Predicting Hemorrhage in ICU Patients with Deep Learning Techniques
Date
2024-10-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Hemorrhage is a leading cause of trauma-related mortality, making early prediction of blood transfusion needs critical for timely intervention. This thesis investigates the application of data-driven modeling techniques to predict hemorrhage in intensive care unit (ICU) patients, aiming to improve patient outcomes through early-stage detection. We developed a Long Short-Term Memory (LSTM)-based architecture to model sequential patient data, capturing the dynamic evolution of vital signs and interventions to predict hemorrhage within a 5-hour window. By using irregularly sampled time series data, Our model achieved an Area Under the Curve (AUC) of 0.99, surpassing existing literature, where AUC values typically range from 0.70 to 0.95. We also compared the performance of our LSTM model with a time series Transformer, finding that LSTM outperformed the Transformer architecture in this study. A key contribution of this research is the comparative analysis of imputation methods, evaluating their impact on data distribution and prediction performance. While imputation techniques significantly altered data distribution, their effect on prediction performance was minimal. Additionally, Shapley values were employed to interpret the model, revealing feature contributions that aligned with surgeons’ understanding of hemorrhage mechanisms, further validating the model. To test external validity, we applied the ICU-trained model to prehospital datasets collected locally in collaboration with Shock Trauma Air Rescue Service (STARS). Although differences in data distribution posed challenges to maintaining high performance outside ICU settings, this research highlights the potential of sequential modeling in hemorrhage prediction and paves the way for future improvements in prehospital care.
Description
Keywords
Citation
Ghias, M. (2024). Predicting hemorrhage in ICU patients with deep learning techniques (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.