UMCA-Net: Uncertainty-aware multi-stage cross-attention for cost-aware multi-omics data classification
Yehong Huang1, Huan Huang1, Selena He1
1Department of Computer Science, Kennesaw State University, 680 Arntson Dr, Marietta, GA 30060, USA.
Computers in Biology and Medicine
|July 21, 2026
Summary
This study introduces an uncertainty-aware framework for disease classification using multi-omics data. It efficiently reduces costs by adaptively acquiring data only when needed, ensuring trustworthy diagnoses.
Area of Science:
- Computational biology
- Bioinformatics
- Medical informatics
Background:
- Multi-omics data integration is crucial for precision medicine.
- High costs and data complexity hinder clinical adoption of multi-omics.
- Existing methods often use static data fusion strategies.
Purpose of the Study:
- To develop an efficient and trustworthy disease classification framework using multi-omics data.
- To address the challenges of high acquisition costs and data complexity.
- To enable cost-sensitive and progressive inference in clinical settings.
Main Methods:
- Proposed an uncertainty-aware multi-view dynamic decision framework.
- Utilized evidential deep learning based on Dempster-Shafer theory.
- Developed a Transformer-based multi-stream architecture (UMCA-Net) with cross-attention for adaptive data acquisition.
Main Results:
- Achieved state-of-the-art performance on four benchmark multi-omics datasets.
- Significantly reduced data acquisition requirements.
- Demonstrated confident classification for over 90% of samples using initial modalities without accuracy loss in some cohorts.
Conclusions:
- The proposed framework offers a scalable and practical solution for balancing diagnostic accuracy and economic cost.
- Facilitates the clinical deployment of multi-omics models.
- Provides principled uncertainty quantification and dynamic decision-making capabilities.