Agentic AI for Clinical Urgency Mapping and Queue Optimization in High-Volume Outpatient Departments: A Simulation-Based Evaluation

1. Problem Statement

The paper addresses the critical issue of overcrowding and inefficient patient flow in Indian public hospital Outpatient Departments (OPDs).

• Context: Public hospitals in tier-2 cities (e.g., Jabalpur, Madhya Pradesh) handle 200–8,000 patients daily with limited specialist physicians.
• Current Limitations: The prevailing First-Come-First-Served (FCFS) token system treats all patients equally, ignoring clinical urgency. This leads to dangerous delays for critical cases (e.g., chest pain) while routine cases (e.g., prescription refills) consume the same resources.
• Gaps in Existing Solutions:
  • Static Rule-Based Triage: Categorizes patients at registration but fails to detect "urgency drift" (worsening conditions while waiting) or utilize longitudinal patient history.
  • Data Fragmentation: Patient records are often paper-based and unavailable, leading to a "clinical vacuum" where hidden risks (e.g., a mild headache in a patient with a history of TIA) are missed.
  • Multilingual Barriers: Existing systems struggle with diverse local languages (Hindi, Bundeli, etc.).

2. Methodology

The authors propose and evaluate an Agentic AI Framework designed to dynamically map clinical urgency and optimize queues.

A. System Architecture (Six Components)

The framework operates as a unified decision-making system with six coordinated agents:

1. Voice-Based Symptom Capture (Modeled): A multilingual speech-to-text module (Hindi, Bundeli, English) to reduce registration time by 40% and structure symptom data.
2. LLM-Powered Severity Prediction: Uses Anthropic's Claude Sonnet to analyze structured symptoms, vitals, and demographics. It outputs a structured JSON containing:
   • Urgency classification (Critical, High, Medium, Low) aligned with WHO IITT.
   • Acuity scoring (1–10).
   • Clinical reasoning and red-flag detection.
3. Load-Aware Physician Assignment: A weighted scoring engine (50% Specialty Match, 30% Load Balance, 20% Availability) to assign patients to the most appropriate available specialist.
4. Adaptive Queue Optimization: Continuously monitors the queue every 5 minutes. It detects urgency drift (stochastic deterioration probabilities) and recalculates priorities to prevent "starvation" of low-priority patients while escalating worsening cases.
5. Multi-Objective Orchestrator: A decision agent that calculates a final priority score using a weighted function of Urgency, Acuity, Wait Time, and Load Score.
6. Patient Memory System: A persistent, searchable longitudinal health record. It cross-references current symptoms against historical data (e.g., diabetes, COPD, pregnancy history) to detect "hidden urgency" that face-value assessment misses.

B. Simulation Design

• Environment: Discrete-event simulation of a District Hospital in Jabalpur, India.
• Dataset: 368 synthetic patients per session (30 independent runs) modeled on real demographic and epidemiological data (NFHS-5, IHME GBD).
• Parameters: Non-homogeneous Poisson arrival patterns, 6 specialist physicians, and consultation times varying by urgency.
• Baselines: Compared against FCFS (Token-based) and Rule-Based Triage (Static priority).
• LLM Integration: The system uses real API calls (simulated for the bulk of the study) with a cost of ~$1.58 per session.

3. Key Contributions

1. Agentic AI Architecture: A novel six-component framework integrating voice capture, LLM reasoning, dynamic queue optimization, and longitudinal memory.
2. Real LLM Integration: A working implementation using Claude Sonnet with structured JSON output for deterministic, reproducible triage decisions.
3. Dynamic Urgency Recomposition: The system does not just reorder patients; it actively changes the composition of the queue by detecting hidden risks and worsening conditions.
4. Patient Memory System: Demonstrates the critical value of integrating longitudinal history to identify "hidden-critical" cases (e.g., a patient with mild symptoms but a history of immunosuppression).
5. Comprehensive Evaluation: A rigorous 30-run simulation comparing FCFS, Rule-Based, and Agentic AI approaches with ablation studies.

4. Key Results

The simulation results (averaged over 30 runs) highlight significant improvements in patient safety and queue dynamics:

• Critical Patient Safety:

  • FCFS: Only 30.8% of critical patients seen within 10 minutes.
  • Agentic AI: 94.2% of critical patients seen within 10 minutes (avg wait 7.5 min).
  • Hidden Critical Detection: The system identified ~11.9 additional critical patients per session (a 92% increase) by leveraging patient history, which FCFS and Rule-Based systems missed entirely.

• Dynamic Queue Recomposition:

  • The Agentic AI transformed the queue distribution from a static 13 Critical / 161 Low to a dynamic 25 Critical / 50 Low.
  • It detected ~236 urgency drift events per session (patients worsening while waiting) and escalated ~111 patients from Low to higher urgency levels per session.

• Throughput & Efficiency:

  • Throughput: Maintained comparable throughput to FCFS (~40.4 patients/hour), proving that reordering by urgency does not reduce overall capacity.
  • Cost: LLM processing cost was negligible (~$1.58 per 368-patient session).

• Trade-offs:

  • Wait Times: While critical patients waited significantly less, non-critical (Low/Medium) patients experienced longer waits (e.g., Low urgency wait increased from 33.8 min in FCFS to 142.4 min in Agentic AI). The authors argue this is a necessary trade-off to prevent misclassification of life-threatening conditions, though it highlights a capacity bottleneck for "High" urgency patients.

• Ablation Study:

  • Removing the Patient Memory System reduced critical detection by ~92% and dropped the number of detected critical patients from 24.9 to 13.0 per session.
  • Removing Drift Detection eliminated the ability to catch worsening conditions during the wait.

5. Significance and Implications

• Paradigm Shift: Moves OPD management from "static prioritization" to "dynamic clinical state tracking." It reveals that a significant portion of the patient population (approx. 30% in the simulation) carries hidden risks invisible to point-of-care-only assessments.
• Safety vs. Fairness: The framework challenges the traditional "fairness" of FCFS, arguing that true fairness in healthcare requires prioritizing clinical need over arrival time. It exposes that current staffing models are inadequate for the true urgency burden of the population.
• Scalability & Governance: The system is designed with explainability (structured JSON reasoning), human-in-the-loop verification (mandatory for Critical/High cases), and compliance with India's Digital Personal Data Protection Act (DPDPA).
• Future Directions: The authors plan to validate the system with real-world data at NSCB Medical College, Jabalpur, and integrate with India's ABHA (Ayushman Bharat Health Account) for nationwide longitudinal records.

Conclusion: The paper demonstrates that an Agentic AI framework can significantly improve patient safety in high-volume, resource-constrained OPDs by dynamically identifying hidden critical cases and managing urgency drift, provided that capacity planning is adjusted to handle the revealed true volume of high-acuity patients.