Link

Inspiration

Every year, millions of patients leave the hospital and enter one of the most vulnerable periods in healthcare: the transition from hospital to home. During this handoff, critical information can be delayed, overlooked, or fragmented across systems.

Research has shown that many patients are discharged before all laboratory and diagnostic results are finalized. In one landmark study, 41% of discharged patients had test results return after leaving the hospital, and some of those results required urgent medical follow-up. Physicians were often unaware those results had returned.

Medication transitions are another major safety challenge. Patients frequently leave the hospital with medication discrepancies between inpatient prescriptions, outpatient records, and discharge instructions. These inconsistencies are a well-documented contributor to preventable adverse events and avoidable readmissions.

At the same time, healthcare systems are under increasing pressure to improve care coordination and reduce avoidable readmissions through programs such as the CMS Hospital Readmissions Reduction Program.

We built Link because the discharge process should not end when a patient walks out the door. We wanted to create a digital safety net that helps hospitals, physicians, and care teams stay connected to critical clinical information during the high-risk post-discharge window.

What it does

Link is an interoperable AI safety net that automates the "Last Mile" of patient discharge. It uses a specialized architecture to monitor, reconcile, and close the loop on patient care.

• The Digital Sentry (MCP): Link deploys an MCP (Model Context Protocol) Server that "stays behind" at the hospital. It continuously polls the FHIR server for any tests that were "Pending" at discharge. The moment a result turns "Final," Link’s intelligence analyzes it for clinical urgency.
• The Safety Reconciliation: Link automatically compares the new hospital medications with the patient's existing home records (via FHIR MedicationRequest), flagging dangerous overlaps or omissions instantly.
• Contextual Intelligence (SHARP): Link uses the SHARP (Secure Healthcare Agent Resource Propagation) extension pattern to securely propagate patient IDs and FHIR tokens. It ensures that the family doctor receives a "Smart Brief" instead of a 20-page PDF, highlighting only the "Red Flags" and required actions.
• Closed-Loop Action: Instead of a passive notification, Link generates clear, clinically-grounded summaries that lead to real-world medical action, ensuring "ghost results" are finally addressed.

How we built it

We built Link with a focus on interoperability, security, and statelessness:

• Standardized Interoperability: Built on HL7 FHIR R4, ensuring compatibility with major EHR providers like Epic, Cerner, and Google Cloud Healthcare API.
• Model Context Protocol (MCP): Leveraged FastMCP to create a secure, discovery-based toolset that agents can use to "see" into the hospital's data lake.
• SHARP Extension Pattern: Implemented a custom capability patching system in the MCP server to support dynamic FHIR context injection (URLs and Bearer tokens) via metadata.
• Python Stack: Used a modular Python architecture for the MCP server, with specialized tools for DiagnosticReport, MedicationRequest, and Condition resources.

Challenges we ran into

• FHIR Complexity: Mapping nested and often inconsistent FHIR JSON resources into clean, concise snippets that an LLM can reason over without hitting token limits.
• Stateless Security: Implementing the SHARP context propagation was challenging. We had to ensure that the MCP server never "stored" credentials but could still securely authenticate against the FHIR server for every request.
• Closing the Loop: Designing the logic to differentiate between a "Normal" result and a "Red Flag" result that requires immediate PCP intervention.

Accomplishments that we're proud of

• True Interoperability: Successfully querying live FHIR sandboxes and returning structured clinical intelligence.
• Zero-Trust Architecture: Building a system where the AI never sees patient data unless the secure FHIR context is explicitly provided by the host platform.

What we learned

• Standardization is Key: MCP is the "missing link" for enterprise AI. It allows us to build medical tools once and deploy them across any AI platform.
• The Power of Small Tools: We learned that specific, atomic tools (like GetPendingLabs) are much more effective for clinical safety than general-purpose "EHR search" tools.
• Agent Orchestration: The transition from hospital to clinic isn't just a data transfer—it's a human handoff that requires an AI to act as a "Transition Specialist."

What's next for Link

• SMART on FHIR Launch: Integrating Link directly into the clinician's workflow inside the EHR dashboard.
• Predictive Risk Scoring: Using historical FHIR data to predict which patients are at the highest risk of readmission before they even leave the hospital.
• Patient-Facing Instructions: Generating simple, multilingual discharge summaries that are automatically pushed to the patient's mobile device via secure channels.

Built With

• fastmcp
• python