Domain Adaptation Using Convolutional Autoencoder and Gradient Boosting for Adverse Events Prediction in the...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Over 5 million patients annually admitted to US ICUs face high mortality from cardiovascular failure, multi-organ failure, and sepsis.
• Existing data-driven models for predicting adverse ICU events often rely on single-database features, limiting generalizability.
• Domain adaptation techniques are crucial for improving the robustness of predictive models across different healthcare datasets.

Purpose of the Study:

• To propose and evaluate a novel data analysis method combining gradient boosting and convolutional autoencoder (CAE) for predicting ICU mortality and readmission.
• To explore the effectiveness of domain adaptation for enhancing predictive model performance across different ICU patient datasets.
• To identify key temporal and non-temporal features influencing adverse event prediction in the ICU.

Main Methods:

• Retrospective data analysis using patient records from the MIMIC-II database and a local CHOA database.
• Application of gradient boosting for mortality and readmission prediction tasks after novel data imputation.
• Utilizing convolutional autoencoder (CAE) for feature extraction and exploring domain adaptation across datasets.

Main Results:

• Gradient boosting demonstrated effectiveness in predicting both ICU mortality and readmission following data imputation.
• Key temporal and non-temporal features influencing prediction tasks were identified using gradient boosting.
• While CAE alone showed limited feature extraction effectiveness on a single dataset, domain adaptation with CAE across two datasets yielded promising results.

Conclusions:

• The proposed method, integrating gradient boosting with domain adaptation via CAE, shows potential for improving adverse event prediction in intensive care units.
• Novel data imputation techniques enhance the efficacy of gradient boosting for predicting ICU mortality and readmission.
• Domain adaptation is a valuable strategy for developing more generalizable and robust predictive models in critical care settings.