Machine learning aided prediction of family history of depression

Increased risk for psychopathology in the offspring of depressed parents has been widely known. The brain may mediate the effects of risk of familial depression on the offspring via shared genetic and environmental factors. Conventional brain imaging studies to test this mediation effects primarily use a priori knowledge to select a subset of brain imaging-derived features. Despite of the existing positive results supporting the notion of the brain as an endophenotype for familial depression, no quantitative assessment has been performed regarding to what extents the complex brain structure contains information about familial depression. To this end, here we aim to predict whether an individual has a history of familial depression. We propose a data-driven, unbiased, and rigorous machine learning approach using multimodal brain features (e.g., grey matter morphometry based on T1-weighted images and structural connectome based on probabilistic diffusion tractography) to capture the complex representations of brain structure. We implemented logistic regression (LR) with regularization, support vector machine (SVR), and graph convolutional neural network (GCN). Our models show promising cross-validated classification accuracy: 97.78% (LR), 93.67% (SVM) and 89.58% (GCN). Brain features with greatest weights in the models include brain regions previously implicated in the depression literature (e.g., emotion regulation frontal-limbic circuit) as well as new regions. Results suggest a large impact of familial depression on the brain structure and connectome. Results also highlight potentials for prediction of risk for psychopathology in a data-driven fashion using costeffective, simple, ubiquitous brain images.