Using Domain Adaptation and Inductive Transfer Learning to Improve Patient Outcome Prediction in the Intensive Care Unit
Date
2023-12-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Predicting patient outcomes in the intensive care unit (ICU) can allow appropriate allocation of resources, minimize costs, and provide better patient care. Machine learning and deep learning models can predict patient outcomes with a high degree of accuracy, but training those models is both data- and resource-intensive. Deep learning models trained on small datasets tend to overfit and generalize poorly, and transfer learning (TL) helps in such situations by leveraging the knowledge learned from pre-trained models. Transfer learning is a machine learning technique where a model pre-trained on source task is adapted for a different but related target task. Here, source task is trained with a large dataset whereas a small dataset is sufficient for training target task. Notably, TL is widely used in medical image analysis and natural language processing, but it is uncommon in electronic health record (EHR) analysis. Within the TL literature, domain adaptation (DA) is most common, whereas inductive transfer learning (ITL) is rare. This study explores both DA and ITL using EHR data. To investigate the effectiveness of these TL models, we compared them with baseline models of logistic regression (LR), lasso regression, and fully connected neural networks (FCNN) in the prediction of 30-day mortality, acute kidney injury (AKI), hospital length of stay (H_LOS), and ICU length of stay (ICU_LOS). We used two cohorts: (1) eCritical, a multicenter ICU data linked with administrative data from ICUs in Alberta, Canada between March 2013 and December 2019, which has 55,689 unique admission records from 48,672 unique patients admitted to 15 medical-surgical ICUs, and (2) MIMIC-III, a single-center publicly available ICU dataset from Boston, USA between 2001 and 2012. The first admission of adult patient records with more than 24-hour ICU stays were included in this retrospective study. We included common features from both the cohorts. Random data subsets of training data, ranging from 1% to 75%, and the full dataset were used to compare the performances of DA and ITL with FCNN, LR, and lasso. Overall, ITL outperformed baseline FCNN, LR, and lasso models in 55 of the 56 comparisons (7 data subsets, 4 outcomes, and 2 baseline models) using BA and MSE metrics. However, DA models outperformed the baseline models 45 out of the 56 times. ITL performance was comparatively better than DA considering the number of times it outperformed baseline models and the margin with which it outperformed baseline models. Also, in 11 out of the 16 cases (8 of 8 for ITL and 3 of 8 for DA) TL models outperformed baseline models at 1% data subset. This is significant because TL models are useful in data-scarce scenarios. When using EHR data, the similarity of data distributions in source and target domains was crucial, as evident from ITL performing much better than DA, mostly because of the domain mismatch in the two cohorts concerning AKI, H_LOS, and ICU_LOS outcomes. As the pre-trained models will be made available publicly, further research can be conducted with additional outcomes and different cohorts to make these pre-trained models more robust using incremental or cumulative transfer learning. These pre-trained models can be used for predicting patient outcomes at ICU.
Description
Keywords
Transfer Learning, Domain Adaptation, Inductive Transfer Learning
Citation
Mutnuri, M. K. (2023). Using domain adaptation and inductive transfer learning to improve patient outcome prediction in the intensive care unit (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.