DNA-T: Deformable Neighborhood Attention Transformer for Irregular Medical Time Series
This paper proposes a novel Deformable Neighborhood Attention Transformer (DNA-T) for irregular medical time series. The proposed Missingness-Aware Embedding layer integrates the timestamps, masks, and intervals to model the missing patterns, which are used to enhance the latent representation of the time series. In addition, a new DNA module is designed that is able to flexibly select relevant neighborhoods using the missing patterns of the neighborhoods at each time point.
1. Framework of the DNA-T
Figure 1. Overall architecture of DNA-T.
This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 62102098, in part by the Science and Technology Planning Project of Guangzhou under Grant 202201010266, in part by the National Natural Science Foundation of China (NSFC) under Grant 62302413, in part by the Natural Science Foundation of Guangdong Province under Grant 2024A1515010186, as well as Regional Joint Fund Project of Basic and Applied Basic Research Foundation of Guangdong Province under Grant 2022A1515140096.