Résumé
RATIONALE: Long-term oxygen therapy (LTOT) delays the progression of Chronic Obstructive Pulmonary Disease (COPD) and other forms of chronic respiratory failure (CRF) and improves survival. We aimed to identify the prognostic factors associated with survival and with the clinical decision of portable oxygen concentrator (POC) type. METHODS: The study applies machine learning to predict two outcomes for LTOT patients: 1) the survival; 2) the choice between POC with autonomy higher/lower than 5 hours, defined herein higher/lower mobility (HM/LM). This is a nationwide retrospective analysis of COPD- and CRF-adult patients based on the French health administrative database SNDS. Patients were included at the first delivery of LTOT device, between 2014 to 2019, and followed up to December 2020. Socio-demographic information (age, sex, deprivation index, residence in rural/urban areas), comorbidities, type of oxygen-delivery equipment and interaction between age and comorbidities were included as features. LTOT patients were stratified into two groups: high (HS) and low (LS) survival. Three machine learning models, classification and regression tree (CART), random forest (RF) and neural networks (NNs), were trained on a balanced random subset (80%) of data. Optimal hyperparameters and accuracy were obtained. RESULTS: Survival analysis was performed on 152,516 LTOT-equipped COPD and CRF patients, equally divided in LS and HS groups. According to the CART algorithm the presence of chronic respiratory disease predicted LS with 73% accuracy, followed by age, number of comorbidities and lung cancer. The RF identified chronic respiratory interaction with age or obesity as the most important variables based on Gini metric. Accuracy for predicting HS was similar between methods, around 78%. For the analysis of the type of POC prescribed, 18,630 patients were identified in the LM and HM groups. In the CART model, the most important variables for prediction of HM device use were the age lower than 79 and chronic respiratory disease, followed by the absence of heart failure and obesity, whereas the age over 88 or neurological degenerative disease predicted the LM use. The RF correctly predict HM use in 70% of cases, LM in 48%, confirming those variables as the most important based on Gini metric. Accuracy for predicting HM POCs was stable between methods, around 57%. CONCLUSIONS: Chronic respiratory disease and age were the strongest predictors of survival and type of POCs. The machine learning approach could be used to support clinical decision-making towards personalized therapies, with the goal of improving LTOT care and survival.