Scholarlink

<p>Diabetes poses a serious threat to people’s physical and mental health. Traditional fasting blood glucose monitoring plays an important role in diabetes diagnosis, but its reliance on single-modality data makes it difficult to comprehensively capture underlying patterns, leading to a certain risk of misdiagnosis. To address the limitations of single-modality diabetes prediction, this study aims to develop a high-accuracy multimodal fusion model (<b>MAMIFusion</b>, <b>M</b>asked blood-test modeling(MBTM), <b>A</b>ttention mechanisms, <b>M</b>utual <b>I</b>nformation maximization and multimodal EHR data <b>Fusion</b>) for early and reliable diabetes diagnosis. This study reviewed over 2,000 admitted patients from the endocrinology department of a tertiary hospital. Key methods: 1) MBTM handles missing blood test data via reconstruction; 2) Hierarchical attention extracts key numerical/textual features while cross-modal attention captures inter-modal interactions to simulate the actual decision-making process of doctors; 3) Mutual information maximization optimizes fused features. The end-to-end model jointly optimizes reconstruction, fusion, and classification through a comprehensive loss function. MAMIFusion achieved 93.56% accuracy on a clinical dataset (tertiary grade-A hospital in Qingdao, n=1,053) and 91.25% on the public dataset (NEISS2023, n=104,129), outperforming single-modality baselines by 16%. The results showed that MAMIFusion enhances both the accuracy and generalizability of diabetes prediction, and thus can serve as a valuable complementary tool for early screening of diabetic patients. Furthermore, the proposed methodological framework demonstrates potential for extension to the prediction of other chronic diseases.</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

MAMIFusion: a multimodal data fusion framework based on EHRs for enhancing diabetes prediction accuracy

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Yu Zhang,
  • Tingqiang Song,
  • Lili Wei,
  • Jingyuan Wang,
  • Yueshuai Pan,
  • Guozheng Wei,
  • Yuyao Guo,
  • Haoran Song,
  • Wenjuan Cao

摘要

Diabetes poses a serious threat to people’s physical and mental health. Traditional fasting blood glucose monitoring plays an important role in diabetes diagnosis, but its reliance on single-modality data makes it difficult to comprehensively capture underlying patterns, leading to a certain risk of misdiagnosis. To address the limitations of single-modality diabetes prediction, this study aims to develop a high-accuracy multimodal fusion model (MAMIFusion, Masked blood-test modeling(MBTM), Attention mechanisms, Mutual Information maximization and multimodal EHR data Fusion) for early and reliable diabetes diagnosis. This study reviewed over 2,000 admitted patients from the endocrinology department of a tertiary hospital. Key methods: 1) MBTM handles missing blood test data via reconstruction; 2) Hierarchical attention extracts key numerical/textual features while cross-modal attention captures inter-modal interactions to simulate the actual decision-making process of doctors; 3) Mutual information maximization optimizes fused features. The end-to-end model jointly optimizes reconstruction, fusion, and classification through a comprehensive loss function. MAMIFusion achieved 93.56% accuracy on a clinical dataset (tertiary grade-A hospital in Qingdao, n=1,053) and 91.25% on the public dataset (NEISS2023, n=104,129), outperforming single-modality baselines by 16%. The results showed that MAMIFusion enhances both the accuracy and generalizability of diabetes prediction, and thus can serve as a valuable complementary tool for early screening of diabetic patients. Furthermore, the proposed methodological framework demonstrates potential for extension to the prediction of other chronic diseases.