Preliminary results on using machine learning methods to predict exercise induced hypoxemia of endurance athletes a posteriori

This study evaluated whether a machine learning (ML) model could retrospectively classify exercise-induced hypoxemia (EIH), based on simple cardiorespiratory parameters measured during a maximal exercise test (MaxT). A dataset comprising 165 athletes (81 with EIH and 84 without) who participated in studies evaluating oxygen saturation levels in conjunction with gas exchange measurements was created. It included maximum, minimum, mean, median, standard deviation and root mean square of demographic, training and exercise data measured during a MaxT, including ventilation (V E ), oxygen consumption (VO 2 ), carbon dioxide production (VCO 2 ), heart rate (HR), and ratios (V E /VCO 2 , VCO 2 /VO 2 , V E /VO 2 , and VO 2 /HR). Both unsupervised and supervised ML methods were employed. Each method was evaluated based on accuracy, Cohen’s kappa coefficient, F1 score, Receiver Operating Characteristics Curve (ROC) and the related Area Under the Curve (AUC). Unsupervised and linear models were not sufficiently robust to predict EIH (accuracy below 60%). Conversely, the non-linear supervised model, particularly XGBoost, achieved 85% accuracy, with k = 0.69, an F1 score of 0.81 and an AUC of 0.79. Maximal V E , minimal VO 2 , the standard deviation of the VCO 2 /VO 2 ratio, as well as maximal and median VCO 2 , emerged as the most informative features in our model. Furthermore, EIH athletes exhibited higher mean values for minimal VO 2 , as well as median and maximal VCO 2 . Our results suggest that ML techniques could be used to develop a model for classifying male athletes as EIH and non-EIH without requiring oxygen saturation measurements. This approach could allow for retrospective re-evaluation of scientific findings in exercise physiology, and to explain inter-individual variations, given that EIH induces specific adaptations to exercise in normoxia and hypoxia.