SimTA++: Simple attention neural network for clinical asynchronous time series.

Abstract

The task of modeling asynchronous time series data in clinical practice is a formidable challenge, made even more difficult by the multi-modal and irregular nature of real-world medical time series data. Despite advances in deep learning for sequential data processing in other domains, e.g., natural language processing and speech recognition, a solution for clinical scenarios is yet to be fully realized. To meet this challenge, we propose Simple Temporal Attention (SimTA), a new temporal attention module, where the relation between asynchronous timesteps is modeled by a simple time-dependent attention mechanism. We further extend the method into SimTA++, by incorporating a non-linear temporal attention mechanism that captures non-monotonic relationships in temporal data. Our experiments demonstrate the superiority of SimTA++ over recurrent neural networks, graph neural networks, temporal fusion transformers and neural differential equations on three benchmark datasets, including a synthetic proof-of-concept dataset, PhysioNet 2019, and an in-house immunotherapy response prediction dataset. Notably, the proposed method achieves promising performance as a predictive multi-omics biomarker in predicting the immunotherapy response. The source code for our PyTorch implementation is available on GitHub at https://github.com/jingyuerY/SimTA_plus.