AI-Powered Medical Document Processing in Long-Term Care

Overview

Guardoc Health operates in the long-term care sector, serving skilled nursing facilities (SNFs) and assisted living facilities (ALFs) where clinical documentation remains a persistent source of operational risk, patient safety concerns, and financial liability. The company’s mission centers on transforming medical data into accurate, complete, and compliance-aligned clinical documentation that empowers nurses and care teams to deliver safer care. The operational challenge is significant: medical records arrive in every conceivable format including multi-page PDFs with handwritten physician annotations, prior authorization forms where checkbox states determine coverage, medication lists in tables or free text, and patient intake forms mixing typed fields with handwriting and stamps. At scale—processing over 1 million documents on peak days—even minor error rates translate to thousands of incorrect records with potential patient safety and compliance implications.

This case study presents a production LLMOps implementation that combines retrieval-augmented generation (RAG), multimodal foundation models, OCR services, and cost-tiered processing architecture to address three core challenges: detecting special medical conditions from patient records with high recall, reliably interpreting PDF checkboxes across dozens of form types, and accurately extracting information from documents that mix formats within a single page. While the case study is presented as an AWS blog post and naturally emphasizes positive outcomes, it provides substantial technical detail about the production architecture and offers insight into real-world deployment considerations for medical document AI systems.

Technical Architecture and LLMOps Implementation

The core production pipeline operates as a multi-stage RAG system built on Amazon Bedrock and supporting AWS services. The architecture reflects a deliberate cost-tiering principle where each processing stage uses the least expensive component capable of handling the work, reserving more computationally intensive multimodal reasoning for final-stage classification where it delivers the most value.

The medical condition classification pipeline begins with Amazon Textract performing initial OCR and structural extraction from incoming patient documents. This first-stage processing extracts text and structural metadata from each page at low per-page cost, going beyond basic OCR by identifying document structure including forms, tables, and layout information. The extracted content is then chunked along boundaries appropriate to the document type rather than arbitrary byte ranges. This chunking strategy ensures that segments correspond to meaningful clinical units such as a medication list, diagnosis section, or physician note, preserving the context needed for accurate downstream classification.

Each chunk is embedded using Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, with critical partitioning by patient to ensure retrieval never crosses patient boundaries—a fundamental requirement in healthcare applications. Before retrieval begins, a custom pre-filter evaluates document type, recency, and patient-context signals to narrow the candidate set early in the pipeline. This pre-filtering reduces the volume of content requiring search operations and lowers overall retrieval costs, a practical LLMOps consideration when operating at the scale of millions of documents.

Within the pre-filtered set, an in-memory k-nearest neighbor search retrieves chunks most relevant to the condition classification query. Because retrieval is scoped per patient, the system scales horizontally with patient volume without relying on a shared vector index. This stage returns only page references, keeping data transfer lightweight before advancing to more intensive multimodal processing.

The retrieved pages then pass through Amazon Nova 2 Lite, a lightweight, cost-efficient multimodal model that performs text-based review to remove obvious non-matches. This intermediate filtering stage sits between retrieval and the final multimodal reasoning step, eliminating documents that don’t require full visual analysis. Only pages that survive all prior filtering stages reach Amazon Nova Pro along with the raw PDF bytes. At this final stage, the Nova model reasons over layout, handwriting, signatures, stamps, and other visual context to produce the classification decision.

A critical production requirement throughout this pipeline is traceability: every classification output resolves back to the specific source page it came from. In clinical settings where decisions carry direct patient safety implications, the system treats this level of traceability as non-negotiable, enabling clinical staff to verify AI-generated classifications against original source documents.

Multimodal PDF Processing

The majority of medical documents exist as PDFs—whether scanned, photographed, or born-digital—and processing them accurately requires models that read pages visually rather than as plain text streams. The case study indicates that Amazon Nova Pro demonstrated meaningful improvements in multimodal PDF handling during evaluation, including interpretation of layout, form structure, checkbox states, signatures, handwritten annotations, and spatial relationships between fields on a page.

