Skip to main content

Scaling AI Thoughtfully: Behind the Scenes of the HealthEdge® GuidingCare Notes Summarizer

Part 2: Technical Implementation

This is part two of a two-part blog series on HealthEdge’s AI Summarizer feature; this post focuses specifically on the feature’s technical implementation. Read Part One here for an overview of the business challenges and operational benefits.

Addressing Information Overload in Care Management

Healthcare organizations face a growing challenge as healthcare professionals struggle with information overload from extensive documentation. Our AI team recently developed the AI Summarizer feature for our GuidingCare care management platform to address this problem. This feature simplifies a care manager’s workflows by analyzing care management notes to generate a concise summary and a list of key action items.

The AI Summarization functionality can reduce care manager review time, translating conversational content from notes into specific tasks and insights.

A Platform-First Architecture

HealthEdge’s approach to AI implementation centers on building reusable platform capabilities rather than one-off solutions. Our AI team created a summarizer API within our broader GenAI platform, which manages the complex technical details that would otherwise burden individual product teams. These include decisions around model choice and temperature settings, as well as comprehensive logging, observability, and security protocols.

The platform also collects user feedback, establishing a foundation for continuous improvement and future model evaluation.

This centralized infrastructure management allows product teams like GuidingCare to focus on what matters most for their specific use cases. Each use case defines its own summarizer configuration, including the input data schema, custom prompts, and desired output formats. This division of responsibilities enables each team to implement AI features as needed without creating bottlenecks. The result is faster deployment cycles and more targeted solutions that address specific business needs.

Embedding AI into GuidingCare Workflows

For GuidingCare, we embedded the AI summarization functionality directly into existing care management workflows through a simple “Summarize” button in the Member Notes interface. The technical architecture follows a secure data flow between components, including the GuidingCare UI, the GuidingCare Cloud API, the WSO2 gateway that connects GuidingCare to external services, and the AI Summarizer. Each layer manages specific responsibilities for data processing and security.

Behind the scenes, the summarization API follows a streamlined processing pipeline that ensures both reliability and traceability. When a summarization request is received, the system first loads and validates the configuration. The prompt template is then dynamically formatted using runtime parameters, the user-defined input data schema, and any optional prompt addenda before invoking Azure AI's chat API. Each operation generates comprehensive logs that capture both the formatted prompt and the resulting output, providing full audit traceability before returning the processed response to the requesting application.

One of the most impactful parts of this project was working closely with our business stakeholders to craft a carefully engineered prompt template. Through direct feedback from clinical teams, we learned that care managers primarily read extensive notes to identify their next actions. This insight drove the design of the prompt to not just summarize notes, but specifically extract actionable follow-ups, a distinction that makes the difference between a generic summary tool and one that directly addresses workflow pain points.

The prompt template uses a structured approach, instructing the AI to analyze clinical notes in JSON format and produce summaries that highlight key diagnoses, reasons for visits, and follow-ups or recommendations, along with a separate, clear list of actionable items. The system processes notes using a schema that includes note type, health notes content, care staff information, and timestamps. To balance comprehensiveness with performance, the summarization feature uses notes from the last 90 days of history. The 90-day lookback limit was established based on both performance and clinical factors—limiting the number of notes improved system efficiency while simplifying the summarization process for the model, reducing the risk of overlooking critical information or producing inaccurate content.

AI Safety and Quality Assurance

As with all AI systems, the inherent risks of inaccurate, incomplete, or unsafe outputs required careful consideration. Our implementation addresses these concerns through multiple layers of protection. The system architecture ensures that AI-generated content is exclusively derived from existing notes within the system, guaranteeing that the outputs are as reliable and secure as the input data. This design principle prevents the AI from introducing external information or making claims beyond the information in the clinical record.

The platform provides comprehensive logging and audit trails for each summarization, enabling ongoing evaluation, monitoring, and debugging. The feature incorporates responsible AI protocols, including explicit AI-generated content disclaimers and human validation, ensuring clinical judgment remains central to care decisions. We also subjected the system to rigorous testing that included both expected inputs and adversarial scenarios, helping to ensure reliability across various input conditions that might occur in real-world healthcare environments.

Security and Compliance

The implementation maintains enterprise-grade security through multiple layers of protection. Data transmission utilizes secure APIs with HIPAA compliance throughout the data flow, while authentication relies on OAuth 2.0 for secure API access. The underlying LLM services are hosted through Azure, leveraging Microsoft’s enterprise security infrastructure and compliance certifications, with all data remaining within HealthEdge’s controlled infrastructure.

Role-based access controls limit functionality to authorized users through a layered permission system that requires both appropriate role permissions and access to the underlying clinical data. This controlled rollout approach ensures proper governance while the feature undergoes early adoption validation.

What’s Next for AI in Care Management?

With intentional platform design, strategic prompt engineering, and robust risk mitigation, we have demonstrated how healthcare technology teams can effectively integrate AI to enhance clinical workflows—and this is just the beginning. Next, we are expanding the summarizer’s capabilities to include interactive features that will allow care managers to ask follow-up questions and engage in a conversational dialogue about member information. This evolution will transform the summarizer from a one-time analysis tool into a dynamic, intelligent assistant that provides deeper insights and even greater operational benefits.

To explore how HealthEdge is shaping the future of care with responsible, practical AI, visit Artificial Intelligence | HealthEdge.

About the Author

Alice Zhan is a principal machine learning engineer at HealthEdge. She holds undergraduate and master's degrees from MIT, where her research focused on machine learning in healthcare. With a passion for leveraging technology to improve patient outcomes, she now has over 7 years of industry experience building ML products for healthcare. Outside work, she enjoys learning new (human) languages and trying new recipes in the kitchen.