Estimation of respiratory volume from thoracoabdominal breathing distances: comparison of two models of machine learning

We compared two models of machine learning (ML) for estimating \(\widehat{V}_\mathrm{RMP}\): a linear model (multiple linear regression—MLR) and a nonlinear model (artificial neural network—ANN), and we used cross-validation to validate these models. Fourteen healthy adults, aged \(24.1 \pm 3.4\) years participated in the present study. The protocol was conducted in a laboratory test room. The anteroposterior displacements of the rib cage and abdomen, and the axial displacements of the chest wall and spine were measured using two pairs of magnetometers. \(\widehat{V}_\mathrm{RMP}\) was estimated from these four signals, and the respiratory volume was simultaneously measured using a spirometer (\(V_\mathrm{SP}\)) under lying, sitting and standing conditions as well as various exercise conditions (working on computer, treadmill walking at 4 and 6 km\(\mathrm{ \ h}^{-1}\), treadmill running at 9 and 12  km \(\mathrm{ \ h}^{-1}\) and ergometer cycling at 90 and 110 W).

The results from the ANN model fitted the spirometer volume significantly better than those obtained through MLR. Considering all activities, the difference between \(\widehat{V}_\mathrm{RMP}\) and \(V_\mathrm{SP}\) (bias) was higher for the MLR model (\(0.00191 \pm 0.141\) L) than for the ANN model (\(0.00158 \pm 0.150\) L).