Can We Avoid Unnecessary Polysomnographies in the Diagnosis of Obstructive Sleep Apnea? A Bayesian Network Decision Support Tool

摘要

Obstructive Sleep Apnea (OSA) affects 2-4% of the population worldwide. The standard test for OSA diagnosis is polysomnography (PSG), an expensive exam limited to urban areas. Furthermore, nearly half of all PSG tests results are negative for OSA. This work aims to reduce these unnecessary exams, by defining an auxiliary diagnostic method that could be used to assess patient's need for PSG, according to their probability of OSA diagnosis. A prospective study was conducted on adult patients with OSA suspicion who performed PSG at our sleep laboratory in Portugal. The studied clinical variables were defined after literature review and collected during consultation. Two comparable cohorts were studied for derivation (n=86) and validation (n=33) of models. Three classifiers were analyzed - a multiple logistic regression classifier (AUC=80.0%) and two Bayesian networks classifiers - Naïve Bayes (AUC=81.3%) and Tree Augmented Naïve Bayes (TAN, AUC=81.4%) - aiming at the best possible specificity (identification of unnecessary exams). Overall, sensitivity-adjusted models could detect normal patients, preventing unnecessary PSG, while keeping sensitivity high. Furthermore, the graphical representation of TAN can be explored by the physician during consultation, making it a helpful tool to assess patients' need to perform PSG.