UC Irvine

Affect, encompassing both emotions and moods, plays a vital role in shaping human experience and behavior. Emotions are intense, short-lived reactions to specific stimuli, while moods are prolonged, less intense affective states influencing overall perception and behavior. Understanding and monitoring affect is crucial for enhancing mental health and emotional well-being. This thesis explores the intersection of affect, mental health, and the autonomic nervous system (ANS), emphasizing the significance of physiological signals in capturing involuntary physiological processes.

With advancements in wearable IoT devices, continuous monitoring of physiological responses such as photoplethysmography (PPG) has become feasible. This research addresses the limitations of traditional self-reporting methods by proposing a comprehensive system architecture for affect monitoring in daily life. The system integrates three essential components: Context-Awareness, Personalization, and Robustness. Context-awareness enables the system to adapt its monitoring and decision-making processes based on the dynamic environment and user-specific conditions, ensuring that the responses are relevant and timely. Personalization ensures that the monitoring is attuned to individual differences through machine learning algorithms. Robustness guarantees reliable performance against daily-life noise and motion artifacts.

First, we introduce pyEDA, a powerful tool that harnesses deep learning techniques to extract more relevant features from physiological signals like PPG and EDA, with a focus on stress-related indicators. We detail the methodology and implementation of pyEDA, emphasizing its ability to handle large datasets and extract meaningful physiological features critical for effective affective monitoring.

Then we focus on the development of a PPG motion artifact removal module. This module is crucial for ensuring robust performance in everyday scenarios where motion artifacts can significantly distort physiological data. We discuss various signal processing techniques employed to mitigate these artifacts, ensuring the reliability of the collected data.

Next, we present a novel module for extracting respiration rate from PPG signals. Incorporating respiratory rate as an additional modality enriches the affective monitoring system. We elaborate on the algorithms and validation processes used to accurately derive respiration rate, demonstrating its significance in understanding physiological responses.

After that we describe the overall closed-loop system architecture, which integrates physiological data, context data, and stress labels. This integration is crucial for developing a context-aware affect monitoring system tailored for daily life. We provide an in-depth explanation of the system components and their interactions, showcasing how this context-aware closed-loop design enhances the accuracy and relevance of affective monitoring.

Finally, we apply a context-aware active reinforcement learning approach to the proposed closed-loop system. This approach aims to enhance both performance and user engagement by incorporating personalization, dynamically adapting to the user's context, preferences, and feedback. We discuss the reinforcement learning framework, the experimental setup, and the results, demonstrating the effectiveness of this approach in real-world applications.

This work aims to develop a context-aware, robust, and personalized affect monitoring system for real-world applications, advancing the field of mental health and affective computing.