Deconstructing Automated Variant Interpretation: A Multi-Modal Ablation Study Comparing LLM Reasoning, Adversarial Agents, and Gradient Boosting in Inherited Cardiac Conditions

This study will employ a six-condition ablation design to evaluate the diagnostic accuracy, safety characteristics, and underlying cognitive mechanisms of three AI architectures in cardiogenetics: (1) Single-Agent Large Language Models (LLMs), (2) Adversarial Multi-Agent LLMs employing Critic–Arbitrator loops (GPT-5.1), and (3) a feature-based machine learning model (CatBoost). Each architecture will be evaluated under two information states—Baseline (ClinVar-blinded) and Enriched (ClinVar-available)—to isolate the effects of deep phenotypic reasoning versus external knowledge retrieval. This factorial framework is designed to test whether advanced reasoning architectures will yield measurable advantages over standard prompting and statistical pattern recognition, and to quantify the degree to which these systems will operate as independent diagnostic reasoners versus consensus-driven decision engines.