468 - Unseen Artificial Intelligence for Neonatal Death Prediction

Poster Number: 468
Publication Number: 468.130

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: Unseen Artificial Intelligence for Neonatal Death Prediction

Objective: Design machine learning systems on dataset A (15.8 M) and the then test their performance on a subset dataset B (3.9 M) which is unseen. Then perform the same experiment by reversing the datasets.

Design/Methods: Two datasets were used, the first one contained birth records of 15,807,801 babies, of which 61,222 were neonatal deaths (life < 28 days) this was from the year 2014-2017 and was obtained with the help of National Center for Health Statistics (NCHS), the second one, subset of 2014-2017 database (for the year 2015) contained birth records of 3,988,733 babies of which 15,577 were neonatal deaths. The features (risk factors) both datasets consisted of place of birth, maternal age, birthweight (gms), gestational age, fetal presentation, model of delivery, maternal race, maternal race by Hispanic classification, Apgar score, single vs. multiple pregnancy, maternal cigarette smoking, congenital anomalies, newborn condition, unmarried, induction of labor, augmentation of labor, antenatal steroid, maternal risk (diabetes, hypertension, previous preterm delivery), infertility treatment, maternal admission to ICU (intensive care unit). A detailed baseline characteristics of the dataset can be seen in Table. The preprocessing involved null value replacement with mean values from the respective columns. Three ML-based risk predictions were implemented (AdaBoost (ADB), Extreme Gradient Boosting (XGB), Logistic Regression (LR)). First the three ML classifiers were trained on the bigger 15.8 M patients’ data, and then these classifiers were used to make predictions on the smaller 3.9 M patients’ data. After this we switched the datasets and trained new ML classifiers on the 3.9 M patients’ data and then used the classifiers to make predictions on 15.8 M patients’ data.

Results: The accuracy and area-under-the-curve (AUC) pairs in Case A (Training on 15.8 M, prediction on 3.9 M) for the three ML ordered models (ADB, XGB, LR) were (99.18%, 0.950), (99.43%, 0.937), and (99.26%, 0.897). In Case B (Training on 3.9 M, prediction on 15.8 M) the accuracy and AUC pairs were (99.61%, 0.952), (99.32%, 0.951) and (99.64%, 0.931). A detailed comparison of ROC curves can be seen in FigureConclusion(s): The results in both the cases were very similar, and the AI systems were able to make highly accurate predictions. AdaBoost was the best performing ML classifier. This test proves that our designed ML systems are highly accurate and can be used in clinical settings for neonatal death predictions and saving lives.
TableBaseline characteristics along with mean values.
FigureLeft: Training on 15.8 M and testing on 3.9M. Right: Training on 3.9 M and testing on 15.8 M.