Exploration of brain-based prediction of early developmental outcomes using deep learning
Poster No:
1419
Submission Type:
Abstract Submission
Authors:
Haitao Chen1,2, Jonathan Stubblefield3, Xiuzhen Huang4,3, Emil Cornea5, John Gilmore5, Wei Gao1,2
Institutions:
1Biomedical Imaging Research Institute (BIRI), Cedars-Sinai Medical Center, Los Angeles, CA, 2Department of Bioengineering, University of California at Los Angeles, Los Angeles, CA, 3Department of Computer Science, Arkansas State University, Jonesboro, AR, 4Department of Computational Biomedicine, Cedars-Sinai Medical Center, Los Angeles, CA, 5Departments of Psychiatry, University of North Carolina Chapel Hill, Chapel Hill, NC
First Author:
Haitao Chen
Biomedical Imaging Research Institute (BIRI), Cedars-Sinai Medical Center|Department of Bioengineering, University of California at Los Angeles
Los Angeles, CA|Los Angeles, CA
Biomedical Imaging Research Institute (BIRI), Cedars-Sinai Medical Center|Department of Bioengineering, University of California at Los Angeles
Los Angeles, CA|Los Angeles, CA
Co-Author(s):
Xiuzhen Huang
Department of Computational Biomedicine, Cedars-Sinai Medical Center|Department of Computer Science, Arkansas State University
Los Angeles, CA|Jonesboro, AR
Department of Computational Biomedicine, Cedars-Sinai Medical Center|Department of Computer Science, Arkansas State University
Los Angeles, CA|Jonesboro, AR
Wei Gao
Biomedical Imaging Research Institute (BIRI), Cedars-Sinai Medical Center|Department of Bioengineering, University of California at Los Angeles
Los Angeles, CA|Los Angeles, CA
Biomedical Imaging Research Institute (BIRI), Cedars-Sinai Medical Center|Department of Bioengineering, University of California at Los Angeles
Los Angeles, CA|Los Angeles, CA
Introduction:
One major goal of developmental neuroimaging research is to build imaging-based models to predict later developmental outcomes so we can identify risks at the earliest timepoint for better intervention. In this study, we aimed at exploring the possibility of predicting 1-/2-year Mullen composite and 4-year IQ scores from earlier/concurrent resting-state functional connectivity maps during the first four years of life using deep learning approaches, specifically AlexNet-based 3D convolutional neural networks [1].
Methods:
623 subjects from the Early Brain Development Study (EBDS, 587 with 1-year Mullen composite (MCOMP_1) score, 479 with 2-year Mullen composite (MCOMP_2) score and 364 with 4-year IQ score(IQ_4) ) with successful resting-state functional MRI (rsfMRI) scans on at least one of the four timepoints (i.e., three-week (MCOMP_1: n=347; MCOMP_2: n=274; IQ_4: n=229), one-year (MCOMP_1: n=348; MCOMP_2: n=269; IQ_4: n=202), two-year (MCOMP_1: n=256; MCOMP_2: n=244; IQ_4: n=167) and four-year (MCOMP_1: n=138; MCOMP_2: n=135; IQ_4: n=125)) were retrospectively identified and included in this study. All rsfMRI datasets underwent standard preprocessing including global signal regression and scrubbing for motion correction. All datasets were registered to the same two-year-old template space for analysis [2].
Ten seed-based functional-connectivity (FC) maps [3] were generated from preprocessed rsfMRI BOLD signals for each age group and served as the input for prediction of later behavioral outcomes (1-year/2-year Mullen composite or 4-year IQ) separately. Behavioral outcomes were classified into 3 classes: High class (score>mean+std), Low class (score<mean-std) and Middle class (the rest). Each task group was split into 3 parts with balanced demographics, where two-thirds of the sample served as training set and one-third served as testing set. AlexNet-based 3D convolutional neural network (Figure 1) was then trained on the training set using default hyperparameters (batch size=16, learning rate=.001) for 400 epochs for the classification task, and then evaluated on the testing set. Specifically, this model was adapted from AlexNet structure [1], where several convolutional layers with batch normalization layers, ReLU activations and max pooling layers were followed by flatten layer and several fully connected dense layers with dropouts.
Ten seed-based functional-connectivity (FC) maps [3] were generated from preprocessed rsfMRI BOLD signals for each age group and served as the input for prediction of later behavioral outcomes (1-year/2-year Mullen composite or 4-year IQ) separately. Behavioral outcomes were classified into 3 classes: High class (score>mean+std), Low class (score<mean-std) and Middle class (the rest). Each task group was split into 3 parts with balanced demographics, where two-thirds of the sample served as training set and one-third served as testing set. AlexNet-based 3D convolutional neural network (Figure 1) was then trained on the training set using default hyperparameters (batch size=16, learning rate=.001) for 400 epochs for the classification task, and then evaluated on the testing set. Specifically, this model was adapted from AlexNet structure [1], where several convolutional layers with batch normalization layers, ReLU activations and max pooling layers were followed by flatten layer and several fully connected dense layers with dropouts.
Results:
Among all prediction tasks (neonates/one-year-olds FC for MCOMP_1, neonates/one-year-olds/two-year-olds FC for MCOMP_2, neonates/one-year-olds/two-year-olds/four-year-olds FC for IQ_4) using ten brain functional-connectivity networks separately, four classification tasks showed relatively strong prediction power ( testing set balanced accuracy >= 0.60 and/or AUC >= 0.60): visual-two brain network in neonates predicted 1-year Mullen composite score, sensory-motor brain network in neonates predicted 2-year Mullen composite score, visual-three brain network in two-year-olds correlated 2-year Mullen composite score, and visual-two brain network in two-year-olds predicted 4-year IQ score. Detailed training/testing curve, confusion matrices and testing metrics are shown in Figure 2 for these four classification tasks. Note that Low and High classes were combined as Outlier class only in calculating testing metrics to obtain more cases in Outlier class.
Conclusions:
This study explored the possibility of predicting later behavioral scores from earlier/concurrent fMRI-based functional-connectivity maps using 3D convolutional neural networks. The preliminary results showed reasonable prediction power of visual / sensory motor functional-connectivity maps, revealing the potential of predicting later IQ / behavioral capabilities from earlier/concurrent brain functional-connectivity topologies in first years of life. Future steps may include adapting deep learning model and employing other models like graph neural networks [4] to explore if better prediction can be achieved beyond the shape and topology of functional network maps.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 2
Higher Cognitive Functions:
Higher Cognitive Functions Other
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural)
Keywords:
Cognition
Development
FUNCTIONAL MRI
Machine Learning
Other - Deep Learning
1|2Indicates the priority used for review
Provide references using author date format
[2] Shi, F., Yap, P.-T., Wu, G., Jia, H., Gilmore, J. H., Lin, W., & Shen, D. (2011). Infant Brain Atlases from Neonates to 1- and 2-Year-Olds. PLoS ONE, 6(4), e18746. https://doi.org/10.1371/journal.pone.0018746
[3] Smith, S. M., Fox, P. M. T., Miller, K. L., Glahn, D. C., Fox, P. M. T., Mackay, C. E., Filippini, N., Watkins, K. E., Toro, R., Laird, A. R., & Beckmann, C. F. (2009). Correspondence of the brain’s functional architecture during activation and rest. Proceedings of the National Academy of Sciences of the United States of America, 106(31), 13040–13045. https://doi.org/10.1073/pnas.0905267106
[4] Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2021). A Comprehensive Survey on Graph Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, 32(1), 4–24. https://doi.org/10.1109/TNNLS.2020.2978386