Recognizing cigarette smoke inhalations using hidden Markov models.

Area of Science:

• Biomedical Engineering
• Data Science
• Respiratory Physiology

Background:

• Smoking detection often relies on self-reporting, which can be inaccurate.
• Wearable systems like the Personal Automatic Cigarette Tracker (PACT) offer objective data collection.
• Distinct temporal breathing patterns during smoking suggest suitability for computational analysis.

Purpose of the Study:

• To investigate the feasibility of using Hidden Markov Models (HMMs) for classifying smoking inhalations.
• To characterize temporal information from respiratory signals and hand-to-mouth proximity for smoking detection.
• To compare HMM performance against existing machine learning methods for smoking recognition.

Main Methods:

• Utilized respiratory signals (tidal volume, airflow) and hand-to-mouth proximity data from the PACT system.
• Developed left-to-right Hidden Markov Models (HMMs) to classify smoking versus non-smoking inhalations.
• Employed leave-one-out cross-validation on a dataset of 20 subjects for model evaluation.

Main Results:

• HMMs demonstrated capability in recognizing smoking inhalations using respiratory and proximity data.
• Achieved average recall of 42.39%, precision of 88.19%, and F-score of 56.38% for smoke inhalation recognition.
• Reported a 7.3% improvement in recall compared to previously published Support Vector Machine (SVM) models.

Conclusions:

• HMMs represent a viable approach for objective smoking detection using wearable sensor data.
• The characterized temporal breathing patterns hold significant potential for improving smoking cessation interventions.
• Further research can refine HMMs for more accurate and robust smoking behavior monitoring.