Breath Analytics: A Sensor-Driven Study of Day-to-Day Human Respiration

Abstract

Breathing is a vital physiological process that serves as a window into both physical and mental health. This dissertation explores the feasibility of measuring and analyzing human breathing patterns using a wearable sensor in real-world, non-clinical environments across a full day. We focus on characterizing the temporal structure of individual breaths—composed of inhale, exhale, and hold periods—and identifying recurring characteristic breaths. These \textit{characteristic breaths} capture the diversity of respiratory behavior while revealing consistent patterns across individuals and their health-states.

Utilizing the Spire Tag, a wearable multi-sensor device equipped with a piezoelectric force sensor, accelerometer, and PPG, we collected high-resolution respiratory and movement data. We conducted a study to examine normal breathing, and include observations on coughing, sneezing, and other respiratory events, to gain deeper insights into respiratory physiology.

We analyzed real-world breathing data from a large public health study where participants wore Spire tags 24/7 in an uncontrolled setting to investigate the distinct changes in breathing patterns that emerge before, during, and after respiratory illness. We developed a system to process and segment the data into individual breaths and select one subject from this cohort and extract time-based features from over 17,000 unique breaths. Through clustering, we derive nine representative characteristic breath types that describe the underlying physiology and account for over 80\% of daily breathing patterns.

We provide a case-study on three subjects from this cohort by analyzing their full day respiratory behavior. Our findings suggest that shifts in the distribution of characteristic breath types specific to the individual as they transition between sick and non-sick days. Our work lays the foundation for a data-driven framework for personalized respiratory monitoring and real-time illness detection.