Emergency Handling

Luna includes multi-stage emergency detection to identify crisis situations and provide immediate safety guidance.

Detection Flow

User Message Received
        │
        ▼
┌───────────────────────┐
│ Stage 1: Keyword      │  Check for emergency keywords
│ Detection             │  (chest pain, overdose, suicide, etc.)
└────────┬──────────────┘
         │ (if detected)
         ▼
┌───────────────────────┐
│ Stage 2: Severity     │  Classify as 'immediate' or 'concerning'
│ Classification        │
└────────┬──────────────┘
         │
    ┌────┴────┐
    │         │
    ▼         ▼
Immediate  Concerning
    │         │
    ▼         ▼
┌──────────┐  ┌──────────────────┐
│ Emergency│  │ Patient Context  │
│ Response │  │ Check            │
│ + 911    │  │ (emergency flags)│
└──────────┘  └──────────────────┘

Stage 1: Emergency Detection

The detectEmergency() function scans the user’s message for crisis keywords:

Immediate Emergencies

Keywords that trigger an immediate emergency response:

• “chest pain”, “heart attack”
• “can’t breathe”, “difficulty breathing”
• “overdose”, “took too much”
• “suicidal”, “want to die”, “end my life”
• “allergic reaction”, “anaphylaxis”
• “seizure”
• “severe bleeding”

Concerning Situations

Keywords that warrant context checking but may not be immediate emergencies:

• “side effects getting worse”
• “extreme pain”
• “dizzy”, “fainting”
• “irregular heartbeat”
• “severe headache”

Stage 2: Emergency Response

For Immediate Emergencies

Luna immediately returns a safety-first response:

🚨 This sounds like a medical emergency.

**Call 911 (or your local emergency number) immediately.**

If you are experiencing [specific emergency type]:
1. [First aid instructions specific to the emergency]
2. Do not attempt to [relevant safety guidance]
3. Stay on the line with emergency services

I am an AI health advisor and cannot provide emergency medical care.
Your safety is the priority — please seek immediate professional help.

The response is generated without waiting for AI model inference to minimize latency.

For Concerning Situations

Luna checks the patient context for emergency flags:

const context = await getPatientContext(userId);
if (requiresEmergencyEscalation(context)) {
  // Return error and log escalation
}

If the patient has emergency flags (e.g., known cardiac condition + chest pain mention), the situation is escalated.

Stage 3: Dismissal Detection

If a user dismisses an emergency concern (“I’m fine, it was just a joke”), Luna uses detectEmergencyDismissal() to determine if the dismissal is genuine or if the user may still be at risk.

Escalation Logging

All emergency detections are logged:

logEscalation({
  userId: user.id,
  type: 'emergency_detected',
  message: userMessage,
  severity: 'critical',
  emergencyType: detectedType,
  timestamp: new Date().toISOString(),
});

Escalations trigger:

1. Entry in compliance logs
2. Staff notification (if configured)
3. Conversation flagged for review

Configuration

Emergency detection patterns are defined in the Luna chat route (apps/luna/src/app/api/chat/route.ts). The detection runs before the AI model is invoked to ensure the fastest possible response for genuine emergencies.