Few-Shots Learning Model for Depression Detection Using Actigraphy

Depression is one of the most serious diseases that have plagued humanity in recent days. Several global health organisations have highlighted the ever-increasing cases of depression and the catastrophic mortality and morbidity rates that could result in the coming years. However, the modest size of the publicly available dataset describing depression behaviour hinders the broad applicability of data-hungry deep learning models in diagnosing this disease. This paper attempts to bridge this gap by presenting a model that addresses the paucity of relevant datasets. The proposed model uses one of the state-of-the-art meta-learning approaches, called the relation network, which can be learned from a small number of samples (few-shot learning). The underlying principle of the relation network is to create a model that learns to compare rather than learns to fit as is the case with traditional deep learning models. The proposed model has redesigned the architecture of the relation network to be suitable for time series classification. Our model uses Long- Short-Term Memory (LSTM) to extract the temporal features of mood swings in depression and the multilayer perceptron (MLP) to facilitate relationship discovery. The proposed model was evaluated using a dataset in which the motor activities of a study group were recorded using body-worn Actigraph accelerometers. The results of this evaluation show that the proposed model performs excellently across 1,5 and 10 shots with an overall accuracy of 99.99, a specificity of 99.98, a precision of 99.99 and F1 score of 99.97 outperforming most related works.