Engineering biomarker representations of vital signs data enhances deep learning mortality prediction

Comprehensive Summary

The ability to accurately predict inpatient mortality in patients admitted to the intensive care unit (ICU) is crucial to approaching patient care. This study differentiated between three methods of predicting patient mortality after 24 hours of ICU admission. These methods included raw data collected every 5 minutes, preprocessed data averaged over an hour, and pre-processed biomarker representations of vital sign data using an expanded version of PhysioZoo Pulse Oximetry Benchmarking (POBM). The expanded PhysioZoo POBM is a platform that analyzes data collected over time including temperature, heart rate, blood pressure, respiratory rate, and SpO2. This vital signs data can be used to train a bidirectional long short-term memory (BiLSTM) classifier that has the ability to predict mortality. The superior AUROC, AUPRC, and Brier scores suggested that biomarker representations are superior in training deep learning models to accurately predict ICU mortality following the initial 24 hours after admission.

Outcomes and Implications

The findings of this study highlight the importance behind the format in which we present data to artificial intelligence models. In this case, biomarker representations are a form of data that can improve the predictive abilities of deep learning models.