Time-Aware Transformer-Based Network for Clinical Notes Series Prediction

Abstract

A patient’s clinical notes correspond to a sequence of free-form text documents generated by healthcare professionals over time. Rich and unique information in clinical notes is useful for clinical decision making. In this work, we propose a time-aware transformer-based hierarchical architecture, which we call Flexible Time-aware LSTM Transformer (FTL-Trans), for classifying a patient’s health state based on her series of clinical notes. FTL-Trans addresses the problem that current transformer-based architectures cannot handle, which is the multi-level structure inherent in clinical note series where a note contains a sequence of chucks and a chuck contains further a sequence of words. At the bottom layer, FTL-Trans encodes equal-length subsequences of a patient’s clinical notes ("chunks") into content embeddings using a pre-trained ClinicalBERT model. Unlike ClinicalBERT, however, FTL-Trans merges each content embedding and sequential information into a new position-enhanced chunk representation in the second layer by an augmented multi-level position embedding. Next, the time-aware layer tackles the irregularity in the spacing of notes in the note series by learning a flexible time decay function and utilizing the time decay function to incorporate both the position-enhanced chunk embedding and time information into a patient representation. This patient representation is then fed into the top layer for classification. Together, this hierarchical design of FTL-Trans successfully captures the multi-level sequential structure of the note series. Our extensive experimental evaluation conducted using multiple patient cohorts extracted from the MIMIC dataset illustrates that, while addressing the aforementioned issues, FTL-Trans consistently outperforms the state-of-the-art transformer-based architectures up to 5% in AUROC and 6% in Accuracy.