KairosMD: Multidisciplinary Ward Round Decision Support

Helping doctors make safer, faster decisions at the bedside.

Comment

KairosMD: Multidisciplinary Ward Round Support System

Inspiration

Acute hospital environments impose high cognitive loads on clinicians. The period required for manual data aggregation from fragmented laboratory, pharmacy, and nursing records contributes to clinician burnout and increases the risk of medical error. KairosMD was developed as a clinical context engine to collapse this information latency, providing a unified, evidence-grounded view of patient data to improve safety and clinical efficiency.

Submission Category

Option 1: Build a Superpower (Powered by MCP)

Try it Out

What it Does

KairosMD is a Multidisciplinary Ward Round Support System (MDS) designed to synthesize complex patient data into high-fidelity briefings.

  • Clinical Synthesis: Aggregates longitudinal vitals, laboratory trends, and medications into concise summaries.
  • Silent Contradiction Detection: Cross-references subjective documentation against objective physiological data to identify deteriorating patients.
  • Evidence Enrichment: Integrates OpenFDA and RxNorm to provide drug-safety alerts grounded in real-world adverse event frequency.
  • Governance: Maintains a chronological audit trail of multidisciplinary team decisions persisted to the clinical record.

Interaction Model

The system operates as a clinical context engine within the PromptOpinion ecosystem. It utilizes SHARP context propagation to receive session-specific patient identifiers, allowing for seamless integration with the clinician's current workspace. The agent orchestrates calls to the KairosMD MCP server to fetch, analyze, and persist data using a deterministic retrieval pipeline that ensures all AI reasoning is grounded in physiological facts.

Technical Implementation

Technical Implementation
Agent Platform PromptOpinion (SHARP Context Propagation)
MCP Framework Python FastMCP (mcp-server-sdk)
Backend Logic Python 3.12, Starlette, Uvicorn
Model NVIDIA Nemotron 3 Super 120B
Performance Disk-backed state persistence (24h TTL)
Data Sync UV Package Manager
Clinical Backend HAPI FHIR R4 (Asynchronous Rate Limiting)
External APIs OpenFDA (FAERS Data Mapping), RxNorm

System Architecture

System Architecture

The KairosMD architecture is a high-throughput Multidisciplinary Ward Round Support System (MDS) that collapses information latency between the clinician and the EHR.

  • Orchestration: Utilizes the SHARP standard to propagate session-specific patient context from the Prompt Opinion Agent to the FastMCP server.
  • Synthesis: The Ward Orchestration Engine collates fragmented FHIR R4 data, running it through deterministic NEWS2 and Conflict Detection modules.
  • Evidence Enrichment: Injects real-world drug safety data from OpenFDA (FAERS) and pharmacological class evidence from RxNorm.
  • Intelligence: Employs NVIDIA Nemotron 3 Super 120B to synthesize raw physiological trends into structured, natural language clinical briefings.
  • Governance: Records MDT decisions via an Audit Persistence Bridge, asynchronously writing signed clinical notes and safety flags back to the FHIR record.
  • Visualization: The MDS Clinical Dashboard provides a real-time visual telemetry stream for the ward census via an SSE interface.

MCP Tool Suite

The KairosMD MCP server exposes nine specialized tools that provide the agent with deep access to clinical intelligence and record persistence:

  1. get_ward_round_summary: Provides a prioritized overview of the entire ward census, including NEWS2 scores, active conflicts, and AI-generated summaries.
  2. get_patient_ward_detail: Executes a deep dive into a specific patient record, returning longitudinal vitals, laboratory trends, and active medications.
  3. get_conflict_report: Aggregates all detected clinical contradictions (e.g., Note vs. Vitals, Allergy vs. Medication) across the ward.
  4. get_discharge_candidates: Evaluates the ward against discharge safety criteria and identifies patients ready for transfer.
  5. record_ward_action: Persists clinical decisions (acknowledgments, notes, escalations) directly back to the FHIR audit trail.
  6. get_action_history: Returns a chronological audit trail of all multidisciplinary decisions made for a patient.
  7. get_drug_safety_info: Queries OpenFDA for real-world evidence, including boxed warnings and FAERS adverse event frequencies.
  8. get_dashboard_access: Generates context-aware deep links to the visual Next.js dashboards.
  9. apply_suggested_plan: A closed-loop tool that allows clinicians to approve and persist AI-generated care plan adjustments.

Clinical Intelligence Layer

The system employs a multi-layered approach to clinical safety:

  • Physiological Deterioration: Automated calculation of NEWS2 (National Early Warning Score) based on validated LOINC codes.
  • Medication Safety: Identification of drug-drug interactions and contraindications, enriched with FAERS adverse event report counts to provide clinical weight.
  • Discharge Validation: Automated assessment of patients against objective clinical criteria (e.g., apyrexial for 24 hours, stable NEWS2) to optimize bed management.

Technical Challenges and Solutions

  • Information Latency: Complex clinical analysis involving 20+ FHIR resources per patient can exceed agent timeouts. We solved this with a disk-backed persistence layer and a background cache warmer that ensures sub-second response times for all tools.
  • Provider Stability: Standard public FHIR servers are prone to 429 rate-limiting. We implemented a sequential request pipeline with jittered delays to maintain connection stability without compromising data depth.
  • Deterministic Grounding: To prevent hallucination, we separate clinical data retrieval from AI reasoning. The LLM only receives slimmed, validated projections of the patient record, minimizing token bloat and ensuring high-fidelity summarization.

Accomplishments

  • Closed-Loop Governance: Successfully implemented a system where clinical decisions made via the agent are persisted back to the FHIR audit trail.
  • High-Fidelity Simulation: Developed a clinical seeder capable of generating 20 distinct, evidence-based inpatient scenarios to validate the system under realistic ward loads.

Lessons Learned

  • MCP Utility: The Model Context Protocol is an effective framework for securely bringing complex, private clinical data into generative AI workflows.
  • Evidence-Based Reasoning: Grounding output in physiological trends and FDA evidence enhances utility compared to generic summarization.

Future Development

  • Diagnostic Integration: Expansion to include diagnostic imaging summaries and pathology reports.
  • Specialty Support: Development of specialized reasoning modules for intensive care (ICU) and cardiology environments.
  • Predictive Analytics: Implementation of length-of-stay and readmission risk modeling.

Built With

Share this project:

Updates