An Explainable Multimodal Artificial Intelligence Model Integrating Histopathological Microenvironment and EHR Phenotypes for Germline Genetic Testing in Breast Cancer

An Explainable Multimodal Artificial Intelligence Model Integrating Histopathological Microenvironment and EHR Phenotypes for Germline Genetic Testing in Breast Cancer

Abstract Genetic testing for pathogenic germline variants is critical for the personalized management of high‐risk breast cancers, guiding targeted therapies and cascade testing for at‐risk families. In this study, MAIGGT (Multimodal Artificial Intelligence Germline Genetic Testing) is proposed, a d...

Full description

Saved in:
Bibliographic Details
Main Authors: Zijian Yang, Changyuan Guo, Jiayi Li, Yalun Li, Lei Zhong, Pengming Pu, Tongxuan Shang, Lin Cong, Yongxin Zhou, Guangdong Qiao, Ziqi Jia, Hengyi Xu, Heng Cao, Yansong Huang, Tianyi Liu, Jian Liang, Jiang Wu, Dongxu Ma, Yuchen Liu, Ruijie Zhou, Xiang Wang, Jianming Ying, Meng Zhou, Jiaqi Liu
Format: Article
Language:English
Published: Wiley 2025-08-01
Series:Advanced Science
Subjects:
BRCA1/2
breast cancer
deep learning
electronic health records
tumor microenvironment
whole‐slide image
Online Access:https://doi.org/10.1002/advs.202502833
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract Genetic testing for pathogenic germline variants is critical for the personalized management of high‐risk breast cancers, guiding targeted therapies and cascade testing for at‐risk families. In this study, MAIGGT (Multimodal Artificial Intelligence Germline Genetic Testing) is proposed, a deep learning framework that integrates histopathological microenvironment features from whole‐slide images with clinical phenotypes from electronic health records for precise prescreening of germline BRCA1/2 mutations. Leveraging a multi‐scale Transformer‐based deep generative architecture, MAIGGT employs a cross‐modal latent representation unification mechanism to capture complementary biological insights from multimodal data. MAIGGT is rigorously validated across three independent cohorts and demonstrated robust performance with areas under receiver operating characteristic curves of 0.925 (95% CI 0.868 – 0.982), 0.845 (95% CI 0.779 – 0.911), and 0.833 (0.788 – 0.878), outperforming single‐modality models. Mechanistic interpretability analyses revealed that BRCA1/2‐mutated associated tumors may exhibit distinct microenvironment patterns, including increased inflammatory cell infiltration, stromal proliferation and necrosis, and nuclear heterogeneity. By bridging digital pathology with clinical phenotypes, MAIGGT establishes a new paradigm for cost‐effective, scalable, and biologically interpretable prescreening of hereditary breast cancer, with the potential to significantly improve the accessibility of genetic testing in routine clinical practice.
ISSN:2198-3844

Similar Items