Predicting Mortality of Intensive Care Patients with Machine Learning Using Electronic Health Record and Non-Invasive Signals
Neissi Shooshtari, Ali Jr (2021)
Neissi Shooshtari, Ali Jr
2021
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2021072316836
https://urn.fi/URN:NBN:fi:amk-2021072316836
Tiivistelmä
This study proposes a novel approach for applying the Electronic Health Record (EHR) data and biomedical signals to predict patient mortality in the Intensive Care Unit (ICU) using machine learning and deep learning models. The study results showed that using a combination of EHR data and waveforms improved the performance of the models compared to using only one of the inputs. A dataset containing 2320 ICU patients was used in this study. A 5-hour data extraction window for EHR data and a 1-hour data extraction window for the waveform were considered. The prediction window was set to 12 hours. Five different models were used in this study. Six vitals along with three non-invasive signals were used and in total, some 67 different features were extracted and used. Area Under Receiver Operating Characteristic Curve (AUROC) and Area Under Precision-Recall Curve (AUPRC) were taken as the metrics. The best performance achieved in this study using both EHR data and wave-forms was an AUROC of 0.877 and an AUPRC of 0.289. The results also showed the model fed by a combination of the EHR data and waveforms outperformed the same models when they were fed only with EHR data by 3% in terms of AUROC and by 10.3% in terms of AUPRC and, the models that were fed only by waveforms by 4.4% in terms of AUROC and 7.03% in terms of AUPRC.