A Novel Graph Convolutional Network for Predicting Cognitive Deficits in Very Preterm Infants
Poster No:
1466
Submission Type:
Abstract Submission
Authors:
Hailong Li1,2, Junqi Wang1, Zhiyuan Li1,2, Kim Cecil1,2, Mekibib Altaye1,2, Jonathan Dillman1,2, Nehal Parikh1,2, Lili He1,2
Institutions:
1Cincinnati Children's Hospital Medical Center, Cincinnati, OH, 2University of Cincinnati, Cincinnati, OH
First Author:
Hailong Li, PhD
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Co-Author(s):
Zhiyuan Li
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Kim Cecil
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Mekibib Altaye, PhD
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Jonathan Dillman
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Nehal Parikh, DO, MS
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Lili He, PhD
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Cincinnati Children's Hospital Medical Center|University of Cincinnati
Cincinnati, OH|Cincinnati, OH
Introduction:
Very preterm (VPT) infants (born at or less than 32 weeks gestational age) are at high risk for adverse cognitive deficits.(Linsell et al. 2018) Given the known benefits of early interventions, accurate prediction soon after birth are urgently needed for at-risk VPT infants. Several studies have applied the brain structural connectome (SC) derived from diffusion tensor imaging (DTI) to predict cognitive deficits. (Kawahara et al. 2017; He et al. 2021; Girault et al. 2019) However, none of these models are specifically designed for graph-structured data, and thus, potentially miss certain topological information conveyed in the brain SC. In this work, we developed graph convolutional network (GCN) models (Kipf and Welling 2016) to learn the SC acquired at term-equivalent age as a graph for early prediction of cognitive deficits at 2 years corrected age in VPT infants. The supervised contrastive learning (SCL) technique(Khosla et al. 2020) is applied to mitigate the impacts of the data scarcity problem. We hypothesize that SCL will enhance GCN models for early prediction of cognitive deficits in VPT infants using the SC.
Methods:
This IRB-approved study utilized a regional VPT infant cohort from Cincinnati Infant Neurodevelopment Early Prediction Study (CINEPS) that includes a total of 393 VPT infants. DTI and T2-weighted MRI data were acquired on 3T MRI scanner (Philips Ingenia) for all infants between 39- and 44-weeks postmenstrual age. VPT infants received standardized Bayley Scales of Infant and Toddler Development, Third Edition (Bayley-III) during the 2-year corrected age follow-up visit. We dichotomize the subjects into high-risk (≤85) and low-risk (>85) groups for developing long-term cognitive deficits. We preprocessed MRI data using the developing Human Connectome project (dHCP) processing pipeline (Bastiani et al. 2019). We used the dHCP neonatal brain atlas to define 82 ROIs, and considered the mean fractional anisotropy over all fiber tracts between the two brain ROIs as the structural connectivity (i.e., edges) of each SC. This results in an adjacency matrix with a size of 82 × 82.
An overview is illustrated in Figure 1. We first augment a given subject's brain SC into multiple ones via a random edge perturbation approach. (Figure 1(A)) The newly generated samples are assigned with the same class label as the original subjects. Next, we trained a graph encoder and an embedding projector to accomplish a pretext contrasting "pull-push" task, which repeatedly pulls together a random subject (i.e., the anchor) and subjects of the same class as the anchor, and pushes apart the anchor and subjects of different classes. (Figure 1(B)) Finally, we develop a GCN model for predicting the risk of cognitive deficits by reusing the pre-trained graph encoder. Figure 1(C) We used a stratified random split strategy to separate subjects into training (60%), validation (20%), and testing (20%) sets. We compared our SCL-GCN model with multiple peer competing models (Kipf and Welling 2016; Hoffer and Ailon 2015; Chen et al. 2020; He et al. 2020). We calculated accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).
