64 - Clinical data predicting neonatal and 2-year outcomes following hypoxic ischemic encephalopathy

Poster Number: 64
Publication Number: 64.429

Rina Bao, Boston Children's Hospital, Cambridge, MA, United States; Yangming Ou, Boston Children's Hospital; Harvard Medical School, Boston, MA, United States

Background: Hypoxic ischemic encephalopathy (HIE) is a brain injury that occurs in 1~5/1000 term-born neonates. Despite therapeutic hypothermia reducing mortality and morbidity, ~60% of patients still die in neonatal stage, or develop neurocognitive deficits by 2 years of age. HIE outcome prediction is urgently needed, to more timely intervene in targeted patients, and to accelerate therapeutic innovations. However, the accuracy of using clinical variables for outcome prediction remains less understood.

Objective: To quantify the accuracy of using clinical data in predicting neonatal and 2-year neurocognitive outcomes.

Design/Methods: We retrospectively collected clinical and outcome data Boston Children’s (BCH) and Massachusetts General Hospitals (MGH). Clinical data of 72 BCH and 74 MGH patients were studied for neonatal-stage outcome prediction (survival vs deceased). Clinical data of 72 BCH and 55 MGH patients were studied for 2-year outcome prediction (normal vs abnormal, defined in Neonatal Research Network criteria). Clinical data included demographics (sex, gestational age, etc.), birth parameters (head circumference, birth weight, APGAR score), treatment and care (length of NICU stay, seizure onset, etc). Outcome predictions were done by both univariate analysis (U test, Spearman correlation, etc.) and multivariate machine learning methods (SVM, Random Forest, Deep Neural Network). Multivariate machine learning was coupled with feature selection algorithms that automatically selected the most predictive variable set. We split BCH and MGH data into half training and half testing. Accuracy was measured by sensitivity, specificity, and classification rate between predicted and actual outcomes.

Results: Multivariate machine learning with feature selection outperformed univariate analysis and machine learning without feature selection. It achieved a 0.65 sensitivity and 0.95 specificity predicting decease in the neonatal stage. The accuracy for predicting 2-year adverse outcome was ~0.70.

Conclusion(s): Clinical variables better predict HIE outcomes when combined than isolated. Future work will quantify the accuracy of machine learning adding MRI into clinical data to predict HIE outcomes, and further validate the accuracy in multi-site/scanner data.