mRisk: Continuous Risk Estimation for Smoking Lapse from Noisy Sensor Data with Incomplete and Positive-Only Labels.

Area of Science:

• Digital Health
• Behavioral Science
• Machine Learning

Background:

• Wearable and mobile sensors enable passive detection of risk factors for unhealthy behaviors.
• Identifying opportune moments for intervention is crucial but challenging due to noisy sensor data and lack of reliable risk state labels.

Purpose of the Study:

• To develop a method for passively detecting rising risk of imminent adverse behaviors using sensor data.
• To improve the effectiveness of behavioral interventions by identifying timely intervention opportunities.

Main Methods:

• Event-based encoding of sensor data to reduce noise.
• Modeling historical influence of sensor-derived contexts on adverse behavior likelihood.
• A novel loss function to address limited negative labels and few positive labels.
• Deep learning models trained on 1,012 days of sensor and self-report data from 92 participants in a smoking cessation study.

Main Results:

• The model produces a continuous risk estimate for impending smoking lapses.
• Risk peaks an average of 44 minutes before a smoking lapse.
• Simulations indicate the model can create intervention opportunities for 85% of lapses with 5.5 interventions per day.

Conclusions:

• Passive sensor data and deep learning can effectively predict adverse behaviors.
• The proposed methods enable timely interventions, significantly improving behavioral intervention strategies.
• This approach holds promise for various behavioral health applications beyond smoking cessation.