ABSTRACT
In the modern world, mental health disorder is one of the serious problems for every nation around the globe, especially in this Covid-19 pandemic period. These illnesses affect not only patient's mental health but also physical health, thus, reduce productivity and quality of human work. Mental health disorder is complex and also takes many forms. In this study, we pay attention to depression (unipolar and bipolar) illness. We conducted our experiment on the open dataset named Depresjon, which was collected from activity motion signals on wearable devices of 32 healthy people and 23 depressed (unipolar and bipolar) patients over several days in a row with a total of 814 samples. Firstly, we did the preprocessing data step in order to make the dataset fitting the DL model input. After that, we deployed many individual DL models to make the first predictions. Next, we generated Deep Stacked Generalization Ensemble Learning (DeSGEL) models which were able to learn how to make the best combination of predictions from previous individual well-trained models. Finally, we made a comparison among the individual and the proposed DeSGEL models. The results showed that the DeSGEL models had outperformed other corresponding individual models. Specifically, among the individual models, VGG16 had the best performance. However, the DeSGEL Resnet based showed an extremely outstanding performance over other individual and ensemble DL models. In detail, these models had Accuracy, Precision, Sensitivity, Specification, F1 score, MCC and AUC of 0.94, 0.91, 0.89, 0.96, 0.90, 0.85 and 0.92 respectively for Individual VGG16 model, and 0.96, 0.96, 0.92, 0.98, 0.94, 0.91 and 0.95 respectively for the DeSGEL Resnet based model. We found that applying Deep learning, especially DeSGEL models using activity motion signal data from wearable devices could be a prospective direction for the early diagnosis of mental health disorders. © 2021 ACM.