A novel machine learning approach to shorten depression risk assessment for convenient uses

Background: Depression is a mental disorder affecting many people worldwide which has been exacerbated by the current pandemic. There is an urgent need for a reliable yet short scale for individuals to self-assess the risk of depression conveniently and rapidly on a regular basis. Methods: We obtained a dataset of responses to the Depression, Anxiety, and Stress questionnaire (DASS-42) from a large sample of individuals worldwide (N = 31,715). With this dataset, important items from the questionnaire were extracted by applying feature selection techniques. Then, using the most important features, various machine learning algorithms were trained, tested, and validated in predicting depression status. Results: This study revealed that only seven items are needed to predict depression status with at least 90 % accuracy of the original full scale. This can be achieved through the Stacked Generalization Ensemble learning method of multiple models. The trained machine learning models from the best algorithm were then implemented as an online Depression Rapid Assessment tool, which allows the user to evaluate their current depression status conveniently and quickly (about 1 min). Limitations: The sample size of the present study is still relatively small and has biases toward certain demographics (e.g., mostly young, Asian, and female). Further, memory issues with Stacked Generalization Ensemble prevent it from being trained in the same way as the other algorithm. Conclusion: It is possible to produce very short assessments that approximate the accuracy of the full scale for convenient and rapid self-assessment of depression risks.