Exploring the Utility of a Machine Learning Approach with Mobile-Based Cognitive Function Tasks for Detecting Depression

Abstract

Self-report questionnaires, used for detecting major depressive disorder (MDD) in daily life, may incur biases stemming from social desirability and repetitive answers. Though detection based on mobile sensing was being developed recently, it cannot sufficiently promote self-help action due to the characteristics of passive feedback. Thus, an active self-monitoring and feedback system is crucial for individuals to recognize and address their malfunctions. In this study, we proposed to predict changes in MDD severity using cognitive tasks monitored on mobile devices. An online survey was conducted to evaluate the severity, incorporating cognitive tasks such as Navon task, Go/No-go task, and n-back task, along with the Quick Inventory of Depressive Symptomatology. Participants completed the survey three times on their mobile devices. The analysis included data from 75 participants, including 21 participants whose MDD score increased by at least one point during the second and third surveys; the first survey was excluded to avoid confounding effects. A random forest classifier was employed for classifying participants whose depression has and has not worsened. The learned model achieved modest accuracy (68.3%) with a significant mean area under the curve of 0.59 (t(9) = 2.98, p = .016, d_z = 0.94), suggesting the potential to predict depressive states based on cognitive domains. Moreover, working memory and attentional inhibition functions contributed to predicting the severity change mostly. Though improvements are required to reduce false negatives for practical applications, our result suggests that MDD aggravation could be assessed by mobile cognitive tasks.