MRNet - A Deep Learning Based Multitasking Model for Respiration Rate Estimation in Practical Settings

Abstract

The explosion of unobtrusive wearable technology has made seamless data aggregation possible, ultimately improving preventive care and diagnosis. Amidst all these burgeoning data points, the respiratory rate remains a crucial descriptor of health, well-being, and performance. While the traditional modes of measurement are accurate, they remain impractical for long-term respiratory rate measurement in an ambulatory setting. Interestingly, respiratory rate can be estimated from physiological signals like Electrocardiogram, Photoplethysmogram, and accelerometer waveforms. While respiration rate estimation from these methods is accurate when the subject is at rest, the estimation is thrown off by motion artifacts and a relatively poor signal-to-noise ratio during ambulatory movement. Addressing this issue, this work presents a novel Deep Learning-based multitasking network that jointly predicts both respiratory rate and the respiratory waveform, thus aiding in an overall reduction in error scores during various activities, including walking, running, etc. Apart from comparisons against the previous state-of-the-art approaches, this work thoroughly discusses the practical aspects of adopting a Deep Learning approach during inference and briefly alludes to the tradeoff between time complexity, parameter counts, and accuracy. While the proposed approach improved overall estimation accuracy, it inevitably requires more parameters and runtime than a traditional approach.