Guardoc Health is using Amazon Nova models via Bedrock to process over one million clinical documents daily, aiming to reduce documentation errors and compliance costs.

Guardoc Health, a provider of documentation software for long-term care, has integrated Amazon Nova models through Amazon Bedrock to process more than one million clinical documents every day. The system is designed to handle diverse document formats, including handwritten annotations and complex forms, to improve accuracy in patient records and reimbursement compliance.

Operational Impact and Reported Results

Guardoc Health reports that its AI-driven pipeline has led to a 46 percent reduction in documentation errors and a 70 percent drop in audit fines. In a quarterly deployment involving two facilities and 200 patients, the company recorded 847 documentation corrections and 86 issues related to Patient-Driven Payment Model (PDPM) reimbursement accuracy. Additionally, the company observed a 74 percent reduction in hospital transfers per 100 admissions during this period. A separate case study across seven facilities and 1,618 residents identified 10,612 documentation issues.

Technical Architecture and Pipeline

The system utilizes a cost-tiered architecture that begins with Amazon Textract to extract text and structural metadata. Data is chunked by clinical boundaries and embedded using Amazon Titan Text Embeddings V2, then stored in Amazon DynamoDB. A retrieval augmented generation (RAG) process uses a pre-filter and k-nearest neighbour search to identify relevant chunks. Amazon Nova 2 Lite performs an initial text-based filter, while Amazon Nova Pro handles the final multimodal reasoning, analyzing raw PDF bytes for handwriting, signatures, and stamps.

Addressing Documentation Challenges

The company identified physician attestation fields on prior authorization forms and patient-reported symptom sections as areas where earlier pipeline versions struggled. These documents often contain handwritten notes that override printed text. For medication extraction, the system employs a hybrid approach, using Amazon Textract for structured tables and Amazon Nova Pro to interpret handwritten additions and non-standard formats that traditional OCR methods may fail to parse.