Optimized Risk Scores for Early Detection of Bipolar Disorder Based on Crowdsourced Text Data

Bipolar disorder is a chronic mental health disorder characterized by periodic manic or depressive episodes with high rates of recurrence and non-recovery. Early intervention for bipolar disorder is critical to prevent progression and complications that lead to societal loss. In this study, we build an early detection model for bipolar disorder based on crowdsourced mental health symptoms. The symptoms gathered through the crowdsourcing process are in the form of free texts. We first extract a list of features using natural language processing techniques to convert free texts into binary features representing mental health symptoms. Based on these features, we create an optimized risk score for detecting bipolar disorder. We formulate this risk score optimization problem as a mixed-integer non-linear programming problem. The proposed model performs better (
$\text{AUC} =0.714, \text{CAL} =0.027, \text{RA} =0.787$
) than the baseline models, and it is interpretable with integer scores assigned to each of the considered symptoms. This model can be used as an early detection tool to recommend for further examination of bipolar disorder.