Comparison of artificial neural network (ANN) and partial least squares (PLS) regression models for predicting respiratory ventilation: an exploratory study

The objective of this study was to assess the potential for using artificial neural networks (ANN) to predict inspired minute ventilation $$ \left( {\dot{V}_{I} } \right) $$ during exercise activities. Six physiological/kinematic measurements obtained from a portable ambulatory monitoring system, along with individual’s anthropometric and demographic characteristics, were employed as input variables to develop and optimize the ANN configuration with respect to reference values simultaneously measured using a pneumotachograph (PT). The generalization ability of the resulting two-hidden-layer ANN model was compared with a linear predictive model developed through partial least squares (PLS) regression, as well as other $$ \dot{V}_{I} $$ predictive models proposed in the literature. Using an independent dataset recorded from nine 80-min step tests, the results showed that the ANN-estimated $$ \dot{V}_{I} $$ was highly correlated (R 2 = 0.88) with $$ \dot{V}_{I} $$ measured by the PT, with a mean difference of approximately 0.9%. In contrast, the PLS and other regression-based models resulted in larger average errors ranging from 7 to 34%. In addition, the ANN model yielded estimates of cumulative total volume that were on average within 1% of reference PT measurements. Compared with established statistical methods, the proposed ANN model demonstrates the potential to provide improved prediction of respiratory ventilation in workplace applications for which the use of traditional laboratory-based instruments is not feasible. Further research should be conducted to investigate the performance of ANNs for different types of physical activity in larger and more varied worker populations.