Transmuting Wearable Sensor Data: From Inertial to Electrical-Like Measurements of Heart Activity

Wearable sensor technology is revolutionizing several research fields, ranging from healthcare to fitness monitoring or biometric recognition, thanks to its many advantages against potential alternatives, such as non-invasiveness and long-term operation capabilities. In more detail, seismocardiography (SCG) and gyrocardiography (GCG) are emerging as useful tools for cardiovascular assessment, relying on inertial measurements of cardiac activity. However, the knowledge and confidence about these signals is still limited in many fields, including the medical one, where the use of electrical measurements, such as those obtained via electrocardiography (ECG), is largely preferred. This paper presents a pioneering study about the possibility of converting SCG and GCG data into ECG-like representations, with the aim of expanding the applicability of inertial wearable sensors to scenarios where their characteristics could provide relevant benefits, yet there could still be the need to exploit knowledge regarding electrical heart activity measurements. In more detail, the effectiveness of recurrent neural networks (RNNs) in performing such task is here investigated. Extensive experimentation on a public dataset demonstrates the feasibility and efficacy of the proposed method in generating signals that significantly resemble the desired ECG data. The capability of the proposed approach in reproducing relevant characteristics of ECG signals in the created data is evaluated considering two potential real-world applications, regarding heart rate estimation and ECG-based biometric recognition.