Skip to main content Skip to navigation

Workshop / Seminar

Wearable Systems and Machine Learning for Affect Recognition and Interventions by Ramesh Sah – Preliminary Exam


About the event

Dissertation Title: Wearable Systems and Machine Learning for Affect Recognition and Interventions

Advisors: Dr. Diane Cook and Dr. Hassan Ghasemzadeh

Degree: Computer Science, Ph.D.

Abstract: Stress recognition and monitoring from wearable sensor data is an emerging area of research with significant implications for an individual’s physical, social, and mental health. Mobile health interventions that incorporate real-time monitoring of physiological and behavioral markers of stress offer promise for delivering tailored interventions to individuals during high-risk states of heightened stress. Our work aims to enable a mobile health system capable of context recognition, stress detection, and explainability for personalized intervention generation. Stress regulation and detection become even more critical for vulnerable populations such as individuals suffering from substance and alcohol abuse. Previous research has highlighted the role of stress in substance misuse and addiction, particularly for relapse risk from Alcohol Use Disorder (AUD). We conducted a study to establish the feasibility of using wearable sensor systems to continuously monitor stress in an ambulatory real-world setting. We examined the associations between physiological and ecological stress markers and self-reported outcomes of stress, alcohol cravings, and negative emotions. Features of Electrodermal Activity (EDA) and Hear Rate Variability (HRV) were significantly correlated with self-reported outcomes, including the number of stressful events, positive and negative emotions, and experienced pain and discomfort. We also developed polynomial-time sensor channel selection algorithms to determine the best sensor modality. We addressed the challenges of small labeled data and the subjective nature of stress by devising an iterative search algorithm to find the optimal segment length for stress events. Our results confirmed EDA sensor modality is most indicative of stress, and the segment length of 60 seconds around user-reported stress events delivers the top stress classification performance. Using majority undersampling to balance the classes, the binary stress classification model achieved an average accuracy of 99% and an f1-score of 0.99 on the training and test sets. Currently, we are studying the relationship between acoustical features of musical segments and physiological biomarkers of stress before and after stressful events. This study aims to establish the validity of using music listening as an intervention and stress regulation technique. Finally, we aim to study the behavioral implications of stress and investigate the explainability of machine learning algorithms trained for stress detection.


Tiffani Stubblefield
(509) 336-2958