Exploring the Role of Nocturnal Hypoxemia and Sleep Fragmentation in Memory Decline: Insights From Explainable Machine Learning Models.

Introduction: Sleep-disordered breathing (SDB) is linked to memory decline, but the exact relationship between sleep fragmentation, nocturnal hypoxemia, and cognitive impairment remains unclear.

Objectives: This study aimed to investigate the associations between micro-arousal burden, nocturnal oxygen desaturation, and memory decline in patients with moderate-to-severe OSA.

Methods: Data were retrieved from the clinical and overnight polysomnographic (PSG) records of adult patients evaluated for suspected SDB. The primary clinical endpoint was the presence and severity of memory decline, ascertained via a standardized Subjective Cognitive Decline (SCD) instrument. A multidimensional array of variables was systematically extracted, encompassing baseline demographic characteristics, cardiometabolic comorbidities, and high-resolution sleep architecture metrics, with a distinct emphasis on stage-specific micro-arousal burdens and the morphological profiles of nocturnal oxygen desaturation. Then, independent t tests and $x^2$ tests were initially utilized to characterize PSG disparities between the memory-normal and memory-decline groups. And interpretable machine learning algorithms, utilizing rigorously partitioned training and validation sets, were deployed to predict cognitive trajectories and elucidate the relative prognostic importance of specific sleep-related parameters.

Results: The final analytical sample comprised 884 participants with complete primary outcome data (memory-normal: N = 408; memory-decline: N = 476). Initial comparative analyses revealed the memory-decline group was older (50.24 vs. 45.95 years, p < 0.001) with a significantly higher prevalence of cardiometabolic comorbidities, including hypertension (47.3% vs. 40.2%, p = 0.035) and diabetes (24.4% vs. 8.8%, p < 0.001). Polysomnographically, this group exhibited a distinct hypopnea-predominant phenotype: despite a comparable overall AHI (45.82 vs. 48.64 events/h, p = 0.099) and global arousal index (26.98 vs. 28.85 events/h, p = 0.172), they demonstrated a significantly higher hypopnea count (122.25 vs. 110.40, p = 0.047) and prolonged awake time with SpO₂ < 95% (33.71 vs. 27.71 min, p = 0.015). Paradoxically, their nadir SpO₂ was elevated (76.68% vs. 74.39%, p = 0.009), maximal obstructive events were shorter (51.42 s vs. 57.49 s, p < 0.001), and obstructive desaturation events were fewer (180.33 vs. 219.70, p = 0.006), indicating a shift toward shallower, persistent desaturation morphologies. Furthermore, interpretable machine learning models, rigorously evaluated on the independent validation set, identified spontaneous NREM micro-arousals, total REM micro-arousals, and obstructive desaturation metrics as the highest-ranking predictive determinants of memory decline.

Conclusions: Memory decline in SDB is more robustly associated with the morphological profile of oxygen exposure rather than absolute event frequencies. A hypopnea-dominant profile with mild, persistent low oxygen levels offers an associative framework for understanding cognitive decline. Future research and clinical interventions should prioritize hypoxic burden as a key factor in phenotype identification and memory decline treatment.