CNN-BiLSTM Model Classifies Sedentary Behavior from Wearables

This study investigates the use of a deep learning model, CHAP (CNN-BiLSTM), for accurately classifying sedentary behavior using data from wearable accelerometers. The research specifically examines the model's ability to generalize from hip-worn sensor data to wrist-worn sensor data, a more challenging scenario due to higher signal variability. Experiments conducted on the iWatch dataset, which includes ground-truth posture labels, confirmed that the hip-trained CHAP model performs robustly on hip data. However, its accuracy significantly decreases when applied directly to wrist data. The study found that fine-tuning the hip-pretrained CHAP model with even limited amounts of labeled wrist data consistently outperformed transformer models trained from scratch on wrist data, highlighting the value of pretraining and the necessity of adaptation for different sensor placements.