Early postnatal brain volumes and two-year cognitive outcomes in preterm infants: An exploratory machine learning study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Child and Adolescent Psychiatry/Psychology, Erasmus MC Sophia Children's Hospital, Rotterdam, the Netherlands. Electronic address: [email protected].
- Department of Pediatric Surgery, Erasmus MC Sophia Children's Hospital, Rotterdam, the Netherlands.
- Department of Neonatal and Pediatric Intensive Care, Division of Neonatology, Erasmus MC - Sophia Children's Hospital, Rotterdam, the Netherlands.
- Department of Child and Adolescent Psychiatry/Psychology, Erasmus MC Sophia Children's Hospital, Rotterdam, the Netherlands.
- Department of Neonatology, Division Woman and Baby, University Medical Center Utrecht, Wilhelmina Children's Hospital, the Netherlands.
- Department of Child and Adolescent Psychiatry/Psychology, Erasmus MC Sophia Children's Hospital, Rotterdam, the Netherlands; Institute of Language Sciences, Utrecht University, Utrecht, the Netherlands.
Abstract
Premature birth is associated with an increased risk of cognitive and motor developmental problems. Early prediction of these outcomes is crucial for timely intervention and adequate follow-up. This study aimed to predict cognitive and motor outcomes in preterm-born infants using brain volumetric features in combination with clinical and demographic perinatal characteristics, and to compare the performance of three different machine learning algorithms. We acquired quantitative structural brain MRI at approximately 30 weeks postmenstrual age (PMA) and clinical and demographic birth characteristics including gestational age, sex, birthweight, and maternal education level. Machine learning techniques were applied to predict cognitive (N = 63) and motor (N = 41) outcomes at two years corrected age, assessed using the Bayley Scales of Infant and Toddler Development III, in infants born before 30 weeks gestation. We evaluated the performance of logistic regression, linear support vector classifier (linear SVC), and k-nearest neighbour (KNN). Employing volumetric features only, KNN achieved a validation accuracy of 0.64, showing modest predictive power for cognitive outcomes. Adding clinical characteristics did not significantly improve accuracy. The models did not perform better than chance when predicting motor outcomes. Volumetric features demonstrated modest predictive ability for cognitive outcomes, and the addition of clinical features did not significantly improve performance. The low to moderate overall validation accuracy underscores, however, the challenges faced in the clinical application of machine learning in neuroimaging. The reliability and clinical applicability of predictions of outcomes from advanced neuroimaging data will largely depend on the quantity and quality of the data, methodological standardisation, and the generalisability of the prediction models.