Predicting in-hospital mortality by combining clinical notes with time-series data

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4026–4031 August 1–6, 2021. ©2021 Association for Computational Linguistics 4026 Predicting in-hospital mortality by combining clinical notes with time-series data Iman Deznabi, Mohit Iyyer, Madalina Fiterau College of Information and Computer Sciences University of Massachusetts Amherst {iman, miyyer, mfiterau}@cs.umass.edu Abstract In intensive care units (ICUs), patient health is monitored through (1) continuous vital signals from various medical devices, and (2) clinical notes consisting of opinions and summaries from doctors which are recorded in electronic health records (EHR). It is difficult to jointly model these two sources of information be- cause clinical notes, unlike vital signals, are collected at irregular intervals and their con- tents are relatively unstructured. In this paper, we present a model that combines both sources of information about ICU patients to make ac- curate in-hospital mortality predictions. We apply a fine-tuned BERT model to each of the patient’s clinical notes. The resulting embed- dings are then combined to obtain the overall embedding for the entire text part of the data. This is then combined with the output of an LSTM model that encodes patients’ vital sig- nals. Our model improves upon the state of the art for mortality prediction, attaining an AUC score of 0.9, compared to the previous 0.87, setting a new standard for mortality prediction on the MIMIC III benchmark.1 1 Introduction One of the major costs in healthcare is critical care in intensive care units. A crucial aspect of critical care is mortality prevention. With advancements in healthcare, a patient’s condition during an ICU visit is monitored through devices that measure many different vital signals, such as heart rate, systolic blood pressure, temperature, etc. Additionally, we have access to clinical notes written by medical professionals during the patient’s stay. This data can be used to predict the condition of patients during their ICU stay, which can help in managing the expensive resources in hospitals and providing 1Our implementation is available at https:// github.com/Information-Fusion-Lab-Umass/ ClinicalNotes_TimeSeries these services to patients who need them most in a more cost-effective way. Most prior work has been focused on predicting patient health using the time-series data gathered from medical devices (Lipton et al., 2016; Che et al., 2018; Narayan Shukla and Marlin, 2020; Xu et al., 2018). More recent work also attempted to leverage clinical notes for making these predictions (Zhang et al., 2019; Ghorbani et al., 2020; Lee et al., 2020; Alsentzer et al., 2019). However, research on combining time series data with clinical notes for outcome prediction has been limited due to the irregularity of the clinical notes compared to the time series data, and the complexity of jointly modeling the two sources of information. In this work, we propose a multimodal neural network that combines time-series data from medi- cal devices with textual information from clinical notes to improve the prediction of in-hospital mor- tality. This task is defined as predicting whether a patient will die before getting discharged from hos- pital based on the first …

Content truncated to 500 words