P2C201: Predicting Neurodevelopmental Outcome in Very Low Birth Weight Infants from MRI Utilizing a Machine Learning Model with Volumetrics Extracted from Infant Freesurfer

Background: Infant Freesurfer was introduced several years ago to solve unmet needs of heavily specialized infant brain. This automated algorithm changed the volumetrics which had been remained research tools that required specialist infrastructures and expertises easier and faster to approach to clinicians.
Purpose of the study is to develop a new model to predict long term neurodevelopmental outcome of preterm infants utilizing automated volumetry extracted from term equivalent age MRI, diffusion tensor imaging and clinical information.

Methods: Preterm neonates hospitalized in Severance Children’s Hospital born from January 2012 to December 2019 were consecutively collected. Inclusion criteria were birth weight under 1500 g who completed both term equivalent age (TEA) MRI and Bayley Scales of Infant and Toddler Development 2nd Edition (BSID-II) at 18 months of corrected age (CA).
The random forest classifier and logistic regression methods were utilized to develop three models. After development, better model was applied to test set. The area under receiver operating curve (AUROC), accuracy, sensitivity, precision and F1 score was evaluated with test set.

Results: Total 150 patient data were enrolled from criteria. For low PDI prediction, random forest classifier was utilized. The AUROC of model with clinical variables, MR volumetry, and both of clinical variables and MR volumetry were 0.8435. 0.7281 and 0.9297, respectively. For low MDI prediction, logistic regression model was selected. The AUROC of model using clinical variables, MR volumetry, and both of clinical variables and MR volumetry were 0.7483, 0.7052 and 0.7755, respectively. The model using both of the clinical variables and MR volumetry showed the highest AUROC in both PDI and MDI prediction

Conclusion: In conclusion, a new prediction model utilizing automated volumetry algorithm significantly differentiates long term psychomotor developmental outcome of preterm infants. This model could give additional information to conventional MRI without expertized efforts in clinical fields.