An overview is illustrated in Figure 1. We first augment a given subject's brain SC into multiple ones via a random edge perturbation approach. (Figure 1(A)) The newly generated samples are assigned with the same class label as the original subjects. Next, we trained a graph encoder and an embedding projector to accomplish a pretext contrasting "pull-push" task, which repeatedly pulls together a random subject (i.e., the anchor) and subjects of the same class as the anchor, and pushes apart the anchor and subjects of different classes. (Figure 1(B)) Finally, we develop a GCN model for predicting the risk of cognitive deficits by reusing the pre-trained graph encoder. Figure 1(C) We used a stratified random split strategy to separate subjects into training (60%), validation (20%), and testing (20%) sets. We compared our SCL-GCN model with multiple peer competing models (Kipf and Welling 2016; Hoffer and Ailon 2015; Chen et al. 2020; He et al. 2020). We calculated accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).
Results:
After data exclusion, we have a total of 282 VPT infants (gestational age at birth 29.2 ± 2.5 weeks, 142 (50.4%) male) with cognitive assessment. The proposed SCL-GCN model achieved a mean AUC of 0.75 on cognitive deficit prediction, significantly higher than Triplet-GCN (0.73, p<0.001) and GCN (0.70, p<0.001). (Table 1) Our SCL-GCN model also achieved 73.1%, 65.9%, and 76.9% on accuracy, sensitivity, and specificity, respectively.
Conclusions:
In a large cohort of VPT infants, we demonstrated that the SCL-GCN model achieved a mean AUC of 0.75 for predicting cognitive deficits. Our results support our hypothesis that the SCL technique is able to enhance the GCN model in our prediction tasks. We also demonstrated that the proposed model outperformed several competing models.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural) 2
Diffusion MRI Modeling and Analysis
Methods Development
Keywords:
Development
Informatics
Machine Learning
PEDIATRIC
Other - Connectome
1|2Indicates the priority used for review
Provide references using author date format
Chen, M., H. Li, J. Wang, W. Yuan, M. Altaye, N. A. Parikh and L. He (2020). "Early Prediction of Cognitive Deficit in Very Preterm Infants Using Brain Structural Connectome With Transfer Learning Enhanced Deep Convolutional Neural Networks." Front Neurosci 14: 858.
Girault, J. B., B. C. Munsell, D. Puechmaille, B. D. Goldman, J. C. Prieto, M. Styner and J. H. Gilmore (2019). "White matter connectomes at birth accurately predict cognitive abilities at age 2." NeuroImage 192: 145-155.
He, L., H. Li, M. Chen, J. Wang, M. Altaye, J. R. Dillman and N. A. Parikh (2021). "Deep Multimodal Learning From MRI and Clinical Data for Early Prediction of Neurodevelopmental Deficits in Very Preterm Infants." Front Neurosci 15: 753033.
He, L., H. Li, J. Wang, M. Chen, E. Gozdas, J. R. Dillman and N. A. Parikh (2020). "A multi-task, multi-stage deep transfer learning model for early prediction of neurodevelopment in very preterm infants." Sci Rep 10(1): 15072.
Hoffer, E. and N. Ailon (2015). Deep metric learning using triplet network. Similarity-Based Pattern Recognition: Third International Workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, Springer.
Kawahara, J., C. J. Brown, S. P. Miller, B. G. Booth, V. Chau, R. E. Grunau, J. G. Zwicker and G. Hamarneh (2017). "BrainNetCNN: Convolutional neural networks for brain networks; towards predicting neurodevelopment." Neuroimage 146: 1038-1049.
Khosla, P., P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu and D. Krishnan (2020). "Supervised contrastive learning." Advances in neural information processing systems 33: 18661-18673.
Kipf, T. N. and M. Welling (2016). "Semi-supervised classification with graph convolutional networks." arXiv preprint arXiv:.02907.
Linsell, L., S. Johnson, D. Wolke, H. O'Reilly, J. K. Morris, J. J. Kurinczuk and N. Marlow (2018). "Cognitive trajectories from infancy to early adulthood following birth before 26 weeks of gestation: a prospective, population-based cohort study." Arch Dis Child 103(4): 363-370.