跳转至

文章背景与核心概要

心电图(ECG)对于心血管疾病的诊断至关重要,然而,如何利用多导联记录自动生成可靠、内容丰富的临床报告仍然是一个巨大的挑战。传统的深度学习方法主要集中在简单的分类任务上,或者生成的文本输出无法满足实际医疗应用的需求。

为了克服这些障碍,本文作者推出了 ECG-LENS,这是一个端到端的框架,旨在将多导联信号建模、诊断感知表征以及具有临床依据的文本生成有机结合。此外,论文还提出了 F1-ECGBERT,这是一种基于 ECG 专有 BERT 的评估指标,用于准确衡量生成诊断标签与参考诊断标签之间的一致性,从而克服了标准词汇指标的局限性。实验结果表明,无论是在域内(PTB-XL)还是跨域(MIMIC-IV-ECG)基准测试中,ECG-LENS 的性能均显著优于当前的尖端模型。


ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation

Authors: Akanta Das, Tasinul Islam Ahon, Ahmed Mahir Sultan Rumi, Md Mahbubur Rahman, Tausif Amim Shadly, Tanzima Hashem
ArXiv ID: arXiv:2608.05893 [cs.AI]
Submitted: 6 August 2026

Authors: Akanta Das, Tasinul Islam Ahon, Ahmed Mahir Sultan Rumi, Md Mahbubur Rahman, Tausif Amim Shadly, Tanzima Hashem
ArXiv ID: arXiv:2608.05893 [cs.AI]
Submitted: 6 August 2026


📌 Summary

📌 Summary

心电图(ECG)对于心血管疾病的诊断至关重要,然而,如何利用多导联记录自动生成可靠、内容丰富的临床报告仍然是一个巨大的挑战。传统的深度学习方法主要集中在简单的分类任务上,或者生成的文本输出无法满足实际医疗应用的需求。

Electrocardiography (ECG) is vital for diagnosing cardiovascular diseases, but automating the generation of reliable, dense clinical reports from multi-lead recordings remains a significant challenge. Traditional deep learning approaches largely focus on simple classification or produce text outputs inadequate for practical medical use.

为了克服这些障碍,作者推出了 ECG-LENS,这是一个端到端的框架,旨在将多导联信号建模、诊断感知表征以及具有临床依据的文本生成有机结合。此外,论文还提出了 F1-ECGBERT,这是一种基于 ECG 专有 BERT 的评估指标,用于准确衡量生成诊断标签与参考诊断标签之间的一致性,从而克服了标准词汇指标的局限性。实验结果表明,无论是在域内(PTB-XL)还是跨域(MIMIC-IV-ECG)基准测试中,ECG-LENS 的性能均显著优于当前的尖端模型。

To overcome these hurdles, the authors introduce ECG-LENS, an end-to-end framework designed to jointly integrate multi-lead signal modeling, diagnosis-aware representations, and clinically grounded text generation. Furthermore, the paper proposes F1-ECGBERT, an ECG-specific BERT-based metric to accurately measure agreement between generated and reference diagnostic labels, overcoming the limitations of standard lexical metrics. Experimental results demonstrate that ECG-LENS outperforms state-of-the-art models significantly across both in-domain (PTB-XL) and cross-domain (MIMIC-IV-ECG) benchmarks.


🧠 Key Contributions & Methodology

🧠 Key Contributions & Methodology

  • 导联感知信号建模: 结合了导联编码器(以保留局部的波形形态)与全局编码器(旨在捕获导联间的依赖关系)。
  • 临床依据生成: 将丰富的信号表征与富含临床信息的文本提示相融合,从而有效地对 GPT-2 解码器进行条件约束。
  • 心电图专用预处理: 实施了定制化的报告预处理策略,将模型的关注点严格引导至具有临床意义的发现上。
  • F1-ECGBERT 指标: 引入了一种新型的基于 BERT 的评估指标,该指标与真实的临床诊断结果高度吻合,而不仅仅依赖于表层文本的相似性。
  • Lead-Aware Signal Modeling: Combines lead-wise encoders to preserve localized waveform morphologies with a global encoder designed to capture inter-lead dependencies.
  • Clinically Grounded Generation: Fuses rich signal representations with clinically enriched textual prompts to effectively condition a GPT-2 decoder.
  • ECG-Specific Preprocessing: Implements a tailored report-preprocessing strategy that directs the model's focus strictly toward clinically meaningful findings.
  • F1-ECGBERT Metric: Introduces a novel BERT-based evaluation metric that aligns closely with real-world clinical diagnostic findings rather than relying solely on surface-level text similarity.

📊 Performance Highlights

📊 Performance Highlights

在与强大的基准模型进行对比评估时,ECG-LENS 实现了显着的绝对性能提升: * METEOR 提升 +4.0% * ROUGE-L 提升 +6.3% * F1-ECGBERT 提升 +11.5%

When evaluated against strong baselines, ECG-LENS achieves remarkable absolute performance gains: * +4.0% in METEOR * +6.3% in ROUGE-L * +11.5% in F1-ECGBERT