Participants

This prospective study was approved by the institutional review board (2020-0352) and registered in the University Hospital Medical Information Network Clinical Trials Registry (UMIN-CTR, R000045230). From December 2020 to December 2024, patients visiting the Department of Psychiatry with suspected OSA were included based on the following criteria: symptoms suggestive of OSA (e.g., snoring, obesity, sleepiness), agreement to undergo MRI and PSG, and age ≥ 20 years. In total, 64 patients were screened; 50 were eligible for the initial MRI and PSG, and 10 others agreed to retesting after initiating continuous positive airway pressure therapy, resulting in 60 MRI–PSG sets. Each consisted of MRI at 4 PM, overnight PSG (≥ 7 h), and MRI at 9 AM the next morning. After excluding 10 datasets due to incomplete data (T1-weighted images, 4; diffusion, 5; PSG, 1), 50 datasets remained (MRI, 100 sessions; PSG, 50). The final cohort (n = 50) had a mean age of 50.0 (SD 14.0) years, with 33 men and 17 women. All participants were diagnosed with obstructive sleep apnea (OSA) based on overnight PSG. Among them, 8 patients had no psychiatric comorbidities and were classified as OSA-only. The remaining 42 participants had comorbid psychiatric diagnoses, including bipolar disorder (n = 14), depression (n = 14), schizophrenia (n = 7), attention-deficit/hyperactivity disorder (ADHD; n = 6), and obsessive–compulsive disorder (OCD; n = 1). (Fig. 1). Among the final 50 participants, 9 underwent follow-up MRI after initiating CPAP therapy. As a result, the dataset includes both pre- and post-CPAP cases. However, the effect of CPAP was not modeled explicitly in this analysis due to sample size limitations.

Fig. 1

Study participant flow and procedure overview. This flowchart illustrates the recruitment pathway, inclusion/exclusion details, and the final dataset composition for the analyses

Overnight PSG

PSG data were analyzed to evaluate sleep architecture, respiratory function, and autonomic responses. All participants underwent standard PSG using Embla N7000 (Natus Neurology Incorporated) or PSG-1100 (Nihon Kohden Corporation). The results were evaluated according to the American Academy of Sleep Medicine Scoring Manual (version 2.5) for the night, between the evening and morning MRI examinations [33]. Sleep architecture was assessed based on the total sleep time, sleep latency, rapid eye movement (REM) sleep latency, wake after sleep onset (WASO), sleep efficiency, and proportion of time spent in each sleep stage, including Stage N1 (light sleep with reduced muscle tone), Stage N2 (intermediate sleep characterized by sleep spindles and K-complexes), Stage N3 (deep slow-wave sleep), and REM sleep. Respiratory parameters included the arousal index, apnea index, hypopnea index, apnea–hypopnea index, lowest recorded oxygen saturation, 3% oxygen desaturation index, and total duration of time with oxygen saturation below 90%.

MRI acquisition

MR images were obtained using a 3 T clinical scanner (Vantage Centurian, Canon Medical Systems, Otawara). The first MRI session was performed at 4 PM before the PSG study (MRI/4PM). This was followed by overnight PSG. The second MRI session was performed at 9 AM the following morning after the PSG (MRI/9AM). Various images were obtained during the MRI session, including T1-weighted images and diffusion tensor images using echo-planar imaging. Diffusion tensor images was acquired by echo-planar imaging, repetition time = 6600 ms, echo time = 85 ms, diffusion time = 35.7 ms, motion-proving gradient = 12 axes, b-value = 1,000 s/mm2, anterior-to-posterior phase encode direction, FOV = 200 mm, matrix = 128 × 128 with interpolation to 256 × 256, slice thickness = 3 mm, slice number = 50, number of averaging = 2. We apply the top-up process to correct susceptibility-induced distortions; b = 0 images with posterior-to-anterior phase encoding were acquired. 3D T1-weighted MPRAGE scans were performed using the following parameters: repetition time = 6000 ms, echo time = 2.5 ms, inversion time = 800 ms, flip angle = 9°, bandwidth = 279 Hz, FOV = 250 mm, matrix = 256 × 256, slice thickness = 1.5 mm, slice number = 260.

Image analysisALPS index

DTI data were processed using the FMRIB Software Library version 6.0 (FSL; Oxford Center for Functional MRI of the Brain, Oxford, UK). Preprocessing steps included eddy current correction using FSL’s “eddy” tool and susceptibility-induced distortion correction using the “topup” method with reversed phase-encoded b0 images. All diffusion datasets were visually inspected for motion artifacts, and five datasets showing substantial artifacts were excluded prior to analysis. Diffusivity maps were obtained for each participant along the x- (right left; Dxx), y- (anterior–posterior; Dyy), and z-axes (inferior-superior; Dzz). The fractional anisotropy map for all individuals was converted into the FMRIB58_FA standard space using linear and non-linear transformations.

