Predicting Clinical Outcomes with Waveform LSTMs

TL;DR

The paper addresses improving ICU mortality prediction by leveraging high-frequency waveform data, focusing on ECG-derived signals within the first 48 hours of ICU admission. It compares logistic regression, a standard LSTM, and a channel-wise LSTM, augmenting traditional clinical features with 12 waveform-derived features and applying an A-SUWO oversampling strategy to address class imbalance. The main findings show modest gains for logistic regression and notable gains for the standard LSTM when waveform information is added, with mixed results for the channel-wise LSTM, particularly in ROC-AUC but meaningful improvements in precision-recall. This work demonstrates the practical potential of multimodal, waveform-informed approaches to improve early mortality prediction and suggests avenues for extending coverage to additional waveform channels and time-windowed features to further boost performance.

Abstract

Data mining and machine learning hold great potential to enable health systems to systematically use data and analytics to identify inefficiencies and best practices that improve care and reduce costs. Waveform data offers particularly detailed information on how patient health evolves over time and has the potential to significantly improve prediction accuracy on multiple benchmarks, but has been widely under-utilized, largely because of the challenges in working with these large and complex datasets. This study evaluates the potential of leveraging clinical waveform data to improve prediction accuracy on a single benchmark task: the risk of mortality in the intensive care unit. We identify significant potential from this data, beating the existing baselines for both logistic regression and deep learning models.