Related Experiment Video
Updated: Jul 9, 2025
14:48
Generation of Dynamical Environmental Conditions using a High-Throughput Microfluidic Device
Published on: April 17, 2021
4.1K
Improving mixed-integer temporal modeling by generating synthetic data using conditional generative adversarial
Alireza Rafiei1, Milad Ghiasi Rad2, Andrea Sikora3
1Department of Computer Science and Informatics, Emory University, Ste. W302, 400 Dowman Dr., Atlanta, GA, 30322, USA.
Computers in Biology and Medicine
|November 27, 2023
Summary
Integrating synthetic data with real ICU medication data significantly improved machine learning model predictions for fluid overload. This approach enhances model performance and sensitivity, offering a promising solution for critical care outcomes.
Area of Science:
- Critical Care Medicine
- Data Science
- Machine Learning
Background:
- Mixed-integer temporal data, common in ICU medication use, challenges predictive model performance.
- Accurate prediction of fluid overload in critically ill patients is crucial for effective treatment.
Purpose of the Study:
- To pilot test the integration of synthetic data with existing complex ICU medication data.
- To enhance machine learning model prediction accuracy for fluid overload.
Main Methods:
- A retrospective cohort study of ICU patients admitted for ≥ 72 hours.
- Development of four machine learning algorithms to predict fluid overload.
- Generation of synthetic data using SMOTE and CTGAN, combined with original data.
- Training models using a stacking ensemble technique with a meta-learner.
Main Results:
- Training on combined synthetic and original datasets improved predictive model performance compared to using original data alone.
- The meta-model achieved an AUROC of 0.83, significantly enhancing sensitivity.
- The meta-learner effectively balanced performance metrics and improved minority class identification.
Conclusions:
- Synthetic data integration is a novel approach for ICU medication data, enhancing machine learning for fluid overload prediction.
- This methodology shows potential for improving predictions of other critical care outcomes.
- Meta-learners can optimize trade-offs between performance metrics for better clinical predictions.
