Co-classification network analysis reveals nodal dysfunction and dimension-specific alterations in obstructive sleep apnea-hypopnea syndrome.

BACKGROUND

Obstructive sleep apnea-hypopnea syndrome (OSAHS) is a prevalent sleep disorder linked to brain alterations, but its brain network patterns and convenient screening methods remain unclear. This study aimed to characterize OSAHS functional brain networks via co-classification network analysis and assess their diagnostic utility.

METHODS

174 participants were stratified by the apnea-hypopnea index (AHI) into none or mild (nm-OSAHS) and moderate or severe (ms-OSAHS) groups, and by the lowest oxygen saturation (SpO2 nadir) into none or mild (nm-hypoxemia) and moderate or severe (ms-hypoxemia) groups. (Registered number: MR-32-25-040098). They underwent out-of-center sleep testing, neuropsychological tests, and MRI. Co-classification networks (via consensus modularity analysis) and traditional functional networks (via graph theory) were analyzed, with random forest model for diagnostic accuracy.

RESULTS

Ms-OSAHS and ms-hypoxemia groups showed worse sleep respiratory parameters. Co-classification network analysis revealed significant alterations in intra-module z score, participation coefficient (PC), and diversity coefficient (SD) in the visual, dorsal attention, salience/ventral attention, and executive control networks in ms-OSAHS patients. PC in the left temporo-occipital junction and right parietal operculum was positively correlated with AHI, while PC in limbic regions was negatively correlated with lowest SpO2. The random forest model using co-classification metrics demonstrates good diagnostic performance for OSAHS, with dorsal attention, control, and salience networks mostly contributing to classification.

CONCLUSIONS

OSAHS induced widespread nodal dysfunction in brain networks, with distinct changes driven by hypoxemia and apnea frequency. Co-classification network analysis outperforms traditional methods in detecting network abnormalities, and machine learning models based on nodal parameters show high diagnostic accuracy, suggesting potential for OSAHS screening.