AI-enhanced hemodynamic prediction of patent ductus arteriosus severity in premature infants
Related Experiment Video
Updated: Oct 31, 2026
Hemodynamic Precision in the Neonatal Intensive Care Unit using Targeted Neonatal Echocardiography
Published on: January 27, 2023
Taeyoung Kim1,2, Jungmin Choi1,3, Myung Jin Chung4,5
1Medical AI Research Center, Samsung Medical Center, Seoul, Republic of Korea.
Background:
Hemodynamically significant patent ductus arteriosus (HSPDA) requires timely assessment in premature very low birth weight (VLBW) infants. This study aimed to develop and validate an artificial intelligence (AI)-based surveillance system for predicting HSPDA and supporting real-time NICU decision-making.
Methods:
We retrospectively analyzed electronic medical record data from VLBW infants with PDA admitted to the Samsung Medical Center NICU between 2013 and 2022. After excluding records with congenital heart disease or missing critical echocardiographic parameters, machine-learning models were trained on 85% of the dataset and tested on the remaining 15%.
Results:
The best-performing model, Extra Trees, achieved an area under the curve (AUC) of 0.829, with a sensitivity, specificity, positive predictive value, and negative predictive value of 0.767, 0.766, 0.694, and 0.825, respectively. For transitional records shifting between hemodynamic states, the AUC, sensitivity, and specificity were 0.700, 0.722, and 0.606, respectively, based on the proposed transition-optimized threshold.
Conclusion:
The AI-based Extra Trees model demonstrated a robust ability to predict the hemodynamic status of PDA in VLBW infants. Thus, it may help NICU clinicians optimize conservative management and improve clinical decision-making.
Impact:
Our AI model achieves 0.829 AUC for predicting hemodynamically significant PDA in premature infants using routine NICU data without requiring echocardiography This study uniquely predicts PDA hemodynamic severity through multimodal clinical data integration, advancing beyond previous binary presence/absence classifications We introduce a transition-optimized threshold that detects dynamic hemodynamic changes, enabling precise monitoring during conservative management The model provides accessible real-time PDA assessment for resource-limited NICUs lacking immediate echocardiographic expertise, optimizing intervention timing and fluid management strategies Our validation across 1033 VLBW infants over 10 years demonstrates robust performance with only 12-h data requirements.
