Identification Of Exercise-induced Expiratory Flow-limitation Using Deep Machine Learning In Healthy Adults Across The Lifespan

BACKGROUND: Expiratory flow limitation (EFL) is an indicator of ventilatory constraint that provides important information in the context of cardiopulmonary exercise testing (CPET), but its widespread use during CPET has been hampered by technical and logistical challenges. To overcome these challenges, we developed a convolutional neural network (CNN) to identify EFL; however, this CNN requires validation in a larger cohort with a wider age range. It is also unclear if accounting for the effects of thoracic gas compression and exercise-induced bronchodilation when assessing EFL influences CNN performance. PURPOSE: To determine the effectiveness of a CNN at identifying exercise-induced EFL in healthy adults across the lifespan. METHODS: n = 82 healthy adults (aged 20-80 y; n = 40 males, n = 42 females) completed pulmonary function testing and an incremental CPET on a cycle ergometer. During CPET, EFL was assessed by overlaying tidal expiratory flow-volume (TEFV) curves with maximal expiratory flow-volume curves generated: i) from a forced vital capacity manoeuvre performed prior to exercise (MEFVPRE), and ii) by combining a series of forced vital capacity manoeuvres performed at varying efforts before and after exercise to account for the effects of thoracic gas compression and exercise-induced bronchodilation (MEFVCOMB). TEFV curves were input into the CNN with one of three labels: 1) no EFL based on MEFVPRE and MEFVCOMB, 2) EFL based on MEFVPRE, or 3) EFL based on MEFVCOMB. A CNN with seven hidden layers and a 4-neuron softmax output layer was used to analyze TEFV curve geometry and classify each curve as EFL or no EFL according to labels 1-3. RESULTS: The CNN processed 6647 TEFV curves; 4453 (67 %) were used for training, and 2194 (33 %) were used for testing. Overall, the CNN correctly classified 85 % of the TEFV curves, with 94 % correctly classified as no EFL based on MEFVPRE and MEFVCOMB, 36 % correctly classified as EFL based on MEFVPRE, and 73 % correctly classified as EFL based on MEFVCOMP. CONCLUSION: These data suggest that our CNN is an effective tool to assess EFL during exercise. However, the CNN’s poor performance in categorizing TEFV curves based on MEFVPRE emphasizes the importance of accounting for the effects of thoracic gas compression and exercise-induced bronchodilation in the assessment of EFL.