Skip to content

Echo-CoPilot: A Multiple-Perspective Agentic Framework for Reliable Echocardiography Interpretation

🕒 Published (v1): 2025-12-06 23:27 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Echo-CoPilot is a multi-agent agentic framework for echocardiography interpretation that runs three parallel ReAct agents (structural, pathological, quantitative) and reconciles their outputs via a self-contrast LLM guided by EchoKG, a clinical knowledge graph encoding ASE/EACVI guideline constraints. It achieves 54.0% accuracy on MIMICEchoQA, outperforming all VLM baselines, while also improving answer stability across repeated stochastic runs. The key insight is that multi-perspective conflict exposure plus guideline-grounded resolution reduces both hallucinated measurement selection and borderline answer flip-flops.

Problem

Existing foundation-model pipelines for echocardiography solve isolated subtasks (view classification, segmentation, EF estimation) without integrating multi-view temporal evidence, quantitative measurements, and guideline-grounded reasoning into a unified, auditable pipeline. Single-trajectory ReAct agents are brittle near clinical severity thresholds, where small measurement noise flips diagnostic conclusions, and they lack mechanisms to prevent metric hallucination (e.g., using EF when cavity size is asked).

Method

Echo-CoPilot operates in two stages:

1. Multi-Perspective Generation. Three independent ReAct agents—Structural (\(P_1\)), Pathological (\(P_2\)), Quantitative (\(P_3\))—each execute up to 10 reasoning steps, invoking a shared tool suite (EchoPrime for view classification/measurements, PanEcho for disease prediction) via a deterministic shared cache that deduplicates identical tool calls. All agents query EchoKG at the start.

2. Self-Contrast Analysis. A dedicated contrast LLM receives all three perspectives, aggregated tool measurements, and EchoKG-retrieved constraints. It generates a structured discrepancy checklist—checking whether each perspective followed EchoKG guidance, flagging severity classification mismatches, comparing quantitative consistency, assessing tool reliability—then produces a final answer that weights EchoKG-compliant perspectives over outliers rather than using majority voting.

EchoKG is a directed graph \(G = (V, E)\) with three node types: Structure (\(V_s\)), Measurement (\(V_m\)), and Pattern (\(V_p\), 16 templates). It encodes dual-edge relationships: requires edges map a clinical pattern (e.g., ventricular dilation) to the measurements that must be used (LVIDd, LVEDV), and avoid edges explicitly exclude irrelevant metrics (e.g., EF for cavity size). Pattern matching is priority-ordered (specific before general) using keyword intersection with optional context and exclusion sets.

Key Contributions

  • EchoKG: A guideline-derived knowledge graph (ASE, EACVI) with explicit requires/avoid edges for measurement selection, serving as a deterministic inference-time guardrail against metric hallucination.
  • Multi-perspective ReAct architecture: Three specialized agents with a shared deterministic tool cache, eliminating redundant tool calls while preserving cross-perspective independence.
  • Self-contrast mechanism: A contrast LLM that generates a structured discrepancy checklist and resolves conflicts by weighting guideline-adherent perspectives, not simple majority voting.
  • SOTA accuracy on MIMICEchoQA: 54.0% vs. 49.0% (next best: Qwen2-VL-7B-Biomed) across 622 echocardiogram videos and 38 views.
  • Stochasticity stress test: Demonstrates higher diagnostic stability across 500 inferences (10 runs Ă— 50 questions) compared to LLM, LLM+Tools, and single ReAct baselines.

Results

  • Accuracy (MIMICEchoQA, zero-shot): Echo-CoPilot 54.0% vs. Qwen2-VL-7B-Biomed 49.0%, GPT-o4-mini 43.9%, GPT-4o 41.6%, Qwen2.5-VL-72B 34.2%.
  • Stability Rate (N=50, 10 runs): Echo-CoPilot 72% vs. ReAct Agent 70%, LLM 40%, LLM+Tools 26%.
  • Avg Changes (answer flips): Echo-CoPilot 0.96 vs. ReAct 1.14, LLM 2.10, LLM+Tools 2.58.
  • Ablation (accuracy): LLM 46.9% → LLM+Tools+EchoKG 49.7% → ReAct Agent 52.3% → Echo-CoPilot (+ SC) 54.0%.
  • Error reduction on hard views: Doppler views 83.0% → 59.3%; PSAX great vessels 71.0% → 59.6% vs. Qwen2-VL-7B-Biomed.
  • LLM pure logic errors: <1% across 500 iterations; tool variability identified as primary bottleneck.

Limitations

  • Evaluation restricted to a single benchmark (MIMICEchoQA, 622 videos); no multi-center or prospective clinical validation.
  • Performance ceiling is constrained by the reliability of underlying perceptual tools (EchoPrime, PanEcho); when tools produce noisy/conflicting outputs, the SC mechanism can produce internally consistent but factually incorrect conclusions.
  • EchoKG covers 16 pattern templates derived from ASE/EACVI guidelines; coverage gaps for uncommon or emerging clinical queries are not quantified.
  • Computational cost of three parallel ReAct loops plus a contrast pass is not analyzed relative to single-agent baselines.

Relevance to Harnesses / Meta-Harnesses

Echo-CoPilot is a concrete example of a meta-harness pattern: a controller layer that instantiates multiple independent sub-agents, collects their outputs, and runs a dedicated synthesis/verification agent over the ensemble—structurally analogous to a verifier or judge harness. The self-contrast mechanism is essentially an adversarial cross-check pass (discrepancy checklist generation + conflict resolution) that mirrors the "adversarial verify" pattern in multi-agent orchestration harnesses. EchoKG functions as an external constraint registry that the harness injects at multiple stages—both during agent execution and during synthesis—demonstrating how structured external knowledge can serve as a guardrail across harness stages. The shared deterministic tool cache is a harness-level optimization (deduplication of sub-agent tool calls) directly applicable to any multi-agent harness that fans out over a shared tool pool.