Deep Neural Network to Remove Motion Artifacts from Heart Rate Sensor Embedded on Handle Cane.

Abstract

Devices worn on the body that track physiological metrics, such as heart rate (HR) and skin conductance, have gained popularity and are typically found in items like smart-watches and bracelets. However, these measurements can be compromised by the movement of the device relative to the skin, which creates artifacts. For certain groups, such as the elderly, embedding sensors into daily-use items, like walking sticks, might offer better adherence. Nonetheless, the issue of motion artifacts becomes particularly challenging in these scenarios. This document presents a method based on a Deep Neural Network to compute the HR from a noisy signal registered by a sensor embedded in a cane. We evaluate our model in a novel dataset obtaining a mean absolute error of 9.81 ± 0.45 beats per minute, which results in a deviation of 10.75% that is in the order of the results obtained by common commercial smartwatches and bracelets.