An interpretable deep learning algorithm for dynamic early warning of posttraumatic hemorrhagic shock based on noninvasive parameter. (August 2022)

• Record Type: Journal Article
• Authors: Zhang, Guang; Xu, JiaMeng; Wang, Huiquan; Yu, Ming; Yuan, Jing
• Is Part Of: Biomedical signal processing and control. Volume 77(2022)
• Publication Date: 2022-08
• Subjects: Dynamic early warning model -- Deep learning -- Post-traumatic hemorrhagic shock -- Non-invasive parameters -- Interpretability analysis; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103779
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: In this paper, a dynamic early warning algorithm is proposed, which can process the flow data of patients in hospital and realize dynamic early warning. The dynamic early warning model established in this paper reserves a lead time so that medical staff have enough time to intervene patients after discovering the high incidence of post-traumatic hemorrhagic shock. This paper proposes and verifies the feasibility of using only non-invasive parameters to establish dynamic early warning model. This article verifies the feasibility of interpretable analysis of patient flow data. Abstract: Objectives: Hemorrhage and hemorrhagic shock are common causes of death after an acute trauma. The mortality from hemorrhagic shock can be significantly reduced through the prophylactic administration of red blood cells or the use of 100% mechanical ventilation. In this study, a dynamic early warning system based on non-invasive parameters was developed in this study and was evaluated using deep learning algorithms, aiming to predict the incidence of hemorrhagic shock in patients over the next 4–8 h. Design: An observational cohort study. Setting: The data set was collected from three data sets from 210 different hospitals in the United States and the Netherlands. One of them was publicly available for model development and two were used for testing. Patients: 9659 patients from eICU database, 2942 patients from MIMICIII database, 1055 patients from AmsterdamUMC database. Interventions: None. Measurements and main results: A deep learning model, constructed using Convolutional Neural Networks (CNN), Bi-directional Long-Short Term Memory (BiLSTM), and Attention Mechanism, was employed to dynamically predict the incidence of hemorrhagic shock in patients over the next 4–8 h based on 4 h patient data. A dynamic early warning model was trained with non-invasive data from the eICU database. The test set, composed of the data of the MIMICIII and AmsterdamUMC databases, was utilized for cross-database validation of model performance, and the AUC value reached 0.8104. When the model parameters were updated with 5% of data, the AUC value increased to 0.8591 in the test set composed of other data. The results from the interpretability analysis showed that gender was crucial for judging whether hemorrhagic shock would occur in patients following trauma. Conclusions: The deep learning model was used to validate the feasibility of constructing a dynamic early warning model for post-traumatic hemorrhagic shock based on non-invasive parameters. The interpretability analysis results were consistent with clinical study results.