrieval-augmented generation (RAG) instead of model fine-tuning to improve output quality.
- Visual Analysis: To process complex clinical guidelines containing diagrams, Color used GPT-4 Vision to describe screenshots of those diagrams.
- Rapid Prototyping: The team used the standard ChatGPT interface and custom GPTs to validate clinical workflows before committing full engineering resources.
Clinical Impact and Performance Metrics
Color is currently implementing the tool in a limited phase-in for its own clinicians and is partnering with the University of California, San Francisco Helen Diller Family Comprehensive Cancer Center (UCSF HDFCCC) for retrospective evaluation and targeted rollout.
Initial performance data indicates significant efficiency gains:
- Increased Detection: Healthcare providers using the copilot identify 4x more missing labs, imaging, or biopsy and pathology results compared to those without the tool.
- Reduced Analysis Time: Clinicians can analyze patient records and identify gaps in an average of 5 minutes, whereas manual processes involving fragmented data can lead to weeks of delay.
Color intends to provide AI-generated personalized care plans, under physician oversight, for over 200,000 patients through the second half of 2024.