Using domain adaptation and inductive transfer learning to improve patient outcome prediction in the intensive care unit

Abstract

Abstract Predicting patient outcomes in the intensive care unit (ICU) can allow for more effective and efficient patient care. Deep learning models are effective in learning from data to accurately predict patient outcomes; however, they require huge amounts of data to train and massive computational power. Transfer learning (TL) helps in scenarios when data and computational resources are scarce. TL is commonly used in medical image analysis and natural language processing but is comparatively rare in electronic health record (EHR) analysis. In medical image analysis and natural language processing, domain adaptation (DA) is the most commonly used TL method in the literature while inductive transfer learning (ITL) is quite rare. This study explores DA as well as rarely researched ITL for predicting ICU outcomes using EHR data. To investigate the effectiveness of these TL models, we compared them with baseline models of fully connected neural networks (FCNN), logistic regression, and lasso regression in the prediction of 30-day mortality, acute kidney injury (AKI), hospital length of stay (H_LOS), and ICU length of stay (ICU_LOS). TL models transfer the knowledge gained while training for the source prediction task on the source domain to improve the prediction performance of the target prediction task on the target domain. Whereas baseline models were trained directly on the target domain for the target prediction task. Two cohorts were used in this study for the development and evaluation. The first was eCritical, a multicenter ICU data linked with administrative data with 55,689 unique admission records from 48,672 unique patients admitted to 15 medical-surgical ICUs in Alberta, Canada, between March 2013 and December 2019. The second was MIMIC-III, a single-center, publicly available ICU dataset from Boston, USA, acquired between 2001 and 2012. Random subsets of training data, ranging from 1% to 75%, as well as the full dataset were used to compare the performances of DA and ITL with FCNN, logistic and lasso regression. Overall, the ITL outperformed baseline FCNN, logistic and lasso regressions in 55 out of the 56 comparisons (7 data subsets, 4 outcomes, and 2 baseline models), whereas DA models outperformed the baseline models in 45 out of 56 cases. 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. In 11 out of 16 cases (8 out of 8 for ITL and 3 out of 8 for DA), TL models outperformed baseline models when trained using the 1% data subset. This is significant because TL models are useful in data-scarce scenarios. The publicly available pre-trained models from this study can be used to predict ICU patient outcomes and serve as building blocks in further research for the development and validation of models in other cohorts and outcomes.