In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability.

Area of Science:

• Wearable sensor technology
• Machine learning for health monitoring
• Biomedical signal processing

Background:

• Deep learning (DL) models for Human Activity Recognition (HAR) using wearable Inertial Measurement Unit (IMU) sensors offer potential for continuous health monitoring and early disease detection.
• Current DL HAR models often lack robustness due to training on limited, lab-controlled data, failing to generalize to real-world conditions.

Purpose of the Study:

• To investigate the impact of subject, device, position, and orientation variabilities on the performance of DL HAR models.
• To quantify data distribution shifts caused by these variabilities using Maximum Mean Discrepancy (MMD).
• To establish the relationship between distribution shifts and DL HAR model performance.

Main Methods:

• Utilized the HARVAR and REALDISP datasets to isolate and analyze variability effects.
• Employed Maximum Mean Discrepancy (MMD) to measure data distribution shifts.
• Correlated MMD values with DL model performance metrics.

Main Results:

• Variability in subject, device, position, and orientation significantly degraded DL HAR model performance.
• An inverse relationship was observed between the magnitude of data distribution shifts (MMD) and model performance.
• The compounding effects of multiple variabilities, as studied in REALDISP, demonstrated substantial challenges in real-world generalization.

Conclusions:

• Real-world variabilities pose significant challenges to the generalization of DL HAR models.
• MMD is a valuable metric for assessing distribution shifts and explaining performance degradation in HAR data.
• Developing more robust DL HAR models capable of handling real-world variability is crucial for effective health monitoring applications.