UC Irvine

Non-invasive assessment of physical and mental health has been the core of an extensive body of research in recent years. Due to the development of wearable bio-sensors and the feasibility of wearing them in long time-frames (e.g. smartwatches), a tremendous opportunity lies in development and improvement of monitoring and detection methods and models for physical and mental health, based on biophysical data.We start by developing an advanced personalized data-driven approach that uses a three-layer deep neural network to estimate Blood Pressure (BP) based on photoplethysmography (PPG) signals. The model takes windows of PPG signals, and automatically extracts features which it uses for the regression problem. The proposed model outperforms prior methods and passes two long-standing standards for BP monitoring systems. In the next steps, we move forward with the development of a system for mental health assessment. Most existing health signal based tracking and monitoring approaches achieve better results presuming a large pre-labeled dataset is available and can be utilized for training the models. This is not a challenge-free assumption since these datasets usually do not exist and are not easy to collect particularly for new applications. Furthermore, these signals and labels are highly subjective and show different data distributions among different subjects. We propose a Q-learning based human-in-the-loop active learning model which releases the restriction of having a large pre-labeled dataset and aims at collecting labels for instances that are expected to improve the model's performance more efficiently. The goal is to not only minimize human annotation while maximizing the model performance through personalizing the models, but also analyzing temporal correlations to optimize the query times. Given that the experiment occurs in everyday settings (also known as \textit{in-the-wild}) and over long periods of time (time frame of few weeks to few months), maximizing user engagement is another important factor we considered in the proposed model. The proposed Q-learning based agent that selects certain instances to be labeled by the users, through analyzing the instance itself as well as the contextual and behavioral information from the subject, all in real time. The framework iteratively updates the agent based on subjective response behavior and also updates the model with the new subjective data. We performed three rounds of field studies with groups of volunteers, to collect data and evaluate the performance of the proposed framework. Extensive experiments demonstrate the superiority of the proposed model to existing unsupervised and active learning models. The proposed context-aware agent is able to provide the same level of personalization as random selection and a standard active learning method, with up to 88% and 32% fewer queries from users, respectively.