A machine learning approach on identification of people with major depressive disorder using neuroimaging data

Abstract

Major Depressive Disorder (MDD) is a prevalent psychological disorder which causes significant impacts on physical and mental health. Although there have been diagnostic criteria and auxiliary screening manipulations, an automatic system to identify people with MDD using their neuroimaging data is still expected to facilitate the detection, management and prevention of this mental disorder. Along with the power of machine learning for the fast organizing and processing of large datasets, a deep learning neural network model was applied due to its considerable performance and ability to discover the most important input dimension over other machine learning methods. This study employed a deep learning neural network (DNN) method to classify MDD patients versus healthy controls using brain structural and functional measures. The participants were part of a larger dataset including 161 adults aged 60 or above recruited in a cross-sectional case-control study. And 41 samples with MDD and 20 without were included in the current study. A meta-analysis method was conducted to reduce the data dimension into a smaller size of seven regions within the cerebrum to improve the efficiency of model training. Along with bilateral pons, a structural feature set including grey matter volumes of nine regions and another functional feature set containing the functional connectivity of each pair of regions within the nine regions was separately built up for DNN, as well as two datasets with an identical data form but balanced sample size. Eventually a 5,000-step bootstrapping procedure was executed on each dataset for a stable result. The model achieved satisfactory and statistically significant prediction accuracies based on both structural (72.8%) and resting-state connectivity (52.9%) features. The prediction performance, particularly that based on grey matter volumetric data, was comparable or even superior to past studies attempting to separate MDD patients from healthy controls or patients with other mental illnesses. This is a noteworthy achievement given that this model was constructed based on a pre-selected small number of brain features. Consequently, compared to other deep learning models, the one in the current study had a much-simplified structure and greatly reduced computational cost, while still achieving satisfactory prediction performance. The feature weights generated by the models highlighted the importance of the left inferior frontal cortex, right amygdala, pons and left dorsolateral prefrontal cortex grey matter volumes in distinguishing MDD patients and healthy controls. These regions have all been previously associated with emotion processing and regulation. While the role of the pons in MDD has not been much investigated before, the current findings provide valuable additional evidence supporting the importance of this region in affective dysregulation. It is further revealed that the resting-state functional connectivity between the left and right inferior frontal cortex, the left and right amygdala, and between the pons and the inferior frontal cortex are critical to MDD identification. These results highlight the importance of the inter-hemispheric coherence in normal emotional function.