Abstract W P128: Development and Validation of Clinical Prediction Model for Obstructive Sleep Apnea among Patients with Ischemic Stroke and TIA

Introduction: Obstructive sleep apnea (OSA) screening instruments that perform well in the general population have demonstrated poor correlation with OSA diagnosis in stroke and TIA populations. Given that OSA is common post-stroke/TIA, has been associated with recurrent vascular events, and is treatable, we sought to develop and validate a cerebrovascular disease specific prediction model for OSA. Methods: We developed and validated a clinical prediction model using data from two separate multi-site randomized, controlled strategy trials: the development population included chronic stroke/TIA patients; the validation population included acute stroke/TIA patients. Data on patient demographics, vital signs, anthropomorphic measurements, past medical history, medications, NIHSS, sleep questionnaires, and polysomnography (PSG) were obtained. OSA was defined as present if the apnea-hypopnea index (AHI) was ≥5 events/hour. Variables considered clinically important and required to remain in the multivariable logistic regression model were gender, weight, neck circumference (women >16”, men >17”), history of CHF, DM and NIHSS. Variables associated with OSA in bivariate analyses (p≤0.25) were also included: age, history of chronic obstructive COPD, automatic implantable cardioverter defibrillator (AICD), Charlson comorbidity, Epworth Sleepiness Scale (ESS), Berlin Questionnaire ≥10, and patient health questionnaire (PHQ8) depression scale. Backward elimination was used to identify factors that were independently associated with OSA (p<0.05). Results: Sleep apnea occurred commonly in the development (119/194; 61%) and validation (75/96; 76%) cohorts. The final model included: gender, CHF, DM, large neck, and the quadratics of weight, NIHSS, and ESS. The model showed reasonable discrimination in the development (c-statistic 0.73) and validation (c-statistic 0.70) cohorts. This model correctly classified 68% of patients in the validation sample as having OSA. Conclusions: A model based on readily available clinical information predicted the presence of OSA reasonably well. Future work is required to determine the utility of this model in clinical practice for the identification of patients who should receive OSA management.