362 - Combining 47 clinical and neuroimaging variables to predict adverse 18-22-month outcomes for Hypoxic Ischemic Encephalopathy in Neonatal Research Network trials

Publication Number: 362.4550

Yangming Ou, Boston Children's Hospital, Boston, MA, United States; Ankush Kesri, Boston Children's Hospital, Boston, MA, United States; rina bao, Boston Children's Hospital, Boston, MA, United States; Chuan-Heng Hsiao, Boston Children's Hospital, Brookline, MA, United States; Rutvi Vyas, Boston Children's Hospital, Boston, MA, United States; Mohammad Arafat Hussain, Boston Children's Hospital/Harvard Medical School, Quincy, MA, United States; Scott A.. McDonald, RTI International, Raleigh, NC, United States; Jeanette O. Auman, RTI, Durham, NC, United States; Anna N.. Foster, Boston Children's Hospital, Boston, MA, United States; Janet Soul, Boston Children's Hospital, Boston, MA, United States; Erfan Darzidehkalani, Boston Childrens hospital, Boston, MA, United States; Matheus Dorigatti Soldatelli, Boston Children's Hospital, Boston, MA, United States; Seetha Shankaran, The University of Texas at Austin, Ann Arbor, MI, United States; Abbot Laptook, Women & Infants Hospital of Rhode Island, Providence, RI, United States; Michael Cotten, Duke university, Durham, NC, United States; Patricia Ellen. Grant, Boston Children's Hospital, Boston, MA, United States

Background: Hypoxic ischemic encephalopathy (HIE) affects in 1~5/1000 term infants. Despite treatment, 1/3 of neonates with HIE shown adverse outcomes by age 2 years. The Consortium Of Biomarker In Neonatal Encephalopathy (COMBINE, since 2022), aims to predict 2-year adverse outcomes in the neonatal stage, toward improving care in early time windows. We merged 2 HIE trials - Late Hypothermia (2008-16, late/standard cooling or control arms, 21 sites) [1] and Optimizing Cooling (2010-14, standard/longer/deeper cooling arms, 18 sites) [2].

Objective: Develop a multivariate machine learning predictor and quantify the accuracy in predicting 18-22-month adverse outcomes.

Design/Methods: We started from 47 variables based on clinical knowledge (Table 1). Experts scored severity of brain injury in neonatal brain MRI, using the Neonatal Research Network (NRN) scoring system: 0 no injury, 1a, 1b, 2a, 2b, and 3 devastating injuries [3]. Outcome was adverse (N=312, moderate/severe disability, or death), or normal (N=95) by 18-22 months. Our algorithm automatically selected the subset of variables with the highest prediction accuracy compared to other variable combinations, by a consensus of linear support vector machine, gaussian SVM and random forest. Accuracy was defined as percentage of neonates being predicted correctly, in 5-fold cross validation (80% patients training; 20% patients testing; repeated 5 times so every patient has been tested once and only once).

Results: Expert NRN score of neonatal brain MRI alone achieved 0.84 accuracy. When NRN score of MRI was not used, machine learning selected 12 out of 46 clinical variables for a 0.84 accuracy. The highest accuracy, 0.89, was obtained by combining NRN expert score and clinical variables. Of the 9 variables finally selected to predict outcomes, the most predictive variables were NRN expert score of MRI, treatment variables (Sarnat scores at discharge and post-treatment, intubation), neonatal characteristics (1min Apgar, gestational age), maternal demographics (age, race), and delivery mode. In contrast, [4] found 0.85 accuracy predicting motor impairment at 1-2 years in 117 infants, combining MRI injury in putamen/globus pallidus, gestational age, and umbilical cord pH.

Conclusion(s): Expert NRN score of neonatal brain MRI, treatment, neonatal characteristics, maternal demographics, and delivery mode jointly predicted 18-22-month adverse outcomes with 0.89 accuracy. Future work will add more sophisticated MRI analysis, more comprehensive clinical variables, and more tests of accuracy stability across multi-site data.

Table 1
Table 1. The 47 clinical and MRI variables we have included as predictors in this study.

Fig 1
Fig 1. Accuracies and auto-selected variables in three experiments for outcome prediction.