This multimodal capability proves essential across multiple document types that Guardoc processes: checkboxes on prior authorization forms, handwriting in physician attestation fields, multi-column medication lists, stamped attestations, and struck-through edits. The ability to process these elements as visual documents rather than text streams forms the foundation for the rest of Guardoc’s document processing stack. For example, physician attestation fields on prior authorization forms may contain handwritten notes that override printed checkboxes, requiring visual reasoning to determine the authoritative information. Similarly, patient-reported symptom sections often contain handwriting that carries information not captured elsewhere in the record, and missing this text represented a significant source of information loss in earlier pipeline versions according to the case study.

Hybrid OCR Pipeline for Medication Extraction

Medication data represents some of the most consequential content in patient records, feeding directly into clinical decision-making, medication reconciliation, prior authorization, and coverage decisions. The challenge lies in format variety: medications appear in structured tables, unstructured prose within physician notes, handwritten additions to printed lists, and scanned pages that have been faxed or photographed through multiple channels. A single patient chart can contain all these formats across different pages.

Guardoc’s production solution is a hybrid pipeline pairing Amazon Textract with Amazon Nova models. Amazon Textract performs OCR and produces structured output with bounding boxes, table structures, and form field recognition. For clean tables and printed medication lists, Textract provides fast, cost-efficient, and accurate extraction. Amazon Nova Pro receives both the original PDF and the Textract output, which includes machine-readable text alongside the visual layout of the source page. The Nova models complement Textract by resolving complex extraction scenarios including medications split across wrapped table columns, handwritten additions to printed lists, non-standard table formats, and inline mentions buried in physician notes.

This division of labor represents a practical LLMOps pattern: each component handles what it does best, with high-volume structured extraction handled by specialized OCR services and complex multimodal reasoning reserved for cases that require it. The result is medication data that downstream clinical workflows can act on directly without manual reconciliation—a critical requirement when operating at production scale.

Cost Optimization and Production Considerations

The architecture demonstrates several production-oriented cost optimization strategies. The cost-tiering principle runs throughout the pipeline: Amazon Titan Text Embeddings and Amazon Nova 2 Lite handle high-volume, lightweight tasks such as embedding, indexing, and coarse filtering, while Amazon Nova Pro is reserved for final-stage multimodal reasoning over raw PDF bytes. The pre-filtering stage removes irrelevant documents before they reach expensive operations, and the intermediate filtering with Nova 2 Lite further reduces the volume reaching the most computationally intensive Nova Pro stage.

The patient-scoped retrieval architecture allows horizontal scaling without shared vector indexes, and the decision to return only page references from the k-NN search rather than full document content minimizes data transfer before multimodal processing. These are practical considerations that become significant when processing millions of documents at production scale.

Business Impact and Production Performance

The case study reports substantial business impact, though as with any vendor-presented case study, these figures should be considered in context. A quarterly deployment across two facilities covering 200 patients reportedly drove 847 documentation corrections, addressed 86 PDPM-impact issues (Patient-Driven Payment Model, where Medicare reimbursement is tied to documentation accuracy), and was associated with a 74% reduction in hospital transfers per 100 admissions. The reduction in hospital transfers is particularly notable given that such transfers represent costly and destabilizing events in long-term care.

In a broader deployment spanning seven facilities and 1,618 residents, the system identified 10,612 issues and improved visibility into facility-level risk patterns. A separate quarterly analysis reported an average 46% reduction in documentation errors, with outcomes including over $400K in annual ROI for a single facility, a 70% reduction in audit fines, and a 65% reduction in litigation exposure.

These metrics suggest that improvements in documentation accuracy translate to measurable financial and compliance outcomes. However, without independent verification or details on measurement methodology, these should be understood as reported outcomes rather than validated benchmarks. The case study does not provide details on false positive rates, manual review requirements, or edge cases requiring human intervention—all important considerations for production AI systems in healthcare.

Model Selection and Evaluation

The case study indicates that Amazon Nova Pro was selected based on evaluation showing meaningful improvements in multimodal PDF handling compared to alternatives. However, the text does not provide details on the evaluation methodology, baseline comparisons, or specific metrics used to make this determination. For practitioners considering similar implementations, the practical guidance offered is straightforward: benchmark on actual document types, measure recall and precision separately, and set confidence thresholds that reflect the real cost of each error type in the specific clinical context.

