Expert-level classification of ventilatory thresholds from cardiopulmonary exercising test data with recurrent neural networks

Abstract

First and second ventilatory thresholds (VT₁ and VT₂) represent the boundaries of the moderate-heavy and heavy-severe exercise intensity. Currently, VTs are primarily detected visually from cardiopulmonary exercise test (CPET) data, beginning with an initial data screening followed by data processing and statistical analysis. Automated VT detection is a challenging task owing to the high signal to noise ratio typical of CPET data. Recurrent neural networks describe a machine learning form of Artificial Intelligence that can be used to uncover complex non-linear relationships between input and output variables. Here we proposed detection of VTs using a single neural network classifier, trained with a database of 228 laboratory CPET data. We tested the neural network performance against the judgement of 7 couples of board-certified exercise-physiologists on 25 CPET tests. The neural network achieved expert-level performances across the tasks (mean absolute error was 9.5% (r = 0.79) and 4.2% (r = 0.94) for VT₁ and VT₂, respectively). Estimation errors are compatible with the typical error of the current gold standard visual methodology. The neural network demonstrated VT detecting and exercise intensity level classifying at a high competence level. Neural networks could potentially be embedded in CPET hardware/software to extend the reach of exercise physiologists beyond their laboratories.

Keywords:

• Retrospective study
• cardiopulmonary exercise testing
• ventilatory threshold detection
• machine learning

Acknowledgements

We are grateful to the Fondazione Cassa di Risparmio of Trento and Rovereto (CARITRO) for partially funding this research. We thank Matteo Ragni and Paolo Rota (Department of Industrial Engineering, University of Trento, Trento, Italy) for their valuable technical help during the design of the neural networks. We thank the expert CPET evaluators from: the University of Verona (CeRiSM Research Centre, Trento, Italy and Department of Neuroscience, Biomedicine and Movement, Verona, Italy), the University of Brescia (Department of Molecular and Translational Medicine, Brescia, Italy) and Pro Motus Research Department, Bolzano/Bozen, Italy.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Andrea Zignoli http://orcid.org/0000-0003-1315-5573

Barbara Pellegrini http://orcid.org/0000-0002-9626-2632

Francesco Biral http://orcid.org/0000-0001-8098-7965

Additional information

Funding

This work was supported by the Fondazione Cassa di Risparmio di Trento e Rovereto (CARITRO) grant number 2017.0379.