Scholarlink

<p>Chest radiography (CXR) is widely accessible, and its ability to capture subtle manifestations of systemic disease remains underexplored. We developed a contrastively pretrained multimodal CXR foundation model using large-scale image–report pairs and evaluated its capacity to predict diverse human diseases. Using 1074 phecodes derived from electronic health records, we trained linear probes on frozen image embeddings and validated performance across three independent cohorts (<i>n</i> = 90,911; <i>n</i> = 79,786; <i>n</i> = 60,282). The model significantly predicted 554 prevalent and 457 incident phenotypes, with 60 prevalent and 42 incident phenotypes demonstrating consistently high discrimination across all datasets. To enhance interpretability, both diseases and 57 radiologist-curated CXR features were embedded into a shared representation space. This co-embedding analysis identified 28 phenotype clusters driven by recognizable imaging patterns, including cardiomegaly, atherosclerosis, and ground-glass opacities, and enabled reconstruction of most predictions with strong explanatory performance (<i>R</i>² ≥ 0.85). Importantly, the embeddings captured imaging signatures associated with near-term cardiovascular events and critical illness. These findings demonstrate that CXR foundation model embeddings encode rich, clinically relevant information beyond conventional interpretation, providing a scalable and interpretable framework for comprehensive disease profiling and a foundation for future prospective validation.</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

Transparent chest radiograph foundation model enables explainable human disease profiling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Chin Lin,
  • Kai-Chieh Chen,
  • Jun-Wei Huang,
  • Wen-Hui Fang,
  • Wei-Chou Chang,
  • Kai-Hsiung Ko,
  • Yi-Chih Hsu,
  • Chin-Sheng Lin,
  • Shih-Hua Lin,
  • Dung-Jang Tsai

摘要

Chest radiography (CXR) is widely accessible, and its ability to capture subtle manifestations of systemic disease remains underexplored. We developed a contrastively pretrained multimodal CXR foundation model using large-scale image–report pairs and evaluated its capacity to predict diverse human diseases. Using 1074 phecodes derived from electronic health records, we trained linear probes on frozen image embeddings and validated performance across three independent cohorts (n = 90,911; n = 79,786; n = 60,282). The model significantly predicted 554 prevalent and 457 incident phenotypes, with 60 prevalent and 42 incident phenotypes demonstrating consistently high discrimination across all datasets. To enhance interpretability, both diseases and 57 radiologist-curated CXR features were embedded into a shared representation space. This co-embedding analysis identified 28 phenotype clusters driven by recognizable imaging patterns, including cardiomegaly, atherosclerosis, and ground-glass opacities, and enabled reconstruction of most predictions with strong explanatory performance (R² ≥ 0.85). Importantly, the embeddings captured imaging signatures associated with near-term cardiovascular events and critical illness. These findings demonstrate that CXR foundation model embeddings encode rich, clinically relevant information beyond conventional interpretation, providing a scalable and interpretable framework for comprehensive disease profiling and a foundation for future prospective validation.