Identifying Smoking from Smartphone Sensor Data and Multivariate Hidden Markov Models

Abstract

Smoking is one of the foremost public health threats listed by the World Health Organization, and surveillance is a key to informing effective policies. High smartphone penetration and mature smartphone sensor data collecting techniques make smartphone sensor data based smoking monitoring viable, yet an effective classification algorithm remains elusive. In this paper, we sought to classify smoking using multivariate Hidden Markov models (HMMs) informed by binned time-series of transformed sensor data collected with smartphone-based Wi-Fi, GPS, and accelerometer sensors. Our model is trained on smartphone sensor time series data labeled with self-reported smoking periods. Two-fold cross-validation shows \(A_{z}\) (area under receiver operating characteristic curve) for HMMs using five features = (0.52, 0.84). Comparison of univariate HMMs and multivariate HMMs, suggests a high accuracy of multivariate HMMs for smoking periods classification.