Predicting Adverse Neonatal Outcomes for Preterm Neonates with Multi-Task Learning.

Area of Science:

• Neonatal Health
• Machine Learning in Medicine
• Predictive Analytics

Background:

• Accurate diagnosis of adverse neonatal outcomes is critical for improving preterm infant survival rates and enabling timely medical interventions.
• Existing machine learning (ML) models often predict single outcomes, potentially overlooking interdependencies and leading to suboptimal performance and overfitting.
• Electronic Health Records (EHRs) offer rich data for developing predictive models in neonatal care.

Approach:

• Analyzed correlations between three common adverse neonatal outcomes.
• Formulated the joint prediction of multiple outcomes as a multi-task learning (MTL) problem.
• Developed an MTL framework with shared hidden layers and task-specific branches for simultaneous outcome prediction.

Key Points:

• The proposed MTL framework effectively predicts multiple adverse neonatal outcomes concurrently.
• Experimental results using EHR data from 121 preterm neonates validate the framework's effectiveness.
• Feature importance analysis provides valuable insights into model interpretability for each predicted outcome.

Conclusions:

• Multi-task learning offers a superior approach to single-task prediction for adverse neonatal outcomes.
• The developed MTL framework demonstrates significant potential for enhancing neonatal intensive care unit (NICU) decision-making.
• This work contributes to more accurate and interpretable ML-based diagnostic tools in neonatology.