MMHA: A multi-modal data fusion algorithm with hierarchical attention for disease diagnosis and cancer subtype prediction

• Published in: Expert systems with applications Vol. 288; p. 128205
• Main Author: Wang, Xue
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2025
• Subjects: Disease prediction; Hierarchical attention; Multi-modal data fusion; Variable channel; Variable channel; Hierarchical attention; Multi-modal data fusion; Disease prediction
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2025.128205

•Uses hierarchical attention to improve accuracy without relying on graph structures.•Shows the effectiveness of different modalities in disease diagnosis.•Explores neural network interpretability in biomarker identification.•Compiles public datasets for disease diagnosis and cancer subtype classification. [Display omitted] In precision medicine, multiomics data integration provides a comprehensive understanding of biological processes, making it essential for accurate disease diagnosis and biomarker identification. However, most existing approaches rely on graph neural networks to combine and propagate information, often integrating additional models to create multi-modal representations of patients. While effective in capturing common associations between individuals, these graph-based methods often overlook critical differences among disease subtypes. To address this limitation, we propose a novel end-to-end interpretable method for multiomics integration, designed to enhance cancer subtype prediction. Our approach employs a Multi-Modal Hierarchical Attention (MMHA) mechanism to handle the complexity of multi-source data with varying scales and structures. Unlike graph-based methods, our model uses the adaptive weighting capabilities of hierarchical attention to aggregate features from different data modalities, uncovering relationships and complementarities without the need for predefined graph structures. Extensive experiments on six public datasets spanning cancer and neurodegenerative disease subtypes demonstrate that our method significantly outperforms state-of-the-art techniques. For instance, MMHA improves accuracy by 4.5% on the LGG dataset and achieves near-perfect performance with 99.09% accuracy on the KIPAN dataset, highlighting its robustness in diverse scenarios. Moreover, it effectively identifies key biomarkers, offering valuable insights to support the development of personalized treatment plans and advancing precision medicine.