Prediction of neurodevelopmental outcome in preterm neonates with cerebral MR spectroscopy and DTI using feed-forward neural networks

Zielsetzung We aimed to evaluate if proton magnetic resonance spectroscopy (1H-MRS) and diffusion tensor images (DTI) quantified in very preterm infants (VPIs) at term equivalent age (TEA) enhance the predictive role of conventional MRI for their neurodevelopmental outcome (NDO) at the corrected age of 12 months using feed-forward neural-networks (fNNs).

Material und Methoden From 300 VPIs born before 32 completed gestational weeks who received an MRI scan at TEA between September 2013 and December 2017, 173 were excluded due to missing or poor-quality spectroscopy data and/or missing neurodevelopmental tests at 12 months corrected age. The data sets of 127 VPIs were considered for motor and cognitive development, of whom 13 and 7, respectively were categorized as delayed. We evaluated five metabolite ratios and two DTI characteristics, each in six areas of the brain. We performed a feature selection algorithm for receiving a subset of characteristics that were prevalent for the VPIs with developmental delay. To reduce bias by unbalanced classes, only VPIs that shared approximate values of those prevalent characteristics were considered for further calculations. We finally constructed predictors using fNNs.

Ergebnisse Predictors constructed by fNNs achieved a true positive rate of 85.7% and a positive predictive value of 100% for prediction of cognitive developmental delay, and a true positive rate of 76.9 % and a positive predictive value of 90.9 % for prediction of motor developmental delay.

Schlußfolgerungen 1H-MRS and DTI quantified at TEA in VPIs add to the predictive value of conventional MRI for motor and cognitive development at the corrected age of 12 months. The proposed approach applying fNNs is promising for the use in clinical practice for identifying those VPIs that would mostly benefit from early intervention services.