Comparative Uncertainty Estimation in Neural Network Analysis of Wearable Sensor Signal for Cough and Fall Detection.

Area of Science:

• Wearable technology
• Machine learning
• Biomedical engineering

Background:

• Human Activity Recognition (HAR) systems using wearable sensors are crucial for health monitoring.
• Traditional HAR models prioritize accuracy, often neglecting prediction reliability under noisy conditions.
• Uncertainty quantification is vital for critical health-related activity detection.

Purpose of the Study:

• To present a Predictive and Uncertainty Assessment Framework (PUAF) for evaluating uncertainty estimation methods in HAR.
• To compare Monte Carlo (MC) Dropout and Bootstrap models for uncertainty quantification in accelerometer-based HAR.
• To assess the robustness of these methods against varying noise levels in sensor data.

Main Methods:

• Utilized accelerometer data from a chest-worn device for five activity classes (Sit, Sleep, Walk, Cough, Fall), focusing on Cough and Fall.
• Generated synthetic datasets with systematic noise variations across all axes to simulate real-world conditions.
• Applied MC Dropout and Bootstrap methods to estimate class probabilities and 95% confidence intervals for uncertainty quantification.

Main Results:

• Both MC Dropout and Bootstrap enhanced model robustness and uncertainty awareness in noisy environments.
• MC Dropout yielded sharper, more sensitive uncertainty estimates.
• Bootstrap models produced more stable and better-calibrated predictions.

Conclusions:

• The PUAF effectively evaluates uncertainty estimation techniques for wearable HAR systems.
• Each method (MC Dropout and Bootstrap) presents unique advantages for improving prediction reliability.
• Uncertainty quantification is essential for developing dependable wearable-based HAR systems, especially for health applications.