CUDLE: Learning Under Label Scarcity to Detect Cannabis Use in Uncontrolled Environments Using Wearables

Abstract

Wearable sensor systems have demonstrated great potential for real-time, objective monitoring of physiological health to support behavioral interventions. However, obtaining accurate labels in free-living environments remains challenging due to limited human supervision and reliance on self-labeling by patients, complicating data collection and supervised learning. To address this, we introduce cannabis use detection with label efficiency (CUDLE), a novel framework that leverages self-supervised learning with real-world wearable sensor data to automatically detect cannabis consumption in free-living environments. CUDLE identifies consumption moments using sensor-derived data through a contrastive learning framework, first learning robust representations via a self-supervised pretext task with data augmentation. These representations are then fine-tuned in a downstream task with a shallow classifier, allowing CUDLE to outperform traditional supervised methods, especially with limited labeled data. To evaluate our approach, we conducted a clinical study with 20 cannabis users, collecting over 500 h of wearable sensor data and user-reported cannabis use moments through ecological momentary assessment (EMA) methods. Our analysis shows that CUDLE achieves a higher accuracy of 73.4% compared to 71.1% for the supervised approach, with the performance gap widening as the number of labels decreases. Notably, CUDLE not only surpasses the supervised model while using 75% fewer labels but also reaches peak performance with far fewer subjects, indicating its efficiency in learning from both limited labels and data. These findings have significant implications for real-world applications, where data collection and annotation are labor-intensive, offering a path to more scalable and practical solutions in computational health.