Non-Pervasive Monitoring of Daily-Life Behavior to Access Depressive Symptom Severity Via Smartphone Technology

The number of people suffering with mental health disorders is rapidly increasing in recent years and it is very common with individuals who like to live alone and escape social meetings. Amongst various kinds of mental health disorders, depression is very common and serious one. In this paper, we propose a method to assess the depression level of an individual using smartphone by monitoring their daily activities. Smartphone time domain acceleration and gyroscope sensor filtered data were used in LSTM-RNN model to classify four physical activities (i.e., resting, exercising, running, walking) Additionally, the geographical location data was clustered to simplify movement activities. Subsequently, from participant activities, ten features were extracted that corresponded with their weekly reported questionnaire (QIDS-16) depression score. Features were used in the regression model to estimate the participant QIDS score. Among all the features, a subset that showed promising relationship with depressive symptom severity was selected using the wrapper feature selection method. Afterwards, these selected subset features were applied in both linear regression model and quadratic discriminant analysis classifier to estimate depression score as well as depression severity level. Regression model for score estimation showed the error rate of root mean square deviation is 3.117. On the other hand, for depression level classification selected quadratic discriminant analysis classifier method had an accuracy of 92%. This identification system appears to be a cost-effective solution that can be used for long-term and can monitor depressed individuals without invading their personal space or creating any disturbance.