Enabling Clinically Relevant and Interpretable Deep Learning Models for Cardiopulmonary Exercise Testing

Abstract

Cardiopulmonary exercise testing (CPET) provides a safe, objective, and reliable assessment of cardiorespiratory fitness and is a valuable method used by clinical practitioners to predict and improve patient outcomes. However, CPET produces complex data consisting of multiple time-series that requires specialized training to interpret. This paper demonstrates accurate disease diagnosis by the use of deep learning models applied to these data using a small set of patients with known health conditions. Despite limited data availability, data augmentation enabled predictions with that consistently outperformed traditional interpretation methods and produced models that focused on clinically relevant regions of the multivariate time-series. Visual explanations of model decisions, projected through the nine-panel plot commonly used to interpret CPET, demonstrate the clinical relevance of model features, and provide insights that can benefit future training, interpretation, and research.Clinical relevance—This method can assist clinical practitioners by providing interpretable and reliable diagnosis recommendations with CPET data.