Contributors: Filosa, M; Massari, L; Ferraro, D; D'Alesio, G; D'Abbraccio, J; Aliperta, A; Presti, D; Di Tocco, J; Zaltieri, M; Massaroni, C; Carrozza, M; Ferrarin, M; Di Rienzo, M; Schena, E; Oddo, C
نبذة مختصرة : The continuous monitoring of an individual's breathing can be an instrument for the assessment and enhancement of human wellness. Specific respiratory features are unique markers of the deterioration of a health condition, the onset of a disease, fatigue and stressful circumstances. The early and reliable prediction of high-risk situations can result in the implementation of appropriate intervention strategies that might be lifesaving. Hence, smart wearables for the monitoring of continuous breathing have recently been attracting the interest of many researchers and companies. However, most of the existing approaches do not provide comprehensive respiratory information. For this reason, a meta-learning algorithm based on LSTM neural networks for inferring the respiratory flow from a wearable system embedding FBG sensors and inertial units is herein proposed. Different conventional machine learning approaches were implemented as well to ultimately compare the results. The meta-learning algorithm turned out to be the most accurate in predicting respiratory flow when new subjects are considered. Furthermore, the LSTM model memory capability has been proven to be advantageous for capturing relevant aspects of the breathing pattern. The algorithms were tested under different conditions, both static and dynamic, and with more unobtrusive device configurations. The meta-learning results demonstrated that a short one-time calibration may provide subject-specific models which predict the respiratory flow with high accuracy, even when the number of sensors is reduced. Flow RMS errors on the test set ranged from 22.03 L/min, when the minimum number of sensors was considered, to 9.97 L/min for the complete setting (target flow range: 69.231 ± 21.477 L/min). The correlation coefficient r between the target and the predicted flow changed accordingly, being higher (r = 0.9) for the most comprehensive and heterogeneous wearable device configuration. Similar results were achieved even with simpler settings which included the ...
No Comments.