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

This paper presents research on a Predictive and Uncertainty Assessment Framework (PUAF), providing a comparative analysis of two prominent methods, Monte Carlo (MC) Dropout and Bootstrap-based models, used in uncertainty estimation techniques of Neural Network predictions of human activity recognition using accelerometer data. Unlike traditional studies that optimize classification accuracy, this work emphasizes uncertainty quantification to enhance model reliability, particularly for critical health-related activities. Among the five activity classes of Sit, Sleep, Walk, Cough and Fall, this work concentrates on the Cough and Fall cases. The study exploits acceleration data from a wearable device positioned on the user's chest, with features derived from three-axis motion measurements. Synthetic datasets are generated by systematically introducing noise variations, added to the original dataset across all axes, to assess robustness under real-world conditions. Each uncertainty estimation method estimates the probabilities for the five different classes along with the corresponding 95% confidence intervals to quantify the prediction uncertainty. A detailed evaluation is conducted by analyzing the average width of these confidence intervals across different noise levels, identifying the most reliable feature and model combination. Both the MC Dropout and Bootstrap enhance model robustness and uncertainty awareness under noisy sensor conditions. The MC Dropout provides sharper and more sensitive uncertainty estimates, while the Bootstrap yields more stable and better-calibrated predictions. The evaluation using the proposed PUAF demonstrates that each method offers distinct advantages, highlighting the importance of uncertainty quantification for reliable wearable-based HAR systems.