TL;DR
This paper introduces StopWatch, a novel system that uses a low-cost Android smartwatch to passively detect cigarette smoking through hand motion analysis. This eliminates the need for expensive sensors or a connected smartphone, paving the way for more accessible and unobtrusive smoking cessation support.
What Problem Does This Paper Address?
Current smoking cessation support systems often rely on "just-in-time" (JIT) interventions, which require accurate, real-time data about a user's smoking behavior. Traditional methods of collecting this data, such as self-reporting, are prone to recall errors and reporting biases. While some technological solutions exist, they often involve expensive bespoke hardware or require a connected smartphone, limiting their accessibility, especially for lower-income populations who have higher smoking prevalence. This paper addresses the need for a low-cost, unobtrusive, and smartphone-independent system for passive cigarette smoking detection.
Methods
StopWatch utilizes the accelerometer and gyroscope sensors within a commercially available Android smartwatch to capture motion data. A machine learning pipeline, running entirely on the smartwatch itself, processes this raw data. The pipeline consists of three stages:
- Motion Feature Extraction: Raw sensor data is binned and thresholded, and an initial decision tree classifier is applied to identify basic hand movements relevant to smoking (e.g., "hand raising to mouth," "hand stationary at mouth").
- Drag Detection: A second decision tree classifier analyzes these motion features for specific patterns and durations that characterize a single "drag" of a cigarette.
- Smoking Instance Detection: Rules are applied to the sequence and timing of detected drags to identify a complete instance of smoking a cigarette (defined as six drags with less than 80 seconds between them).
This on-device processing is a key differentiator, as it removes the dependency on a smartphone for analysis.
Datasets
The paper describes two phases of data collection:
- Parameter Determination: Data from 38 participants was used to determine the parameters for the analysis pipeline.
- Validation: Data from 14 new participants was used for validation. One participant was excluded, leaving 13 participants for the free-living validation.
The validation involved both laboratory settings and 24-hour free-living periods. During the free-living phase, participants could log false positives and false negatives via the smartwatch, and also completed a paper diary of smoking events.
Key Insights and Ideas
- On-Device Processing is Feasible: The paper demonstrates that sophisticated motion analysis for smoking detection can be effectively performed directly on a low-cost smartwatch, significantly reducing system complexity and cost.
- Smartphone Independence is Crucial: By removing the smartphone requirement, StopWatch becomes accessible to a broader demographic, including those with lower incomes who may not own a smartphone. This is critical for widespread adoption of smoking cessation technologies.
- Passive Detection Reduces Bias: The system offers a way to gather objective smoking behavior data, bypassing the inherent inaccuracies of self-reporting.
- Potential for Advanced JIT Interventions: The detailed, passive data collected by StopWatch can fuel more sophisticated context-triggered and adaptive interventions for smoking cessation.
Key Results
In the free-living validation phase with 13 participants:
- Precision: 86% (95% CI: 78% to 93%)
- Recall: 71% (95% CI: 63% to 78%)
The system achieved substantial agreement with participant paper diaries, with a Cohen's Kappa of 0.73. In laboratory settings, the system achieved 75% precision, 92% recall, and 90% accuracy.
Limitations and Future Work
Limitations: * Performance: The current performance (especially recall) is lower than some more expensive and complex systems. * Smartwatch Specificity: The system has only been validated on an LG G-Watch running Android Wear v1.5.
Future Work: * Enhancing system performance (recall and precision). * Validating the system on a wider range of smartwatches. * Testing the feasibility of using StopWatch in various smoking behavior change interventions. * Exploring the detection of electronic cigarette use and distinguishing it from traditional cigarette smoking.
Summary updated