60 - A machine learning-based model for early mortality and poor outcome prediction in preterm infants

Poster Number: 60
Publication Number: 60.429

Yun-Hsiang Yang, NCKUH, Tainan, Tainan, Taiwan (Republic of China); Yuh-Jyh Lin, National Cheng-Kung U Hospital, Tainan, Tainan, Taiwan (Republic of China); Ts-Ting Wang, Ditmanson Medical Foundation Chia-Yi Christian Hospital, Tainan, Tainan, Taiwan (Republic of China); Yi-Han Su, national cheng kung university hospital, tainan, Tainan, Taiwan (Republic of China); WEIYING CHU, National Cheng Kung University Hospital, Tainan, Tainan, Taiwan (Republic of China); Wei-Ting Lin, National Cheng Kung University Hospital, Tainan City CentralWest distinction, Tainan, Taiwan (Republic of China); Yen Ju Chen, National Cheng Kung University hospital, Tainan, Tainan, Taiwan (Republic of China); Yu-Shan Chang, National Cheng Kung University Hospital, Tainan, Tainan, Taiwan (Republic of China); Yung-Chieh Lin, National Cheng Kung University Hospital, Tainan, Tainan, Taiwan (Republic of China); Chyi-Her Lin, E-Da Hospital, Kaohsiung, Kaohsiung, Taiwan (Republic of China)

Background: Prediction of mortality and severe intra-ventricular hemorrhage (IVH) is crucial both for the individual preterm infant (e.g., to adjust postnatal care and to advise parents for decision-making) and to compare risk-adjusted outcomes between groups of preterm infants.

Objective: We aimed to use machine learning techniques to build effective models for predicting early mortality, severe intra-ventricular hemorrhage, and early poor outcomes (occurrence of either two) among preterm infants.

Design/Methods: We collected data on very-low-birth-weight preterm infants born between 2016 and 2022 obtained from the Taiwan Neonatal Network. Twenty-four attributes including initial management in the delivery room, maternal and neonatal characteristics were used in our analysis. The target outcomes were early mortality, severe IVH, and early poor outcome. The performance of the model was evaluated using the area under the receiver operating characteristic curve (AUROC).

Results: A total of 6,730 infants were enrolled. The logistic regression algorithm yielded the highest AUROC (e.g., 0.881 when predicting early mortality, 0.827 when predicting severe IVH, and 0.850 when predicting early poor outcome). After attribute selection, the AUROC of these simplified models, namely Neural Network, Logistic Regression, and Gradient Boosting models were favorable (e.g., 0.873, 0.866 and 0.865 when predicting early mortality, 0.819, 0.821 and 0.815 when predicting severe IVH, and 0.840, 0.840 and 0.837 when predicting early poor outcomes).

Conclusion(s): We built predictive models using machine learning algorithms to predict mortality, severe IVH, and early poor outcome. The logistic regression model was the most optimal. Through attribute selection, Neural Network, Logistic Regression, and Gradient Boosting models could build simplified estimators for clinical application.
BB77421C-A3FB-42C2-B6D1-8D4BC8F6A05F.jpeg

881A8183-0C61-480F-8544-93C57054C6B6.jpeg

040E8AEE-46F0-462D-9758-5EE8B82DFCED.jpeg