New Framework Enhances Trustworthy Multimodal Fusion for Medical Prognosis

Multimodal learning holds significant promise for improving predictive accuracy in fields such as medical prognosis. However, its practical deployment, particularly in safety-critical environments, is often hindered by poor performance when data streams are noisy or uninformative. Existing fusion methods typically lack robust mechanisms for dynamically assessing data quality or providing reliable confidence scores for their predictions, which is a major barrier to clinical adoption. To address these critical limitations, a new framework called Adaptive Confidence-weighted Expansion (ACE) has been developed. ACE works by first enriching the multimodal data space through the generation of new, complementary modalities derived from intra-modality correlations. Following this, it implements a sophisticated dual-level confidence system. This system adaptively reweighs all modalities based on their individual reliability before fusion and then calculates a comprehensive trust score for the final, fused decision. Evaluations using four challenging multi-omics datasets (BRCA, KIPAN, LGG, and ROSMAP) demonstrate that ACE significantly outperforms current state-of-the-art algorithms. It shows improvements in both classification performance and the calibration of confidence scores, offering a more stable and robust data fusion method suitable for high-stakes problems.