“Ai-Orchestrated Workflow For Refractive Surgery And Keratoconus Diagnosis: A Three Ai-Agents System For Comprehensive Patient Assessment”

• Accuracy in keratoconus diagnosis and refractive surgery depends on a systematic, Al-driven workflow to provide precise risk stratification and individualized treatment planning.
• The study presents an LLM-coordinated workflow that combines multimodal imaging (Pentacam, AS-OCT, epithelial mapping) with systematic patient history evaluation. Through the utilization of novel Al-derived metrics, the platformaugments diagnostic certainty, optimizes decision-making efficiency, and ultimate treatment option optimization.
• The ultimate goal is to optimize patient safety, surgical planning, and long-term outcomes witha systematic data-driven strategy.

This research assesses a three-agent Al-assisted diagnostic workflow for keratoconus evaluation and refractive surgery planning.

The system consists of: (1) General Ophthalmology Agent for taking history, (2) Pentacam & AS-OCT Agent for interpretation of corneal imaging, and (3).

Anterior Segment Specialist Agent for final diagnosis and treatment planning.

Various LLMs are compared in accuracy, decision influence, and expert agreement on real ophthalmic cases

A three-agent Al-assisted workflow is tested to evaluate patients for refractory surgery and diagnose keratoconus.

Each agent-General Ophthalmology, Pentacam & AS-OCT, and Anterior Segment Specialist-is evaluated using various Large Language Models (LLMs).

The performance of models is benchmarked on accuracy, diagnostic concordance, and decision alignment with specialist ophthalmologists.

Actual patient cases are run through the workflow, contrasting LLM outputs with latest guidelines and standard clinical evaluations.

Outcome metrics are diagnostic accuracy, response variation, and Al effectiveness in informing surgical decisions.

The workflow was tested on 50 keratoconus and refractive surgery cases using several LLMs on 3 agents.
General Ophthalmology Agent:
Accuracy of patient history: 84% (LLM range: ±6%)
Risk factor consistency: 78% (Some misunderstood steroid-induced IOP risk)
Pentacam & AS-OCT Agent:
Imaging accuracy: 87% (LLM variability: ±5%)
Detection of keratoconus: 92% (Enhanced in structured models)
PTA accuracy: 81% (Borderline values resulted in variation)
Anterior Segment Specialist Agent:
Agreement on diagnosis: 89% (LLM variability: ±7%) Surgery recommendation: 76% (Variability in borderline LASIK cases)Total: 85% accuracy, LLM variation: ±9%, best time: 11.2s, worst time: 22.4s.

Results emphasize LLM variation, indicating varying models for different agents.

• This research effectively deployed an Al-based ophthalmology workflow, maximizing patient history extraction, corneal image analysis, and surgical decision.
• Cross-testing three specialist agents with multiple LLMs, the process resulted in 85% diagnostic accuracy, accurate keratoconus detection (92%), and high concordance (89%) with expert ratings.
• Findings illustrated task-specific LLM optimization boosts Al precision, minimizing inconsistency in borderline LASIK referrals.
• This achievement establishes Al as a valid ophthalmology application, opening the door to LLM-based precision diagnostics and Al-guided surgical planning.