76 - Effect of Neonatal Death Risk Factors on Artificial Intelligence-based Risk Predictive Models

Poster Number: 76
Publication Number: 76.437

Jagjit S. Teji, Ann & Robert H. Lurie Children's Hospital of Chicago, Oak Brook, IL, United States; Jasjit S. Suri, AtheroPoint, Roseville, CA, United States; Suneet K. Gupta, Bennett University, Gr. Noida, Uttar Pradesh, India; Suneet Jain, Bennett University, Bhopal, Madhya Pradesh, India

Background: World Health Organization, WHO, suggests a set of five minimum prenatal are maternal age, place of delivery, mode of delivery, birth weight, and gestational age that can predict the birth outcome.

Objective: Suggest additions to WHO’s minimum prenatal indicators on National Center for Health Statistics (NCHS) 2015 birth-infant death files and observe the performance with incremental feature selection on (NCHS) 2014-2017 birth-infant death files.

Design/Methods: The risk factors from dataset were: Place of birth, maternal age, birthweight (grams), gestational age, fetal presentation, mode of delivery (vaginal or C-section), maternal race, maternal race by Hispanic classification, Apgar score, single or multiple pregnancies, maternal cigarette smoking, congenital anomalies, newborn condition, unmarried, induction of labor, augmentation of labor, antenatal steroid, maternal risk, infertility treatment, and maternal admission to incentive care unit including those recommended by WHO. We used three different ML-based risk predictive algorithms (Logistic Regression (LR), Extreme Gradient Boosting (XGB), and Adaboost (ADB)) and their individual feature selection to study the effects of these risk factors. First, we evaluated the performance of these classifiers on the five WHO indicators and then with additional risk factors identified by the Algorithms to suggest additions in the WHO list. Secondly, we studied the performance of incremental feature selection (top 10%, top 20%, top 30% features, and so on) on the 2014-2017 NCHS files (birth records: 15,807,801, infant deaths: 91,773 (life < 1 year) with (LR) classifier with Mutual Information (MI) feature selection method.

Results: Out of 3,988,733 babies, 15,577 were neonatal deaths. WHO’s minimum set of prenatal indicators with the three ML classifiers (LR, XGB, and ADB) gave area-under-the-curve (AUC) scores of 0.9080, 0.9316, 0.9282, respectively. With the addition of Apgar Score to WHO’s five risk factors, the performance increased by 3.2%. AUC score with XGB increased by 3.2%, and similarly, the AUC score with ADB went up by 3.6%. With incremental feature selection (top 10%, top 20%, top 30% features, and so on) with Mutual Information, there is a 60% increase in AUC scores up to top 40% feature selection and then the AUC score increases gradually. The effect of incremental feature selection can be seen in Table and the ROC curves can be seen in Figure.Conclusion(s): The addition of risk factors to WHO’s five minimum prenatal indicators can help drastically improve death prediction. This is useful in poor countries where all factors cannot be recorded.
Table Incremental Feature Selection.
Figure Effect of incremental feature selection using MI