Aegis Perioperative Copilot

AI-driven clinical decision-support system to assist perioperative teams in identifying anesthesia risks, optimization gaps, airway concerns, and surgery cancellation risk before elective procedure

Comment

AEGIS PERIOPERATIVE COPILOT

Project: Aegis Perioperative Copilot
Hackathon: PromptOpinion Agents Assemble Hackathon
Submitted by: Dr. Bhavna Gupta

The AI copilot that turns unstructured preoperative data into a
structured, evidence-based triage decision — in under 30 seconds.

INSPIRATION

Perioperative assessment is one of the most critical yet consistently
underserved moments in surgical care. Before every operation, an
anesthesiologist must synthesize a patient's laboratory results,
airway anatomy, comorbidity burden, medication risks, and surgical
urgency — often in 15 minutes or less, often without a complete
clinical picture, and always under institutional pressure to proceed.

The consequences of this cognitive overload are real and measurable:

  • Same-day surgical cancellations cost Indian hospitals ₹12,000 to
    ₹40,000 per theatre slot, with preventable causes accounting for
    a significant proportion.

  • Missed hyperkalemia before induction has caused fatal cardiac
    arrhythmias. Unrecognised difficult airways have led to
    cannot-intubate-cannot-oxygenate crises. Uncontrolled HbA1c
    doubles surgical site infection rates.

  • Junior residents and trainees routinely see complex ASA III–IV
    patients in pre-operative clinics with no structured decision
    support and no reliable way to know what they might have missed.

I am an Associate Professor of Anaesthesiology and Critical Care
at AIIMS Rishikesh, with 237+ publications and over a decade of
perioperative clinical experience. I built Aegis because I know
exactly what happens when preoperative assessment is rushed, fragmented,
or incomplete — and I wanted to build something that could actually
prevent it, rather than document it after the fact.

The name Aegis comes from Greek mythology — the shield of protection
carried before battle. Perioperative anaesthesia is exactly that:
anticipation, vigilance, and protection before surgery begins.

WHAT IT DOES

Aegis Perioperative Copilot is a two-agent AI system built on
PromptOpinion, powered by a FastMCP server deployed via NGROK on local server, and
FHIR R4 native from the ground up.

It accepts patient data in three formats:

  • Live EHR patient data via PromptOpinion FHIR context headers
  • Raw FHIR R4 bundles (transaction or collection type)
  • Canonical patient JSON or natural language summary

And produces a fully structured preoperative assessment in seconds.

AGENT 1 — AEGIS SUPERVISOR (Primary Triage Engine)

Computes four validated clinical risk scores:

  • RCRI — Revised Cardiac Risk Index
  • STOP-BANG — OSA screening
  • Apfel Score — Post-operative nausea/vomiting risk
  • ASA Physical Status

Detects 15+ hardcoded clinical safety flags with specific thresholds:

K+ ≥ 6.0 mmol/L → CRITICAL_HYPERKALEMIA → POSTPONE

K+ 5.5–5.9 mmol/L → HYPERKALEMIA → OPTIMIZE

Hb < 7.0 g/dL → SEVERE_ANEMIA → POSTPONE

SBP ≥ 180 mmHg → SEVERE_HYPERTENSION → POSTPONE

HbA1c ≥ 9.5% → UNCONTROLLED_DIABETES → OPTIMIZE

Mallampati 3–4 + OSA + BMI > 40
→ PREDICTED_DIFFICULT_AIRWAY → ESCALATE

Active anticoagulant not held → ESCALATE

Metformin not held day-of-surgery → OPTIMIZE

Retrieves evidence-based guideline recommendations from:

  • ESA 2022
  • ASA 2022
  • DAS 2025
  • AIDAA 2024
  • ADA 2024
  • ASRA 2023

Issues a final recommendation:

PROCEED / OPTIMIZE / ESCALATE / POSTPONE

with clinical reasoning and cancellation risk estimation.

