Predicting mortality risk for preterm infants using deep learning models with time-series vital sign data

Area of Science:

• Neonatal Medicine
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Extremely preterm birth poses a significant mortality risk.
• Current prediction models rely on static variables, limiting real-time accuracy.
• Vital sign time-series data offer potential for more precise outcome prediction in Neonatal Intensive Care Units (NICUs).

Purpose of the Study:

• To develop a novel deep learning model for real-time prediction of mortality risk in preterm infants during initial NICU hospitalization.
• To integrate time-series vital sign data with static clinical variables for enhanced predictive power.
• To address limitations of existing static models in capturing dynamic physiological changes.

Main Methods:

• Development of DeepPBSMonitor, a deep learning model designed to process both time-series vital signs and fixed clinical variables.
• Implementation of techniques to handle noise and imbalanced data inherent in clinical datasets.
• Comparative evaluation of DeepPBSMonitor against other predictive approaches using real-world infant data.

Main Results:

• DeepPBSMonitor demonstrated superior performance in predicting mortality risk.
• Achieved high performance metrics: accuracy of 0.888, recall of 0.780, and Area Under the Curve (AUC) of 0.897.
• The model effectively integrated dynamic vital sign data and static variables.

Conclusions:

• The developed DeepPBSMonitor model is effective for real-time mortality risk prediction in preterm infants in the NICU.
• This novel approach shows promise for improving clinical decision-making and patient outcomes.
• Deep learning offers a powerful tool for leveraging complex vital sign data in neonatal critical care.