Scholarlink

<p>Artificial intelligence (AI) models capable of detecting a patient’s reported race from medical data raise important concerns around fairness and equity. In this study, we investigated whether machine learning models could identify race-based differences in electrocardiograms (ECGs) from healthy Black and White individuals and explored the origins of these differences. We analyzed approximately 10 million ECG traces from 1.76 million subjects across multiple institutions. A convolutional neural network (CNN)-based classifier was developed and optimized for various configurations, including network depth, fusion strategy, and input format. The best-performing model, a 1-layer late fusion CNN using median beats as input, achieved an AUC of 86.17 (±0.34). Performance was consistent across sexes (AUC≈86%), and analysis suggested race-related ECG signatures appear after birth. Socioeconomic status influenced model accuracy, and interpretability analyses revealed the QRS complex as a key contributor. These findings highlight the presence of non-genetic, race-associated patterns in ECGs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Non-genetic factors determine deep learning identified ECG differences between black and white healthy subjects

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月10日

查看全文
  • Sandeep Chandra Bollepalli,
  • Eric M. Isselbacher,
  • Jagmeet P. Singh,
  • Antonis A. Armoundas

摘要

Artificial intelligence (AI) models capable of detecting a patient’s reported race from medical data raise important concerns around fairness and equity. In this study, we investigated whether machine learning models could identify race-based differences in electrocardiograms (ECGs) from healthy Black and White individuals and explored the origins of these differences. We analyzed approximately 10 million ECG traces from 1.76 million subjects across multiple institutions. A convolutional neural network (CNN)-based classifier was developed and optimized for various configurations, including network depth, fusion strategy, and input format. The best-performing model, a 1-layer late fusion CNN using median beats as input, achieved an AUC of 86.17 (±0.34). Performance was consistent across sexes (AUC≈86%), and analysis suggested race-related ECG signatures appear after birth. Socioeconomic status influenced model accuracy, and interpretability analyses revealed the QRS complex as a key contributor. These findings highlight the presence of non-genetic, race-associated patterns in ECGs.