SAFETY REVIEWER

The Safety Reviewer is the architectural innovation that separates
Aegis from other clinical AI systems.

It operates on a ONE-WAY SAFETY RATCHET:

It can ONLY upgrade the Supervisor's recommendation

It can NEVER downgrade it

The Safety Reviewer independently:

  • checks a 21-point clinical checklist
  • performs medication cross-checks
  • counts difficult airway predictors
  • adds mandatory safety addendum

If the Supervisor says:

PROCEED

and the Safety Reviewer finds:

K+ 6.1

the final decision becomes:

POSTPONE

No override. No exception.

WHAT AEGIS HANDLES THAT NO OTHER TOOL DOES

  • Sparse data mode
  • FHIR R4 native support
  • Transaction + collection bundle handling
  • LOINC-coded and text-coded observations
  • valueString parsing
  • ServiceRequest fallback
  • Unit-tested workflow
  • Clinically validated thresholds

HOW I BUILT IT

ARCHITECTURE OVERVIEW

Patient Data (FHIR / JSON / prose)
       ↓
FastMCP Server (Python 3.11, Railway)
       ↓ 6 MCP Tools
PromptOpinion Agent Orchestration
       ↓
Agent 1: Aegis Supervisor
       ↓ assessment JSON
consult prompt
       ↓ final assessment + safety addendum
Structured Markdown Report → Clinical Team

MCP SERVER (Python 3.11 + FastMCP 3.x, deployed on Railway)

1. fetch_patient_from_fhir

Fetches live patient data via PromptOpinion FHIR context headers.

2. assess_current_patient

fetch → normalise → score

Primary entry point for live workflows.

3. normalize_fhir

FHIR R4 Bundle → canonical Aegis patient JSON.

4. compute_risk_scores

Deterministic Python scoring engine for:

  • RCRI
  • STOP-BANG
  • Apfel
  • ASA

5. fetch_guideline

Hardcoded guideline retrieval from authoritative sources.

6. format_aegis_report

Converts assessment JSON into structured markdown report.

FHIR CONTEXT INTEGRATION

The MCP server declares:

ai.promptopinion/fhir-context

during the MCP initialize handshake.

SMART scopes:

patient/Patient.rs
patient/Condition.rs
patient/Observation.rs
patient/MedicationStatement.rs

Optional scopes:

MedicationRequest
Procedure
AllergyIntolerance

KEY ENGINEERING DECISIONS

Deterministic Scoring

Risk scores are computed by Python functions, never by the LLM.

Operator-Precedence Bug Fix

A Python precedence bug in FHIR procedure extraction was identified and fixed.

One-Way Ratchet

Safety Reviewer can only upgrade risk recommendations.

Graceful Degradation

Sparse data mode prevents crashes and unsafe assumptions.

CHALLENGES I FACED

THE FHIR CONTEXT EXTENSION

PromptOpinion's FHIR context support required declaring:

ai.promptopinion/fhir-context

during MCP initialize handshake.

THE FASTMCP IMPORT BUG

RequestContext import pattern changed in FastMCP 3.x.

THE PROCEDURE EXTRACTION BUG

FHIR procedure extraction returned None due to operator precedence.

CLINICAL SCOPE DISCIPLINE

The hardest design decision was what not to build.

WHAT I LEARNED

1. DETERMINISTIC SCORING IS NOT OPTIONAL

The LLM cannot reliably compute RCRI or other risk scores.

2. SAFETY ARCHITECTURE MUST BE STRUCTURAL

Safety guarantees must exist at architectural level.

3. FHIR IS HARDER THAN IT LOOKS

Real-world FHIR bundles are inconsistent and messy.

4. MISSING DATA IS THE CLINICAL NORM

Sparse-data handling is essential for clinical trustworthiness.

5. THE PROMPT IS THE LAST LINE OF DEFENCE

Architecture matters more than prompting.

