In the transformation toward data-driven healthcare, disease diagnosis faces challenges in multimodal data fusion and uncertain decision-making. To address this, this paper proposes an adaptive threshold-optimized multimodal three-way decision model integrating neighborhood rough sets with simulated annealing (SA). First, a multimodal hybrid decision information system is defined, achieving unified representation of multimodal data through modality-aware attribute neighborhood partitioning. Second, attribute weights are characterized using an information gain-driven objective weighting method, combined with a cross-modal weighted fusion mechanism to construct a multimodal weighted neighborhood decision rough set. Finally, integrating SA and multi-layer perceptron (MLP) constructs a two-stage sequential three-way decision model with adaptive threshold optimization that dynamically optimizes decision thresholds, solves the boundary domain sample retention issue in single-stage decision-making, and forms a positive cycle of ‘data accumulation → uncertainty resolution’. Experimental results on real clinical data show that the proposed model effectively handles multimodal medical data, significantly improves the accuracy of uncertain decisions in multimodal disease diagnosis, and provides data-driven auxiliary diagnostic support for physicians.