Towards a generalizable method for detecting fluid intake with wrist-mounted sensors and ...
summary
A method detects fluid intake from wrist accelerometer traces by expanding variable-length windows around local peaks rather than using fixed windows, leveraging adaptive data segmentation to match the 2–6s variation in drink gestures. Evaluation uses 30 participants and 561 drinking instances with leave-one-participant-out cross-validation, yielding 90.3% precision and 91.0% recall. Implementation uses an off-the-shelf Wavelet Health tracker and only accelerometer channels, and an anonymized labeled dataset is released.
AI-written summary · about