04: Machine Learning Models for Risk Stratification in Infants with Abnormal Development

Music City Center 

Introduction: To address the gap in knowledge of developmental outcomes of infants ≥34 weeks gestation requiring NICU care, 1183 infants were evaluated. Variables collected for each infant included demographics, delivery course, NICU diagnoses and management, and developmental screening results. Of interest was which variables predict abnormal developmental screening. Methods: Traditional logistic regression analyses results were unstable because of the multicollinearity problem between the large number of predictor variables. CART analysis was ultimately used. Other machine learning methods, including random forest, TreeNet gradient boosting, MARS and regularized logistic regression methods, were also used for comparison. Results: NICU variables selected as most important in the CART analysis were PO feeding, number of medications prescribed, and number of follow-up appointments at NICU discharge. Other four machine learning models produced similar results to the CART model with slightly higher AUC and specificity, but lower sensitivity. Conclusion: Machine learning algorithms can provide useful tool in health care to assist providers in making complex clinical decisions.

Machine learning

Prediction

Health care 

Section on Statistical Consulting