Regional Brain Age is Decreased in Children with Sickle Cell Anemia.
Authors
Affiliations (6)
Affiliations (6)
- Washington University in St. Louis, Saint Louis, Missouri, United States.
- Washington University in St. Louis School of Medicine, Cleveland, Ohio, United States.
- Washington University in St. Louis, St. Louis, Missouri, United States.
- Washington University, Saint Louis, Missouri, United States.
- Washington University School of Medicine, St. Louis, Missouri, United States.
- Washington University, St. Louis, Missouri, United States.
Abstract
DeepBrainNet, a machine learning tool, uses MRI to predict an individual's brain age, allowing calculation of the brain age gap (predicted - chronological age) for use as a biomarker of brain health. We tested the DeepBrainNet tool in sickle cell anemia (SCA) on 210 brain MRIs from 130 children (90 SCA, 40 control) to investigate the hypothesis that children with SCA would have a younger predicted brain age than healthy controls as an index of aberrant brain development. DeepBrainNet estimates brain age from 80 axial slices of each brain with the median utilized as global brain age. Global brain age gap was not different between SCA and controls (p = 0.956). However, the estimated difference in regional brain age gap between SCA without stroke and controls was -1.26 (95% CI -2.34, -0.19, p=0.021) years after adjusting for chronological age, indicating regionally reduced brain age in SCA. The DeepBrainNet model architecture was modified to create a SCA classifier that distinguished axial brain slices from controls versus SCA without stroke history (accuracy=0.69, AUC=0.7). We conclude that there is a regional decrease in brain age in children with SCA without stroke compared controls, suggesting altered brain development. Furthermore, DeepBrainNet can be used to train a classifier to accurately classify the presence of SCA in children without infarcts with only the input of clinically available MRI sequences. These data highlight that machine learning tools could potentially be used to improve upon risk prediction algorithms and assessment of treatment effect with further development.