Can a prediction model combining self-reported symptoms, sociodemographic and clinical features serve as a reliable first screening method for sleep apnea syndrome in patients with stroke?

OBJECTIVE: To determine whether a prediction model combining self-reported symptoms, sociodemographic and clinical parameters could serve as a reliable first screening method in a step-by-step diagnostic approach to sleep apnea syndrome (SAS) in stroke rehabilitation.
DESIGN: Retrospective study.
SETTING: Rehabilitation center.
PARTICIPANTS: Consecutive sample of patients with stroke (N=620) admitted between May 2007 and July 2012. Of these, 533 patients underwent SAS screening. In total, 438 patients met the inclusion and exclusion criteria.
INTERVENTIONS: Not applicable.
MAIN OUTCOME MEASURES: We administered an SAS questionnaire consisting of self-reported symptoms and sociodemographic and clinical parameters. We performed nocturnal oximetry to determine the oxygen desaturation index (ODI). We classified patients with an ODI ≥15 as having a high likelihood of SAS. We built a prediction model using backward multivariate logistic regression and evaluated diagnostic accuracy using receiver operating characteristic analysis. We calculated sensitivity, specificity, and predictive values for different probability cutoffs.
RESULTS: Thirty-one percent of patients had a high likelihood of SAS. The prediction model consisted of the following variables: sex, age, body mass index, and self-reported apneas and falling asleep during daytime. The diagnostic accuracy was .76. Using a low probability cutoff (0.1), the model was very sensitive (95%) but not specific (21%). At a high cutoff (0.6), the specificity increased to 97%, but the sensitivity dropped to 24%. A cutoff of 0.3 yielded almost equal sensitivity and specificity of 72% and 69%, respectively. Depending on the cutoff, positive predictive values ranged from 35% to 75%.
CONCLUSIONS: The prediction model shows acceptable diagnostic accuracy for a high likelihood of SAS. Therefore, we conclude that the prediction model can serve as a reasonable first screening method in a stepped diagnostic approach to SAS in stroke rehabilitation.