Artificial neural network identification of exercise expiratory flow-limitation in adults

Abstract

Abstract Identification of ventilatory constraint is a key objective of clinical exercise testing. Expiratory flow-limitation (EFL) is a well-known type of ventilatory constraint. However, EFL is difficult to measure, and commercial metabolic carts do not readily identify or quantify EFL. Deep machine learning might provide a new approach for identifying EFL. The objective of this study was to determine if a convolutional neural network (CNN) could accurately identify EFL during exercise in adults in whom baseline airway function varied from normal to mildly obstructed. 2931 spontaneous exercise flow-volume loops (eFVL) were placed within the baseline maximal expiratory flow-volume curves (MEFV) from 22 adults (15 M, 7 F; age, 32 yrs) in whom lung function varied from normal to mildly obstructed. Each eFVL was coded as EFL or non-EFL, where EFL was defined by eFVLs with expired airflow meeting or exceeding the MEFV curve. A CNN with seven hidden layers and a 2-neuron softmax output layer was used to analyze the eFVLs. Three separate analyses were conducted: (1) all subjects (n = 2931 eFVLs, [GRALL]), (2) subjects with normal spirometry (n = 1921 eFVLs [GRNORM]), (3) subjects with mild airway obstruction (n = 1010 eFVLs, [GRLOW]). The final output of the CNN was the probability of EFL or non-EFL in each eFVL, which is considered EFL if the probability exceeds 0.5 or 50%. Baseline forced expiratory volume in 1 s/forced vital capacity was 0.77 (94% predicted) in GRALL, 0.83 (100% predicted) in GRNORM, and 0.69 (83% predicted) in GRLOW. CNN model accuracy was 90.6, 90.5, and 88.0% in GRALL, GRNORM and GRLOW, respectively. Negative predictive value (NPV) was higher than positive predictive value (PPV) in GRNORM (93.5 vs. 78.2% for NPV vs. PPV). In GRLOW, PPV was slightly higher than NPV (89.5 vs. 84.5% for PPV vs. NPV). A CNN performed very well at identifying eFVLs with EFL during exercise. These findings suggest that deep machine learning could become a viable tool for identifying ventilatory constraint during clinical exercise testing.