DEPRESSION DETECTION THROUGH ACTIVITY RECOGNITION: DEEP LEARNING MODELS USING SYNTHESIZED SENSOR DATA

Abstract

Despite having a very different kinesthetic sensibility from smartphone sensors, human bodies assess variations and address specific sensor values. Human Activity Recognition (HAR) has been used extensively in various applications; however, despite this, HAR cannot currently be used to correlate activity patterns to identify biomarkers for any disease. This provisional scrutiny leverages HAR to detect depression-symptomatic activities. Data collection was carried out in various ways, combining outdoor and indoor activities, while the smartphone was kept in the slash pocket. The Generative Adversarial Network (GAN) model generates synthesized sensor data to enhance the size of the original dataset. The dataset was improved via preprocessing, including the Butterworth low-pass filter. Since the data was linear, deep learning models such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) were used. The evaluation procedure is divided into three sections. To begin with, LSTM outperformed GRU on the combination of actual and generated data, with an accuracy of 96.48%. Furthermore, the selected dataset was then analyzed to identify the impact of the Butterworth low-pass filter, which produced a higher accuracy of 2.01%. Finally, the obtained dataset was compared to two publicly accessible datasets, WISDM and MHEALTH, regarding the two suggested models, with the conjunction of the depression dataset and the LSTM model achieving a greater accuracy of 1.80%.