Back to Portfolio

Case Study: Medical Triage Agent

Autonomous AI agent for scalable, medically-precise triage system

Challenges

Healthcare providers and hospitals constantly face the challenge of managing high volumes of patient inquiries efficiently. Patients often experience long wait times for initial assessment, while clinical staff spend significant time on routine triage tasks.

Key Challenges

  • High Volume: Managing large numbers of patient inquiries with limited staff
  • Accuracy Requirements: Ensuring medically precise assessments with zero tolerance for errors
  • Scalability: Creating a solution that could handle fluctuating demand
  • Integration: Connecting with existing healthcare systems and EHR standards

Solution & Architecture

The agent was engineered to mimic a clinician's stepwise reasoning process, combining multiple AI technologies for comprehensive patient assessment.

Medical Triage Agent Architecture

Architecture diagram showing the multi-component medical triage system

Key Components

  1. Symptom Intake & Triage

    Dialogflow CX initiates structured, multi-turn conversations to collect symptoms, vitals, and history. Healthcare Natural Language AI maps free-text symptoms to SNOMED-CT / ICD-10 codes for precision.

  2. Diagnosis Support

    Gemini 1.5 Pro applies chain-of-thought prompting to rank urgency, while Med-PaLM 2 augments with specialized medical knowledge for differential diagnoses.

  3. Treatment Planning

    Vertex AI Agent Builder orchestrates guideline retrieval, conflict detection, and treatment suggestions in a cohesive flow with evidence-based recommendations.

  4. Scalability & Safety

    GPU-accelerated A2 VMs on Google Kubernetes Engine handle inference bursts, while Cloud Run manages stateless components with zero-downtime deployments.

Implementation Highlights

Multi-Model Reasoning

Combined Gemini 1.5 Pro's reasoning capabilities with Med-PaLM 2's medical expertise for clinically sound assessments, with chain-of-thought prompting for transparent decision-making.

Real-Time Medical Coding

Automated mapping of free-text symptoms to standardized SNOMED-CT and ICD-10 codes, ensuring interoperability with existing healthcare systems and EHR platforms.

Clinical Scenario

Real-World Use Case: Chest Pain Assessment

  • Patient Presentation: Sharp chest pain, shortness of breath on exertion, and fatigue for five days
  • Agent Assessment: The agent asks clarifying questions about onset, radiation, and medication history, then maps symptoms to SNOMED codes
  • Clinical Reasoning: Chain-of-thought reasoning ranks "acute coronary syndrome" highest and retrieves latest ACC/AHA guidelines
  • Outcome: Treatment plan includes immediate ECG & troponin order, discontinuation of Lisinopril pending physician approval, and cardiology referral

Results & Impact

Quantitative Results

  • 42% Faster Triage: Reduced average triage time from 12 minutes to 7 minutes
  • 96% Accuracy: Concordance with senior clinician decisions on 5,000-case test set
  • High Scalability: Sustains 1,100 concurrent user sessions with <1.2s median latency
  • Full Interoperability: 100% of outputs delivered in SNOMED-CT/ICD-10 for EHR ingestion

Qualitative Benefits

  • Reduced Burden: Significantly decreased workload for clinical staff
  • Improved Access: 24/7 availability for initial patient assessments
  • Enhanced Consistency: Standardized triage process across all patients
  • Better Documentation: Automated medical coding and record keeping

Technology Stack

Gemini 1.5 Pro Med-PaLM 2 Vertex AI Dialogflow CX Google Kubernetes Engine Cloud Run SNOMED-CT ICD-10 Healthcare NLP API Python TensorFlow