Predictive Analytics for Neonatal Health Risk Assessment: A Machine Learning Approach to Early Detection of Critical Health Conditions in Newborns

Infant morbidity and mortality in the first month of life is a significant contributor to infantmorbidity and mortality in the world. In this paper, NeoRisk, a machine learning model topredict the level of neonatal risk (Healthy or At Risk) on a daily basis based on longitudinalmonitoring data of 100 newborns (vital signs, growth metrics, feeding patterns, jaundice values)is presented, and initial results with tabular models (Logistic Regression, Random Forest,XGBoost) showed an almost perfect result (ROC AUC ≈ 1.000). However, when the explicit useof jaundice values in cases of leukocyte counts Removing leaking features systematically therealistic performance of ROC AUC was found in the 0.85-0.94 range. To address linkedphysiological dynamics, a time-series model based on LSTMs and using 7-day historicalsequences was constructed, which plays a crucial role in avoiding the leakage of information andthe need to handle the class imbalance in neonatal predictive analytics. Of the most importantpredictors were jaundice (pre-leakage) and weight change / gestational age (post-leakage).NeoRisk provides a clinically relevant reproducible risk stratification pipeline in the early periodand demonstrates the benefits of longitudinal deep learning strategies in the neonatal care setting.