Co-leading an AI value-based care application

At Lumeris I co-led development of an AI application for value-based care. It ran as a Windows desktop application next to the Electronic Medical Record (EMR) system physicians already used, and its job was to identify patients and surface information that might otherwise be missed.

Why it mattered

In value-based care, providers are rewarded for patient outcomes rather than the volume of services. That only works when the right information reaches the physician at the right time, and important details are often buried in documents nobody has time to read in full.

How it worked

• OCR with AWS Textract turned scanned and unstructured documents into text and structured data.
• LLMs on AWS Bedrock read that text and pulled out what was relevant to the patient.
• Embeddings in Pinecone made the extracted information searchable by meaning, so related records could be found even when they used different wording.
• Electron packaged it as a Windows app that runs alongside the existing EMR instead of replacing it.

Making the output trustworthy

An AI tool in a clinical setting is only useful if physicians can trust it. Much of the work went into confidence score calibration, so the app's confidence reflected how often it was actually right, and into prompt engineering to keep the model's answers focused and consistent.

Results

During beta testing, the application helped prevent two potential ER visits by surfacing information in time for physicians to act on it.