Mobile ascertainment of smoking status through breath: a machine learning approach

The ability to remotely and automatically detect whether someone has been smoking can develop systems that promote smoking cessation via behavioral management or gamification. This article advances automated methods for ascertaining smoking status via a mobile system comprised of an environmental carbon monoxide (CO) gas sensor paired with a smartphone. We apply several machine learning strategies to the CO time courses derived from our system to predict recent smoking status in a group of smoking (n=11) and non-smoking (n=9) adult women. The study investigates i) a range of feature types conventionally used in smoke analysis studies, ii) the contribution of each feature type in overall prediction accuracies and iii) the possibility of improving the overall predictions by identifying segments of time during which the most informative data is recorded. The analysis is composed of two stages of i) conventional classification via Logistic Regression (LR) and Support Vector Machine (SVM) and ii) feature-ensemble and variations of boosting, bagging, and mixture of expertise methods to investigate the effects of feature mixture, learner mixture, and extra training on overall performances of the predictions. The results indicated that features extracted from the middle segments (end exhalation period) of the recordings led to the highest classification success. The results indicated that having higher number of training samples, and stronger mixed learners are influential to the performance across segments while the feature mixture did not improve the performances across segments.