Leveraging Activity Recognition to Enable Protective Behavior Detection in Continuous Data
summary
A hierarchical HAR-PBD architecture combines GC-LSTM graph convolutions with LSTM to detect protective behavior in continuous wearable data where protective labels make up ~21% of samples. A class-balanced focal categorical cross-entropy loss (class-balanced focal categorical cross-entropy) addresses imbalance. Evaluated on a chronic-pain patient dataset the system raises macro F1 to 0.81 vs. 0.66 and PR‑AUC to 0.60 vs. 0.44, degrades with fewer sensors but remains above chance, and targets real‑time, activity‑contextualized detection for rehabilitation support.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:27Motivation
- 1:56Backbones
- 2:18The Baseline
- 2:32Aleviating Data Imbalance
- 3:19Combining Activity Recognition
- 3:59The Impact of Sensor Size
related talks
11mDetection and Differentiation of Obstacles in Repeated Adaptive Human-Computer Interactions from ...
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning
09mAn Activity Recognition System for Taking Medicine Using In-The-Wild Data to Promote Medication ...
01h 01mD1T2 - Applying Machine Learning to User Behavior Anomaly Analysis
06mWeakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables
06mAdversarial Multi-view Networks for Activity Recognition