Real-Time AI Assistant for Clinical Decision Support

Overview

InnovateCo operates a network of outpatient clinics where physicians manage complex patient cases across multiple specialties. Clinicians were spending significant time navigating between electronic health record (EHR) systems, searching medical literature, and cross-referencing treatment protocols -- tasks that fragmented their attention during patient encounters.

The clinical team needed a tool that could serve as an intelligent assistant during consultations: answering questions about patient history, surfacing relevant clinical research, and suggesting applicable treatment protocols without requiring the clinician to leave their primary workflow.

We designed and built a real-time AI assistant powered by an agentic architecture, giving clinicians a conversational interface that orchestrates multiple data sources and tools behind the scenes.

Approach

The project began with an in-depth compliance review. Healthcare AI systems operate under strict regulatory requirements, and any tool interacting with patient data must meet HIPAA security standards and maintain comprehensive audit logging. We worked closely with InnovateCo's compliance team to establish the security architecture before writing any application code.

Our agent architecture uses a supervisor pattern where a central reasoning agent coordinates specialized tool agents. When a clinician asks a question, the supervisor determines which tools are needed, dispatches parallel requests where possible, and synthesizes the results into a coherent response.

Three primary tool agents handle the core functionality. The patient data agent connects to the EHR system via FHIR APIs, retrieving relevant medical history, lab results, medications, and encounter notes. The research agent queries a curated vector database of clinical literature and treatment guidelines. The protocol agent matches patient conditions against institutional care protocols and formulary data.

A key technical challenge was achieving sub-three-second response times. Clinical workflows demand near-instant responses; any noticeable delay disrupts the consultation flow. We optimized through aggressive caching of frequently accessed patient summaries, parallel tool execution, and streaming responses so clinicians see results progressively rather than waiting for complete processing.

Technical Details

The assistant is built as a Next.js application embedded within InnovateCo's existing clinical portal. The frontend provides a chat-style interface that clinicians can open alongside their EHR view. WebSocket connections maintain real-time communication, and responses stream token-by-token for perceived responsiveness.

The agent orchestration layer uses the Claude API for reasoning and tool dispatch. Each tool agent is implemented as a typed function with strict input and output schemas, ensuring predictable behavior and comprehensive error handling. The supervisor agent sees tool descriptions and decides at each step which tools to invoke, supporting multi-step reasoning chains for complex clinical queries.

Patient data retrieval uses FHIR R4 APIs with scoped OAuth2 tokens. Each clinician's session carries their access credentials, ensuring the assistant can only access data the clinician is authorized to view. All queries and responses are logged with full audit trails including timestamps, data accessed, and the reasoning chain that led to each tool invocation.

The vector database stores embeddings of clinical guidelines, drug interaction databases, and peer-reviewed research. We implemented a retrieval-augmented generation pipeline that combines semantic search with metadata filtering, allowing queries scoped by specialty, date range, and evidence level.

Results

The pilot deployment with 15 clinicians ran for six weeks before expanding to the full clinical network. Adoption was strong from the outset, with 92% of pilot clinicians using the assistant regularly within the first two weeks.

Clinical workflow efficiency improved by 40%, measured as the reduction in time spent switching between applications and manually searching for information during patient encounters. Clinicians reported that the assistant was particularly valuable for complex cases involving multiple comorbidities, where cross-referencing patient history with current treatment guidelines would previously require significant manual effort.

Response times consistently met the sub-three-second target, with median response time at 1.8 seconds for patient data queries and 2.4 seconds for research lookups. The streaming interface meant clinicians perceived near-instant responses for the initial content, with additional detail appearing progressively.

The system now serves as the foundation for InnovateCo's broader clinical AI strategy, with planned extensions for automated pre-visit summaries and post-encounter documentation assistance.