Severe viral rash progression prediction and surveillance early warning based on local-global attention and temporal convolutional network.

Aiming at the problems of delayed early warning for severe viral exanthems and insufficient prediction accuracy based on a single data source, this study proposes a dual-branch cross-modal fusion prediction model based on TCN-TransDAF. The model adopts a Dilated-TCN branch to capture the long-term evolution law of clinical temporal indicators within 72 h after diagnosis, and a local-global attention Transformer branch to mine the variation characteristics of virulence loci in viral genomes. With the aid of a cross-modal fusion mechanism, it realizes the deep coupling of dual-dimensional information of host and pathogen, and simultaneously completes the tasks of severe risk determination, subtype identification and onset time window prediction. The experimental results show that the model achieves an AUC-ROC of 0.942 for severe case prediction and a MAE of 5.8 h for severe case onset time window prediction on the test set. Compared with the optimal baseline model, the AUC-ROC is increased by 5.79%, the MAE is reduced by 26.34%, and the recall is increased by 10.09%, which enables accurate early warning 12-18 h before the onset of severe symptoms. This study provides an efficient technical solution for the early warning of severe viral exanthems, and its cross-modal modeling idea also offers a reference for the prediction research of other infectious diseases.