Predicting Depressive Behavior with Monitoring Activity Data Using Machine Learning and Feature Selection Approaches

Depression is a common mental health disorder. It significantly affects millions of people's daily lives all around the world. Despite the high incidence of depression and limited treatment options, many people are hesitant to seek professional help. Researchers are now exploring the use of wearable technology to collect personalized physiological data, which has recently been applied to provide healthcare services. Depression can be recognized with the help of machine learning using motor activity data collected by the ActiGraph wristband. Large data samples and higher effectiveness are necessary for developing a reliable and robust machine-learning model. However such quantitative and relevant clinical data are not openly available. This study used the depresjon (depression in Norwegian) dataset consisting of time-series data on the activity of only 23 depressed and 32 healthy individuals. However, a data augmentation technique was used that increased the number of samples from 693 to 5500. Among these algorithms, XGBoost obtained an impressive accuracy of 98.8% in 10-fold cross-validation, which met the requirements to accurately detect depression. In addition, this investigation highlighted different hidden patterns included in the time series of the two classes. Furthermore, this study utilized Explainable AI approaches, such as SHAP. These approaches further improve model transparency and interpretability. The explanation highlighted the significance of the statistical feature Median activity among features like Mean, Zero crossing rate, Standard deviation, etc. that significantly influence the model's decision-making process.