Abstract
Purpose
This study aims to develop sleep apnea screening models using a large clinical sleep dataset of SpO2 data, with the goal of achieving better performance and generalizability compared to existing models.
Methods
We utilized SpO2 recordings from the Sleep Heart Health Study database (N = 5667). Probabilistic ensemble machine learning was employed to predict sleep apnea status at three AHI cutoff points: ≥5, ≥ 15, and ≥ 30 events/hour. To investigate the impact of data granularity, SpO2 data were resampled to 1/30, 1/60, and 1/300 Hz. Model performance was evaluated across various decision boundaries ranging from 0.05 to 0.95.
Results
Our models demonstrated good to excellent performance, with AUC values of 0.82, 0.85, and 0.90 for cutoffs ≥ 5, ≥15, and ≥ 30, respectively. Sensitivity ranged from good to excellent (0.76, 0.84, 0.89), while specificity ranged from good to excellent (0.87, 0.86, 0.90). Positive predictive values (PPV) ranged from fair to excellent (0.97, 0.83, 0.66), and negative predictive values (NPV) ranged from low to excellent (0.43, 0.87, 0.98). Both decision boundaries and data granularity had a significant impact on model performance, with optimal decision boundaries aligning with the prevalence of positive cases in the cohort. Lower data granularity resulted in decreased model performance.
Conclusion
Our models demonstrated superior performance across all three AHI cutoff thresholds compared to existing large sleep apnea screening models, even when considering varying SpO2 data granularity. The use of probabilistic ensemble machine learning shows promises for developing generalizable sleep apnea screening models with overnight SpO2 data.