Comparing machine learning techniques for neonatal mortality prediction: insights from a modeling competition.

Area of Science:

• Neonatal Medicine
• Machine Learning in Healthcare
• Clinical Informatics

Background:

• Predicting mortality in neonatal intensive care units (NICUs) is challenging due to complex patient data.
• Machine learning (ML) presents an opportunity for improved risk stratification in NICU settings.

Purpose of the Study:

• To compare the predictive performance of various ML models for NICU mortality.
• To assess model performance in a team-based modeling competition format.

Main Methods:

• Five neonatologist-led teams applied ML techniques (logistic regression, CatBoost, neural networks, random forest, XGBoost) to a dataset of over 6,000 NICU admissions.
• Models predicted mortality risk using demographic, clinical, heart rate, and oxygen saturation data.
• Model performance was evaluated using the area under the receiver operator characteristic curve (AUC).

Main Results:

• Logistic regression achieved the highest AUC on test data, outperforming more complex models.
• Audience preference favored a complex model (CNN) for perceived real-world applicability.
• Teams utilized diverse strategies for feature selection, hyperparameter tuning, and model evaluation.

Conclusions:

• Model complexity does not guarantee superior predictive performance in NICU mortality prediction.
• The choice of modeling approach should prioritize data characteristics, interpretability, and team expertise.
• This study advocates for a thoughtful selection of ML methods rather than solely pursuing complexity.