Deep Learning for Depression Symptomatic Activity Recognition

Discovering solutions to a handful of hazards associated with human health through analyzing physical activities have elevated the aspiration of prognosis of various diseases. This novel approach employs Human Activity Recognition (HAR) to devise depression symptomatic activities as a repercussion. The accumulation of data is conducted versatilely, including outdoor and indoor activities, maintaining all desirable postures while the smartphone is located in the slash pocket. Concerning the augmentation of the original dataset, synthetic sensors data are generated by the Generative Adversarial Network (GAN) model.Data preprocessing, including Butterworth low pass filter, enhanced the dataset. Considering the type of data being sequential, deep learning models like Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) were employed. The evaluation process is held in three stages. Firstly, LSTM outperformed GRU on the amalgamation of actual and generated data with an accuracy of 96.48%. Secondly, the selected dataset was further experimented to recognize the significance of the Butterworth low pass filter, where it achieved greater accuracy of 2.01% with its presence. Finally, two publicly available datasets WISDM and MHEALTH were compared with the acquired dataset based on the two proposed models, where the combination of the depression dataset and LSTM model acquired higher accuracy of 1.80%.