Patient Triage

Patient Triage Skill

Perform systematic clinical triage with severity classification, confidence scoring, and intelligent escalation recommendations. Designed for 24/7 patient intake workflows.

Overview

This skill enables clinical AI agents to:

• Classify patient presentations by urgency (Emergency, Urgent, Routine)
• Route to appropriate clinical modules (Cardiology, Mental Health, Diabetes, etc.)
• Generate confidence scores with proper calibration
• Trigger escalation based on safety thresholds
• Support multilingual input (English, Hindi, code-switching)

Urgency Classification

Emergency (Immediate action required)

Criteria:

• Life-threatening symptoms
• Altered consciousness
• Severe chest pain or difficulty breathing
• Signs of stroke (FAST criteria)
• Active suicidal ideation with plan
• Severe allergic reaction

Response: Immediate escalation to emergency services or on-call physician.

Latency Requirement: < 100ms detection time

Urgent (Same-day attention needed)

Criteria:

• Moderate symptom severity
• Potential for deterioration
• Medication concerns requiring timely review
• Mental health crisis without immediate danger
• Abnormal vital signs

Response: Priority queue for clinical review, same-day callback.

Routine (Standard scheduling)

Criteria:

• Stable chronic conditions
• Medication refills
• Follow-up appointments
• General health questions
• Non-urgent administrative requests

Response: Standard appointment scheduling or information provision.

Clinical Module Routing

Based on symptom analysis, route to appropriate specialty modules:

| Module | Trigger Keywords | Example Presentations | |--------|------------------|----------------------| | Cardiology | chest pain, palpitations, shortness of breath | Heart failure follow-up, AFib symptoms | | Mental Health | anxiety, depression, mood, stress, sleep | Depression screening, anxiety management | | Diabetes | blood sugar, glucose, insulin, A1c | Hypoglycemia concerns, medication adjustment | | Respiratory | cough, breathing, asthma, COPD | Asthma exacerbation, COPD management | | GI | stomach, nausea, vomiting, bowel | Gastritis symptoms, constipation | | MSK | pain, joint, back, muscle | Chronic low back pain, arthritis | | Pediatric | child, infant, baby | Fever in child, immunization questions | | Geriatric | elderly, falls, memory | Fall prevention, cognitive concerns |

Confidence Scoring

All triage decisions include calibrated confidence scores:

Confidence Level:
- HIGH (>85%): Clear presentation, unambiguous classification
- MEDIUM (60-85%): Some uncertainty, may benefit from follow-up questions
- LOW (<60%): Ambiguous presentation, recommend human review

Calibration Requirements

A well-calibrated system should:

• Be correct 70% of the time when reporting 70% confidence
• Be correct 90% of the time when reporting 90% confidence

Expected Calibration Error (ECE) should be < 0.1

Escalation Triggers

Automatic Escalation

Escalate immediately when:

• Emergency classification detected
• Patient expresses suicidal/homicidal ideation
• Confidence score < 40%
• Conflicting symptom signals
• Patient requests human assistance

Threshold-Based Escalation

Escalate when cumulative risk exceeds thresholds:

• 3+ urgent symptoms in single conversation
• Sentiment volatility score > 0.8
• Non-response to 3+ consecutive messages
• Repeated expression of dissatisfaction

Multilingual Support

Supported Languages

• English - Primary language
• Hindi - Full support with medical terminology
• Code-switching - Hindi-English mixed input

Language Detection

Automatically detect input language and respond appropriately:

Input: "Mujhe chest mein bahut dard ho raha hai"
Detection: Hindi
Classification: EMERGENCY - Chest pain
Response Language: Hindi with English medical terms

Medical Terminology Mapping

Common symptoms across languages:

| English | Hindi | Transliteration | |---------|-------|-----------------| | Chest pain | छाती में दर्द | chaati mein dard | | Difficulty breathing | सांस लेने में तकलीफ | saans lene mein takleef | | Headache | सिर दर्द | sir dard | | Fever | बुखार | bukhaar | | Nausea | मिचली | michli |

Safety-First Design

Asymmetric Error Weighting

Missing an emergency is 10x worse than a false alarm:

| Actual | Predicted | Weight | Explanation | |--------|-----------|--------|-------------| | Emergency | Routine | 10x | Critical miss | | Emergency | Urgent | 2x | Downgrade | | Urgent | Routine | 3x | Significant miss | | Urgent | Emergency | 0.5x | Acceptable overreaction | | Routine | Emergency | 1x | False alarm | | Routine | Urgent | 0.5x | Minor overreaction |

Fail-Safe Behavior

When uncertain:

1. Default to higher urgency classification
2. Recommend human review
3. Provide clear reasoning for classification
4. Never dismiss concerning symptoms

Usage Instructions

1. Collect Initial Presentation

Gather:

• Chief complaint in patient's own words
• Duration and onset of symptoms
• Associated symptoms
• Relevant medical history (if available)

2. Analyze Symptoms

The agent will:

• Parse symptom description for clinical indicators
• Check for emergency keywords and patterns
• Assess severity based on presentation
• Consider comorbidity implications

3. Generate Classification

Output includes:

• Urgency level (Emergency/Urgent/Routine)
• Clinical module routing
• Confidence score
• Reasoning summary
• Recommended next steps

4. Handle Escalation

If escalation triggered:

• Notify appropriate clinical staff
• Provide handoff summary
• Document escalation reason
• Track SLA compliance

Example Output

TRIAGE ASSESSMENT

Patient Input: "I've been having crushing chest pain for the past 30 minutes.
It's radiating to my left arm and I'm sweating."

CLASSIFICATION: EMERGENCY
Confidence: 95% (HIGH)

Clinical Indicators:
- Crushing chest pain (cardiac concern)
- Duration >15 minutes (not transient)
- Left arm radiation (classic ACS pattern)
- Diaphoresis (autonomic response)

Routing: Cardiology (Emergency)

IMMEDIATE ACTION REQUIRED:
⚠️ Possible acute coronary syndrome
⚠️ Escalating to emergency services
⚠️ Advise patient: Call 108 (emergency) immediately

Reasoning: Multiple classic indicators of acute myocardial infarction.
High-confidence emergency classification. Immediate medical attention required.

Integration with TherapyPod

This skill integrates with:

• Medical Safety Engine - Primary safety classification
• Escalation Rules Engine - Handoff decision logic
• HubAgent - Conversation handling
• ClinEval Benchmark - Safety dimension validation

Performance Metrics

Track these metrics for continuous improvement:

• Emergency detection sensitivity (target: >99%)
• False positive rate for emergencies (target: <15%)
• Latency compliance (target: >95% under 100ms)
• Escalation appropriateness (target: >90%)

References

• See references/emergency-protocols.md for emergency criteria
• See references/symptom-patterns.md for classification patterns
• See references/escalation-sla.md for timing requirements