Understanding smoking behavior using wearable sensors: Relative importance of various sensor modalities

Abstract

The Personal Automatic Cigarette Tracker (PACT) system, which consists of abdominal (AB) and thoracic (TC) breathing sensors, and a RF hand-to-mouth proximity sensor (PS), has proven to be useful in the detection and characterization of cigarette smoke inhalations. In this research, we further analyze the impact of subjects' anthropometric characteristics on the quality of sensor signals and evaluate the contribution of each sensor modality to the accuracy of the classifier for smoke inhalations detection. Results indicated that subjects with medium BMI, high BMI, and in a standing position were, respectively, 1.91, 4.74 and 4.32 times more likely to affect the quality of the breathing signal. Features extracted from TC+AB+PS, TC, AB, and PS sensors for individual detection models, resulted in F-scores of 94%, 85.39%, 88.54% and 90.48% respectively. For group models, the F-scores were 67.12%, 41.46%, 46.56% and 59.14%. This indicates higher contribution of abdominal breathing and hand gestures to detection of smoke inhalations.