Hypoxic burden combination model in screening low-ArTH subtype of obstructive sleep apnea

<bold>Objective:</bold> Hypoxic burden (HB) is a comprehensive measure of obstructive sleep apnea (OSA) severity, but its relationship with arousal threshold (ArTH) subtype remains unclear. This study aims to investigate this relationship and explore HB's potential in identifying low-ArTH subtype Method: Using data from the Sleep Heart Health Study (SHHS1, n=5495), SHHS2 (n=2642), and Ruijin dataset (n=393), participants were categorized into none-OSA, OSA with low-ArTH, or OSA without low-ArTH groups. Participants were categorized into three groups including none-OSA, OSA with low-ArTH, or OSA without low-ArTH based on predefined criteria. Various oxygen parameters were derived from the SpO2 signal, encompassing oxygen desaturation index (ODI2/3/4/5) and hypoxic burden (HB2/3/4/5) according to different desaturation criteria (2%, 3%, 4%, 5%), percent total recording time with different oxygen saturation (T95/90/85/80/75/70), average and nadir SpO2 (minsat). Machine learning approaches were trained using any three oximetry parameters combination. <bold>Results:</bold> Random forest (RF) emerged as a stable training method. HB and ODI consistently hold a prominent position in the top-performing combinations. The "ODI4-minsat-HB3" combination showed robust performance across the three datasets, with an F1-score of 95.2%, 78%, and 78.6% in SHHS1, SHHS2, and Ruijin dataset using RF, respectively. <bold>Conclusion:</bold> The “ODI4-minsat-HB3” machine learning model effectively identifies low-ArTH subtype in OSA, suggesting the potential of simple oximetry approaches in OSA subtype differentiation.