Using Domain Adaptation and Inductive Transfer Learning to Improve Patient Outcome Prediction in the Intensive Care...

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Prediction
• Electronic Health Record (EHR) Data Analysis

Background:

• Accurate prediction of patient outcomes in the Intensive Care Unit (ICU) is crucial for effective patient care.
• Deep learning models require substantial data and computational resources, often unavailable in real-world clinical settings.
• Transfer learning (TL) offers a solution for data scarcity by leveraging pretrained models, but its application in EHR analysis, particularly inductive transfer learning (ITL), is underexplored.

Purpose of the Study:

• To investigate the efficacy of domain adaptation (DA) and inductive transfer learning (ITL) for EHR-based ICU patient outcome prediction.
• To evaluate the performance of these TL methods under varying degrees of simulated data scarcity.

Main Methods:

• Utilized two large ICU patient cohorts: eCritical (multicenter) and Medical Information Mart for Intensive Care III (single-center).
• Compared DA and ITL models against baseline models (fully connected neural networks, logistic regression, lasso regression).
• Assessed model performance in predicting 30-day mortality, acute kidney injury, and ICU/hospital length of stay using data subsets from 1% to 75%.

Main Results:

• Inductive transfer learning (ITL) models significantly outperformed baseline models in 55 out of 56 comparisons (P < .001).
• Domain adaptation (DA) models also outperformed baseline models in 45 out of 56 comparisons (P < .001).
• ITL models showed superior performance over DA models, particularly when trained on very small data subsets (e.g., 1% data).

Conclusions:

• Transfer learning (TL) models, especially ITL, are highly effective for ICU patient outcome prediction in data-scarce environments.
• The study provides a framework for estimating prediction performance with and without TL across different data scarcity levels.
• Publicly available pretrained models can accelerate future research and model development for diverse ICU cohorts and outcomes.