Machine learning-based prediction of adherence to continuous positive airway pressure (CPAP) in obstructive sleep apnea (OSA).

Continuous positive airway pressure (CPAP) is the "gold-standard" therapy for obstructive sleep apnea (OSA), but the main problem is the poor adherence. Therefore, we have searched for the causes of poor adherence to CPAP therapy by applying predictive machine learning (ML) methods. The study was conducted on OSAs in nighttime therapy with CPAP. An outpatient follow-up was planned at 3, 6, 12 months. We collected several parameters at the baseline visit and after dividing all patients into two groups (Adherent and Non-adherent) according to therapy adherence, we compared them. Statistical differences between the two groups were not found according to baseline characteristics, except gender (P< .01). Therefore, we applied ML to predict CPAP adherence, and these predictive models showed an accuracy and sensitivity of 68.6% and an AUC (area under the curve) of 72.9% through the SVM (support vector machine) classification method. The identification of factors predictive of long-term CPAP adherence is complex, but our proof of concept seems to demonstrate the utility of ML to identify subjects poorly adherent to therapy. Therefore, application of these models to larger samples could aid in the careful identification of these subjects and result in important savings in healthcare spending.

Predictors of long-term adherence to continuous positive airway pressure in patients with obstructive sleep apnoea and acute coronary syndrome.

BACKGROUND: Continuous positive airway pressure (CPAP) is an effective treatment for obstructive sleep apnoea (OSA), but an evaluation of CPAP adherence is rarely carried out among patients with acute coronary syndrome (ACS). The goals of the study are to analyse long-term adherence and identify the predictors of non-compliance with CPAP treatment for patients with non-sleepy OSA and ACS. METHODS: This is an ancillary study of the ISAACC study, which is a multicentre, prospective, open-label, parallel, randomized, and controlled trial (NCT01335087) in patients with hospital admission for ACS. For the purpose of this study, only non-sleepy patients with moderate or severe OSA and randomized to receive CPAP treatment were analysed (n=357). Non-compliance was defined as CPAP dropout or average cumulative CPAP use of <4 hours/night. Multivariable logistic regression analysis was performed to identify predictors of CPAP adherence. RESULTS: Adherence to treatment was 35.3% at 12 months. According to the unadjusted analysis, higher apnoea-hypopnea index (AHI) (P<0.001) and oxygen desaturation index (ODI) (P=0.001) were associated with a lower risk of non-compliance. Multivariable logistic regression analysis showed that high AHI (P=0.0051), high amounts of smoking pack-year (P=0.0170), and long intensive care unit (ICU) stays (P=0.0263) were associated with lower odds of non-compliance. It also showed a significant interaction between ACS history and age (P=0.0131), such that young patients with their first ACS showed significantly lower odds of CPAP non-compliance than patients with recurrent ACS and significantly lower odds of CPAP non-compliance were associated with ageing only in patients with recurrent ACS. CONCLUSIONS: Protective factors against non-compliance with CPAP treatment in non-sleepy patients with ACS were illness severity (high values of AHI or ICU stay length) or smoking amount. Patients with no previous history of ACS showed lower odds of CPAP non-compliance than patients with a recurrent ACS with younger age.