FUTURE SCOPE

PHASE 1 — IMMEDIATE EXTENSIONS

  • Aegis Consent Intelligence Agent
  • Surgical Scheduling Agent
  • Resident Teaching Mode

PHASE 2 — REAL-TIME HOSPITAL INTEGRATION

  • HIS / OT integration
  • Smart alerts
  • FHIR subscriptions

PHASE 3 — PREDICTIVE INTELLIGENCE

  • Cancellation prediction
  • ICU prediction
  • Difficult airway prediction
  • Postoperative complication prediction

PHASE 4 — NATIONAL PERIOPERATIVE NETWORK

  • Regional benchmarking
  • ABDM / ABHA integration
  • Rural deployment

PHASE 5 — FULL PERIOPERATIVE ECOSYSTEM

  • Intraoperative decision support
  • ICU monitoring
  • Recovery intelligence
  • Pain optimisation
  • Regional language expansion

WHY "AEGIS"?

In Greek mythology, the Aegis symbolized protection and guidance.

We chose the name because perioperative anesthesia is fundamentally about anticipation, vigilance, preparation, and protecting patients before surgery even begins.

CORE LANGUAGE & RUNTIME

Python 3.11

Core language for:

  • clinical scoring logic
  • FHIR normalization
  • flag detection
  • report formatting

Python Standard Library

Used modules:

ast
json
logging
datetime
re
typing
dataclasses
math

MCP FRAMEWORK

FastMCP 3.2.4

MCP framework exposing all 6 clinical tools.

https://github.com/jlowin/fastmcp

MCP Protocol

Underlying protocol standard.

mcp (Python SDK)

Low-level primitives used for FHIR extension declaration.

HTTP & ASYNC

httpx 0.28.1

Async HTTP client for parallel FHIR resource queries.

AI PLATFORM & ORCHESTRATION

PromptOpinion

Healthcare AI orchestration platform.

Claude Sonnet (Anthropic)

Used for:

  • orchestration
  • clinical reasoning synthesis
  • safety addendum generation

A2A Protocol

Agent-to-agent communication protocol.

HEALTHCARE STANDARDS & INTEROPERABILITY

FHIR R4

Healthcare interoperability standard.

LOINC

Supported mappings include:

2823-3 → Potassium
718-7 → Hemoglobin
2160-0 → Creatinine
4548-4 → HbA1c

SMART on FHIR

OAuth-based FHIR access framework.

CLINICAL SCORING SYSTEMS

RCRI — Lee et al. 1999

Perioperative cardiac risk score.

STOP-BANG — Chung et al. 2008

OSA screening tool.

Apfel Score — Apfel et al. 1999

PONV prediction model.

ASA Physical Status Classification

Rule-based ASA estimation.

EVIDENCE-BASED GUIDELINES

  • ESC/ESA Perioperative Guidelines 2022
  • ASA Practice Guidelines 2022
  • DAS Difficult Airway Guidelines
  • AIDAA Perioperative Diabetes Guidelines
  • ADA Standards of Care 2024
  • ASRA / ESAIC Anticoagulation Guidelines
  • KDIGO Renal Guidelines
  • AAGBI Perioperative Guidelines

DEPLOYMENT & INFRASTRUCTURE

Local server and via NGROK

Docker

Containerization .

GitHub

Source control and deployment trigger.

DEVELOPMENT TOOLS

VS Code

Primary development environment.

PowerShell

Used for development and testing.

Python unittest

9 unit tests covering:

  • bundle parsing
  • LOINC mapping
  • sparse data handling
  • procedure extraction
  • operator precedence bug fix

Built With

  • 2025
  • 3.11
  • a2a
  • aidaa
  • apfel
  • classification
  • claude
  • code
  • das
  • docker
  • esa
  • fastmcp
  • fhir
  • git
  • httpx
  • mcp
  • on
  • promptopinion
  • protocol
  • python
  • r4
  • railway
  • rcri
  • score
  • smart
  • sonnet
  • stop-bang
  • vs
Share this project:

Updates