Fusion-driven multimodal Learning for Biomedical Time Series in Surgical Care

Jinshan Che^*Mingming SunYuhong Wang

• Xinxiang Central Hospital, Xinxiang, China

The integration of multimodal data has become a crucial aspect of biomedical time series prediction, offering improved accuracy and robustness in clinical decision-making.Traditional approaches often rely on unimodal learning paradigms, which fail to fully exploit the complementary information across heterogeneous data sources such as physiological signals, imaging, and electronic health records. These methods suffer from modality misalignment, suboptimal feature fusion, and lack of adaptive learning mechanisms, leading to performance degradation in complex biomedical scenarios. To address these challenges, we propose a novel multimodal Deep Learning framework that dynamically captures inter-modal dependencies and optimizes cross-modal interactions for time series prediction. Our approach introduces an Adaptive multimodal Fusion Network (AMFN), which leverages attention-based alignment, graphbased representation learning, and a modality-adaptive fusion mechanism to enhance information integration. Furthermore, we develop a Dynamic Cross-Modal Learning Strategy (DCMLS) that optimally selects relevant features, mitigates modality-specific noise, and incorporates uncertaintyaware learning to improve model generalization. Experimental evaluations on biomedical datasets demonstrate that our method outperforms state-of-the-art techniques in predictive accuracy, robustness, and interpretability. By effectively bridging the gap between heterogeneous biomedical data sources, our framework offers a promising direction for AI-driven disease diagnosis and treatment planning.