Predicting Respiration Rate using Acceleration Sensors and LSTM: A novel Approach

The significance of appraising respiratory rate as a determinant of sleep quality has achieved a firm foundation within the literature. This has led to the emergence of numerous technological solutions encompassing both non-contact and wearable modalities, leveraging sensor mechanisms encompassing cameras, radar, wireless communication, and accelerometers, with the objective of estimating respiratory frequencies. Nevertheless, the efficacy of non-contact methodologies has frequently been compromised when confronted with constrained environmental circumstances, exemplified by low illumination levels and tangible obstructions. Conversely, ongoing investigative endeavors targeting wearable apparatuses have exhibited a predilection for machine learning paradigms; however, the potential for more advanced and precise deep learning models remains inadequately harnessed. Identifying the potential for enhancement, this work presented a novel approach for estimating respiration rates. This involves a synergistic fusion of an accelerometer and an LSTM deep learning model, resulting in superior predictive capabilities for respiratory rate compared to their individual performances. In this experiment, data were collected from 10 teenagers with dynamic respiration rates during various sleep postures (lateral left, lateral right, supine, prone, and sitting). The experiment achieved ±1.962 MAE error in estimating respiratory rates.