Development of a Screening Model for Exercise-Induced Desaturation by Machine Learning Method

Exercise-induced desaturation (EID) during the 6-minute walk test (6MWT) is an established marker of adverse outcomes in patients with chronic obstructive pulmonary disease (COPD). We therefore sought to develop a screening-oriented machine learning approach to identify patients at increased risk of EID.

We analyzed data from the nationwide, multicenter Korea COPD Subgroup Study (KOCOSS). EID was defined as peripheral oxygen saturation (SpO2) < 90% with a decrease of ≥ 4%p. The cohort was stratified into training (80%) and test (20%) sets. Candidate predictors were selected using the Boruta algorithm, and models were developed using multivariable logistic regression (MLR), extreme gradient boosting (XGB), random forest (RF), and support vector classification (SVC), with a screening-oriented threshold strategy prioritizing sensitivity.

Among 1,788 patients with COPD, 185 (10.3%) exhibited EID. All models showed high area under the precision-recall curve (PR-AUC) during internal validation. Predictors selected by the Boruta algorithm included body mass index, COPD Assessment Test, St. George's Respiratory Questionnaire-C, mental health indicators, pulmonary function parameters, X-ray-identified bronchiectasis, and hemoglobin level. In the independent test set, PR-AUC declined across models while calibration metrics showed modest differences between datasets. The XGB model achieved the highest sensitivity in internal validation and its sensitivity and specificity remained relatively stable in the test set. Baseline SpO2 and diffusion capacity of the lung for carbon monoxide were the most influential predictors.

A screening-oriented machine learning approach using routinely available variables may facilitate targeted referral for the 6MWT in COPD.