Multi-Scale Alcohol Consumption Behavior Detection from Wearable AI
Wearable sensor-based alcohol detection requires methods that are aligned with the available label structure, sensing modality, and temporal scale. This paper presents a comparative study across two alcohol sensing settings. The first setting uses the Bar Crawl dataset, where smartphone accelerometer data must be interpreted with delayed transdermal alcohol concentration (TAC) measurements. To address this weak supervision, we apply a backward-search framework anchored on TAC rise and prior motion activity. Candidate motion episodes were identified 35-37 minutes before TAC rise in observable cases, while sedentary behavior and limited observability explained non-detected cases. The second setting uses a wrist-worn IMU drinking gesture dataset with explicit gesture annotations under leave-one-subjectout (LOSO) evaluation. Among six models (LR, RF, SVM, XGB, LGBM, and 1D-CNN), LGBM achieved the best semi-controlled DX-I performance (F1=0.699), while RF achieved the best freeliving DX-II baseline performance $(\mathbf{F} \mathbf{1} \boldsymbol{=} \mathbf{0. 3 8 0})$. The results show that label structure, class imbalance, and evaluation protocol strongly influence valid methodology, reported performance, and generalizability in wearable alcohol sensing.