Development of a prediction model for exertional desaturation by machine learning method

It has been well known that exertional desaturation is a poor prognostic factor in chronic obstructive pulmonary disease (COPD). Exertional desaturation can be identified during 6-minute walk test (6MWT). However, 6MWT is not routinely performed in clinical practice. Thus, it is difficult to predict exertional desaturation in patients with COPD. The purpose of this study is to develop a prediction model for exertional desaturation by machine leaning method. The data from the Korean COPD subgroup study (KOCOSS) cohort was utilized. Exertional desaturation was defined as oxygen saturation less than 90% during 6MWT. Gradient Boosting was utilized to develop a prediction model. Among 3,676 patients enrolled in this cohort, 1,807 patients performed 6MWT. The prevalence of exertional desaturation was 9.6% (173/1,807). Age, sex, smoking status, comorbidities, lung function, COPD assessment test score and number of exacerbations were included in the model development. The prediction power of developed model was modest. Area Under the Curve was 0.732 and accuracy was 0.946. The feature importance for this model was illustrated in Figure 1. Lung function, BMI, and age were important factors. Prediction model for exertional desaturation can be useful in clinical practice. This may lead to apply home oxygen earlier in high-risk patients. Funding: Korean Environment Industry and Technology Institute (2022003310008), Korea National Institute of Health (2016ER670100, 2016ER670101, 2016ER670102, 2018ER67100, 2018ER67101, 2018ER67102, 2021ER120500, 2021ER120501 and 2021ER120502)