Imputing Brain Measurements Across Data Sets via Graph Neural Networks.

Area of Science:

• Neuroimaging and Machine Learning
• Computational Neuroscience
• Data Science in Healthcare

Background:

• Publicly available structural MRI datasets often lack crucial brain region measurements (ROIs) needed for machine learning model development.
• Existing methods for addressing missing data, like reapplying software or standard imputation techniques, are computationally intensive or inadequate for entire missing measurements.
• The Adolescent Brain Cognitive Development (ABCD) Study dataset, for instance, does not release Freesurfer curvature scores, hindering specific research applications.

Purpose of the Study:

• To develop and validate a novel deep learning approach for imputing entire sets of missing neuroimaging measurements from publicly available datasets.
• To address the limitations of current imputation methods by reframing the problem as a cross-dataset prediction task.
• To account for demographic variations, such as sex, in brain measurements during the imputation process.

Main Methods:

• Proposed a Demographic Aware Graph-based Imputation (DAGI) algorithm utilizing a graph neural network (GNN) to model dependencies between ROI measurements.
• Employed a parallel architecture that simultaneously trains a graph decoder for imputation and a classifier for predicting demographic factors.
• Trained DAGI on the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) dataset to impute missing Freesurfer measurements for the ABCD dataset.

Main Results:

• Cross-validation on the NCANDA dataset demonstrated that DAGI-imputed scores significantly outperformed those generated by linear regressors and existing deep learning models.
• The inclusion of DAGI-imputed scores enhanced the accuracy of a sex classification model compared to using only the available ABCD Freesurfer scores.
• DAGI successfully imputed missing Freesurfer measurements for a large cohort of the ABCD study (N=3760).

Conclusions:

• DAGI offers an effective and computationally efficient solution for imputing missing neuroimaging data, overcoming limitations of traditional methods.
• The method's ability to model inter-regional dependencies and account for demographics makes it valuable for enhancing large-scale neuroimaging datasets.
• DAGI has the potential to improve the training and performance of machine learning models in neuroscience research by leveraging diverse data sources.