Predicting Depression in Canada by Automatic Filling of Beck’s Depression Inventory Questionnaire

The risk for depression and anxiety increased as people adjusted to a new normal after the COVID-19 pandemic. Early detection and appropriate onset treatment and support can reduce the consequences of depression. Automatic detection of depression in social media has recently become an important area of investigation. However, because of the lack of extensive annotated data, we propose a method for using a model that learns to answer a depression questionnaire and apply it to make population-level predictions. We used the eRisk 2021 Task 3 training dataset to build an automated model to fill the Beck’s Depression Inventory (BDI) questionnaire. We selected the best performing model for each group of questions based on predefined metrics and consolidated those models into one model (called the
$BDI\_{}Multi\_{}Model$
). The
$BDI\_{}Multi\_{}Model$
achieved better performance than the state-of-the-art for this challenging task. Then, we used this model for inference on a Canadian population dataset and compared its predictions with the statistics of the most recent mental health survey conducted by Statistics Canada. The correlation between the inference of the answered questionnaire based on our
$BDI\_{}Multi\_{}Model$
and the official statistics showed a strong Pearson correlation of 0.90.