Using the Long to Short Approach to Develop Rapid Depressions Scales

Clinical depression affects 5-10% of the world population each year and is a serious mental health issue globally. There are many traditional psychological scales that assess levels of depression in adults, where their items are often redundant in the information they carry, and their scoring is not necessarily linear to the item scores. Thus, machine learning techniques can help find the redundancy in the items, as well as the nonlinear relationship between the item scores and the final prediction. Using the Depression Anxiety Stress Scale 42 (DASS-42) as the basis, participants will be asked to fill out an online questionnaire about their demographics information (age, gender, country of residence, race, etc.) and all 42 items of DASS-42 to provide a dataset for this study. Feature selection techniques such as MRMR and Gini feature importance were applied to identify the most important features in the dataset. Then, using machine learning methods such as Logistic Regression, XGBoost, and Ensemble models, models will be fitted on the most important features to develop a shortened depression scale (7-9 items consisting of demographics items and DASS items) that accurately predicted the levels of depression (as measured by the AUC, ROC and F1 scores.