Attention-Enhanced Cross-Modal Temporal Alignment with Directional Metric Learning for Multimodal Diabetes Detection

Abstract

Integrating clinical information with wearable measurements may support diabetes prediction, but the modalities differ in sampling density and representation. This study proposes a multimodal framework that uses transformer-based cross-modal attention to align sparse clinical information with higher-frequency wearable time series. It combines a directional similarity term motivated by the von Mises–Fisher distribution with a learnable gate that integrates early- and late-fusion representations; the alignment, similarity, and prediction objectives are optimized jointly. In the reported test evaluation, the proposed framework achieved an AUROC of 0.923, an AUPRC of 0.871, and an F1 score of 0.802, compared with 0.891, 0.832, and 0.768, respectively, for the cross-modal attention baseline. The reported early-prediction score was 0.813 for the proposed model and 0.742 for that baseline. In a review of 100 test cases, clinicians rated 84% of the attention patterns as clinically plausible (Fleiss’ κ=0.72). Mean inference latency was 0.83 ms per sample on an NVIDIA V100 GPU. These results indicate improved predictive performance and clinically plausible attention patterns under the reported evaluation conditions.