20 - Complexity of Brain MRI Neonatal Research Network Scores in Multivariate Machine Learning Prediction of Hypoxic-Ischemic Encephalopathy Outcomes

Publication Number: 1017.20

Michelle Y. Li, Boston Children's Hospital, Boston, MA, United States; Ankush Kesri, Boston Children's Hospital, Cambridge, MA, United States; Rutvi Vyas, Boston Children's Hospital, Boston, MA, United States; Chuan-Heng Hsiao, Boston Children's Hospital, Boston, MA, United States; Seetha Shankaran, University of Texas at Austin Dell Medical School, Ann Arbor, MI, United States; Abbot R.. Laptook, Women & Infants Hospital of Rhode Island, North Kingstown, RI, United States; Michael Cotten, Duke University School of Medicine, Durham, NC, United States; P. Ellen grant, Boston Children's Hospital, Boston, MA, United States; Yangming Ou, Boston Children's Hospital, Boston, MA, United States

Background: Hypoxic-ischemic encephalopathy (HIE) occurs in 3/1000 neonates in high-income countries due to a lack of blood flow and oxygen to the brain. An NICHD Neonatal Research Network (NRN) score denotes an expert-assigned injury level on MRI, which is associated with death or disability following moderate or severe HIE [1]. Recent expansion from a 6- to 10-level NRN score [2], with further granularity on watershed injury, vascular infarction, or injury to the basal ganglia, thalamus, or internal capsules, aims to better reveal neuroanatomic associations of HIE outcomes [3].

Objective: Compare the accuracy of 11 versions of NRN scores (6-, 10-, or simplified 2-, 3-levels) in predicting 5 HIE outcomes, when combined with clinical variables for multivariate machine learning-based outcome prediction.

Design/Methods: We utilized the Optimizing Cooling data (N=298) [3], using the available 6- and 10-level NRN scores, and 46 clinical variables listed in Table 1. Brain MRIs were scored by consensus of 2 central readers into 6 levels (0, 1A, 1B, 2A, 2B, 3) and 10 levels (0, 1A, 1B, 2A1, 2A2, 2A3, 2B1, 2B2, 2B3, 3). We also converted 6-level consensus scores into 5 versions of 2-level, and 4 versions of 3-level scores (definition and frequency in Table 1). Five outcomes at 18-22 months were predicted: 1. adverse outcome, the primary outcome of the trial, present (moderate-severe disability or death, N=72) or not present (N=226); 2. cerebral palsy present (N=42) or not present (N=228); and 3-5. Bayley-III Cognitive, Motor, and Language scores. We employed a multivariate feature selection with consensus among 3 machine learning predictors (linear & gaussian-kernel support vector machines, and random forest) [4]. The algorithm selected feature subsets with the highest prediction accuracy (sensitivity and specificity for binary outcomes; and mean absolute error (MAE=sum(|predicted-actual|)/N) for continuously-valued outcomes) in 5-fold cross-validations.

Results: For adverse outcome, a 3-level NRN score version (0-1A-1B vs 2A-2B vs 3), combined with clinical variables, exhibited significantly higher prediction accuracy (0.913, Fig 1) than 6- or 10-level scores. For cerebral palsy (Fig 2) and all Bayley-III scores, accuracies from most NRN score versions had no significant difference.

Conclusion(s): When combined with clinical variables in multivariate machine learning, certain 3-level NRN scores predict adverse outcomes, but not cerebral palsy or Bayley scores, significantly more accurately than other versions of NRN scores.

Table 1. 47 variables as input features for outcome prediction and definitions of 6-, 10-, 2-, and 3-level NRN scores with frequencies in each level.
A total of 47 variables (46 clinical and 1 NRN score on neonatal brain MRI) are input features for multivariate machine learning of HIE outcome prediction. We used one NRN score version in each experiment, and there were 55 experiments (11 NRN score versions * 5 outcomes). Simplified NRN score levels were defined by grouping scores into 2 or 3 levels. Acronyms: BGT -- basal ganglia and thalamic; ALIC -- anterior limb of the internal capsule; PLIC -- posterior limb of the internal capsule; WS -- watershed.

Figure 1. Effect of different NRN score versions in predicting adverse outcomes.
Each NRN score version was combined with 46 clinical features as input to multivariate machine learning. The McNemar test was used to evaluate significance. *: p<0.05; **: p<0.01; ***: p<0.001.

Figure 2. Effect of different NRN score versions in predicting cerebral palsy.
Each NRN score version was combined with 46 clinical features as input to multivariate machine learning. The McNemar test was used to evaluate significance. *: p<0.05; **: p<0.01; ***: p<0.001.