Poster
in
Workshop: Workshop on Learning from Time Series for Health
Parallel Time-Sensor Attention for Electronic Health Record Classification
Rachael DeVries · Marie Lisandra Zepeda Mendoza · Ole Winther
Keywords: [ Missing Values ] [ explainability ] [ transformers ] [ classification ] [ novel architectures ] [ noisy/irregular measurements ]
When working with electronic health records (EHR), it is critical for deep learning (DL) models to achieve both high performance and explainability. Here we present the Parallel Attention Transformer (PAT), which performs temporal and sensor attention in parallel, is competitive to state-of-the-art models in EHR classification, and has a uniquely explainable structure. PAT is trained on two EHR datasets, compared to five DL models of different architectures, and its attention weights are used to visualize key sensors and time points. Our results show that PAT is particularly well-suited for healthcare and pharmaceutical applications, which have a strong interest in identifying key features to differentiate patient groups and conditions, and key times for intervention.