The expanded context windows in the Amazon Nova 2 family are mentioned as making multi-document reasoning more tractable, enabling the pipeline to reason across larger volumes of clinical content in a single pass. Continued improvements in multimodal capabilities are described as reducing edge cases such as degraded scans, mixed-format pages, and non-standard layouts that currently require human review.

Production Roadmap and Future Directions

Guardoc’s stated roadmap expands beyond compliance into broader clinical workflow automation, with planned modules including AI NoteAssist, MDSAssist, CarePlanAssist, AdmissionAssist, and Pharmacy Reconciliation. These modules are designed to help teams document accurately, reduce manual workload, and move efficiently through the full clinical workflow within electronic health record (EHR) systems. The expanded context windows in newer model versions are positioned as enabling more comprehensive multi-document reasoning, though specific technical details on how these capabilities will be implemented are not provided.

Critical Assessment and Limitations

This case study is fundamentally a vendor success story published on the AWS blog, and should be read with that context in mind. While it provides substantial technical detail about the production architecture, it naturally emphasizes positive outcomes and does not extensively discuss limitations, failure modes, or challenges encountered during development and deployment.

The reported business metrics—46% error reduction, 70% reduction in audit fines, $400K ROI—are presented without details on measurement methodology, baseline conditions, or independent verification. The 74% reduction in hospital transfers is described as “associated with” the deployment rather than causally attributed, which is appropriate given the complexity of factors influencing such outcomes, but this also means the relationship should not be overstated.

The case study does not discuss false positive rates, how confidence thresholds are set in production, what percentage of documents require manual review, or how the system handles edge cases and failure modes. It does not describe the human-in-the-loop workflows that likely exist for validation, the training requirements for clinical staff using the system, or how the system integrates with existing clinical workflows and EHR systems beyond high-level descriptions.

From an LLMOps perspective, the case study provides limited visibility into monitoring, observability, model versioning, A/B testing, or how the system handles model updates and performance degradation over time. These are critical considerations for production ML systems but are not addressed in the text.

LLMOps Insights and Patterns

Despite these limitations, the case study does illustrate several valuable LLMOps patterns for production document processing systems. The cost-tiered architecture where cheaper models and services handle high-volume tasks and expensive multimodal reasoning is reserved for cases requiring it represents a practical approach to cost management at scale. The patient-scoped retrieval architecture that scales horizontally without shared indexes addresses a real scalability concern in healthcare applications. The hybrid approach combining specialized OCR services with multimodal foundation models leverages the strengths of each component type.

The emphasis on traceability—ensuring every classification resolves to a specific source page—addresses a fundamental requirement for AI systems in clinical settings where verification against source documents is essential. The chunking strategy that preserves meaningful clinical units rather than splitting on arbitrary boundaries demonstrates attention to domain-specific requirements that impact downstream accuracy.

The progression from Amazon Textract to Amazon Titan embeddings to Amazon Nova 2 Lite to Amazon Nova Pro represents a funnel architecture that progressively narrows the candidate set before applying the most expensive processing, a pattern applicable beyond healthcare document processing.

Conclusion

Guardoc’s implementation demonstrates a production LLMOps architecture for medical document processing that combines RAG, multimodal foundation models, OCR services, and cost-optimized processing pipelines. The technical architecture shows thoughtful consideration of cost, scale, and domain-specific requirements such as patient data partitioning and source traceability. The reported business outcomes suggest meaningful impact on documentation accuracy, compliance, and financial performance, though these should be understood as vendor-reported results rather than independently validated benchmarks.

For practitioners considering similar implementations, the case study offers useful architectural patterns but also highlights the importance of rigorous evaluation on actual document types, careful measurement of recall and precision for specific use cases, and setting confidence thresholds that reflect the real cost of errors in the clinical context. Medical document processing remains a high-stakes application where the consequences of errors extend from patient safety to regulatory compliance to financial performance, and production systems must be architected accordingly.