Real-time prediction of smoking activity using machine learning based multi-class classification model

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Smoking cessation interventions can be enhanced by real-time detection of smoking behavior.
• Distinguishing smoking from other daily activities using wearable sensors is a significant challenge.

Purpose of the Study:

• To develop a machine learning framework for real-time smoking activity detection using wrist-wearable inertial measurement unit (IMU) sensor data.
• To create a low-cost wearable device for collecting relevant sensor data.

Main Methods:

• Utilized a sliding window mechanism to process streaming sensor data and extract time-domain, frequency-domain, and descriptive features.
• Performed hyperparameter tuning and feature selection for model optimization.
• Developed and validated multi-class classification models using in-sample and out-of-sample testing.

Main Results:

• Achieved up to 98.7% predictive accuracy (area under the receiver operating curve) in identifying smoking activity.
• Demonstrated the feasibility of real-time detection of smoking among confounding activities of daily living (ADLs).

Conclusions:

• The developed framework offers a novel application for wearable devices in accurately detecting smoking in real-time.
• Findings can aid healthcare professionals in monitoring patients and delivering just-in-time interventions for smoking cessation.
• The framework has potential for broader applications in preventive healthcare and activity detection.