A Personalized Deep Neural Network to Recognize Human Activities in Healthy Subjects

Human Activity Recognition (HAR) using smartphone sensors is among the most demanding Artificial Intelligence (AI) topics for its growing applications in clinical healthcare monitoring. Although implementing an accurate HAR model is essential, developing a personalized one and adapting it to the end-user is challenging in real-life applications. There are various personalization methods in HAR. This research combines two personalization methods to adapt the proposed HAR model to individual subjects. Accordingly, we developed an Android application to capture the accelerometers of fifteen subjects during six predefined daily activities, such as standing, sitting, lying down, walking, going upstairs, and going downstairs. Subjects were instructed to carry the phone in predetermined positions in fixed or free modes. These positions include locations where people mostly carry their smartphones. The proposed personalized HAR model consisted of two main steps: (1) the dataset was divided into multiple groups based on the subjects' physical similarities, e.g., such as weight and height. Then, (2) based on multiple data splitting configurations (user-independent, user-dependent, and hybrid), the HAR model is generalized to the end-user in each group using a Long Short-Term Memory (LSTM). After normalization and windowing, the collected data were fed to the LSTM network. The model was trained and validated by 80% and 20% of the dataset, respectively. Finally, we evaluated our proposed model using the Human Activity Sensing Consortium (HASC) corpus 2011, a large, publicly available dataset. The proposed LSTM network with the user-dependent personalization method produced an average F-measure of 0.99 and 1 for our dataset and the HASC dataset, respectively.