Multivariate Model-based Tool for Predicting Cardiovascular Outcomes Using Multimodal Physiological Sleep Data.

Obstructive sleep apnea (OSA) represents a potential risk factor for cardiovascular disease (CVD), but the clinical diagnostic marker, the Apnea-Hypopnea Index (AHI), may not fully capture its long-term cardiovascular impact. This study examines a wide range of sleep-related physiological features in OSA and non-OSA subjects to identify the most informative predictors of CVD outcomes. The database included multimodal sleep metrics from polysomnographic (PSG) recordings, clinical data, and relevant risk factors. An average of 10 years of follow-up was also available for longitudinal study of CVD outcomes. A two-step feature selection approach was used to reduce the initial database. Irrelevant features were discarded, followed by selection of highly informative features using the least absolute shrinkage and selection operator (LASSO). The final, well balanced model, was selected using a wrapper-based method. Model performance was robustly evaluated using repeated holdout validation, with an accuracy of 75.3% using only 10 features (AUC = 78.0%). In addition to known risk factors such as sex, age, BMI and diabetes, additional physiological markers emerged as significant covariates. In particular, nocturnal average SpO2 values (total and during NREM stages), high heart rates during apneic episodes with SpO2 drops >3%, blood pressure levels, and the frequency of central apneas per hour of REM or NREM sleep in specific sleep positions were identified as potential biomarkers associated with cardiovascular risk. These findings underscore the importance of combining multimodal sleep physiological metrics for improved predictive models of CVD risk stratification beyond AHI in OSA patients, allowing for more targeted clinical interventions.Clinical relevance-Integrating multiple sleep physiological metrics beyond AHI improves cardiovascular risk stratification in OSA patients, enabling more precise clinical interventions.