The participant with the smallest degree of warping was selected for the ROI placement to ensure a uniform ROI setting across all cases. The ALPS-ROIs of the projection and association areas of the left and right hemispheres were set based on previous reports [30, 32, 34, 35]. The ALPS index was calculated as the ratio of the mean x-axis diffusivity in the projection area (projDxx) to the x-axis diffusivity in the association area (assocDxx), mean y-axis diffusivity in the projection area (projDyy), and z-axis diffusivity in the association area (assocDzz), as follows:

$$} = }\left( },}} \right)/}\left( },}} \right)$$

ROI reproducibility was not assessed via inter-rater reliability because a common ALPS-ROI template was used for all participants. Specifically, a single subject with minimal non-linear warping during spatial normalization was selected, and ALPS-ROIs were defined in this subject. These ROIs were then applied to all participants in standard (FMRIB58_FA) space. This approach minimizes inter-individual variability in ROI placement. To ensure accurate registration, all non-linear warps were visually checked, and major distortions were excluded.

ALPS indices in the corpus callosum

Variant ALPS indices were set in three areas of the corpus callosum. Because the direction of the fibers and blood vessels in the corpus callosum is simple, it was thought that it would be possible to confirm the extent to which the direction of the perivascular space affects the diffusivity of the ROIs. The white matter fibers of the corpus callosum run from left-to-right or along the x-axis. In contrast, in the genu and splenium of the corpus callosum, the medullary arteries and veins run mainly in the anterior–posterior direction or the y-axis direction, while in the body of the corpus callosum, the medullary arteries and veins run mainly in the vertical direction or the z-axis direction [36, 37]. Based on these anatomical relationships, the following three variant ALPS indices related to the corpus callosum were calculated as the ratios of diffusivity along the direction of the medullary veins and arteries to that in the direction perpendicular to both the vascular orientation and the main white matter fiber orientation (Fig. 2). In the genu of the corpus callosum, ccgALPS was defined as ccgDyy/ccgDzz; in the body of the corpus callosum, ccbALPS as ccbDzz/ccbDyy; and in the splenium of the corpus callosum, ccsALPS was defined as ccsDyy/ccsDzz.

Fig. 2

Definition of variant ALPS indices. Variant ALPS indices were set for the genu, body, and splenium of the corpus callosum. Whereas the white matter fibers of the corpus callosum run left-to-right along the x-axis, the orientation of the vessels differs depending on the region. In the genu of the corpus callosum, the vessels run primarily in the anteroposterior direction along the y-axis. In the corpus callosum, the vessels mainly ran vertically along the z-axis. In the splenium of the corpus callosum, the vessels run primarily in the anteroposterior direction along the y-axis. The figure also illustrates the ALPS-ROI (projection and association areas in the white matter outside the body of lateral ventricle). Each variant of the ALPS index is defined as the ratio of diffusivity along the vascular direction to that along the perpendicular direction. In the genus of the corpus callosum, ccgALPS was calculated as ccgDyy/ccgDzz. In the body of the corpus callosum, ccbALPS was calculated as ccbDzz/ccbDyy. In the splenium of the corpus callosum, the ccsALPS was calculated as ccsDyy/ccsDzz

Choroid plexus and white matter hypointensity volumes

For the volumetric analysis of the CPV and WMHV, 3D T1-weighted images were automatically segmented using FreeSurfer version 7.4.1 (https://surfer.nmr.mgh.harvard.edu/). The “recon-all” pipelines were used for brain segmentation. These steps included motion correction, normalization, automatic segmentation, and registration. Brain volume segmentation was based on an atlas containing probabilistic information regarding the location of the structures. Images were segmented by normalizing the participant data to a common space and applying a probabilistic atlas to image-specific voxel intensities to select the optimal segmentation [38], [39]. All FreeSurfer segmentations were visually inspected for quality assurance. Cases with obvious segmentation errors, such as ventricular overestimation or cortical mislabeling, were corrected manually using Freeview by modifying the brain mask or placing control points as appropriate. The total intracranial volume, CPV, and WMHV were determined. All automatic segmentations were quality-checked and approved by a neuroradiologist blinded to the sleep data. Relative CPV (rCPV) and relative WMHV (rWMHV) were assessed as ratios to the total intracranial volume, following previously reported methods [31, 32].

Statistical analysisHead motion quantification

Framewise displacement (FD) was calculated for each subject from the motion parameters estimated during eddy current correction (FSL eddy). For each subject, mean and maximum FD values were derived. To assess potential motion-related bias, Pearson’s correlation was calculated between mean FD and the ALPS index.

Correlations between imaging and PSG parameters

To examine how imaging parameters, including the ALPS index, correlated with PSG data, correlation analyses were conducted. For MRI sessions at 4 PM and 9 AM before and after overnight PSG, correlations were assessed between imaging-derived metrics (MRI/4PM and MRI/9AM) and sleep and respiratory measures from PSG. Imaging parameters included the ALPS index, variant indices (ccgALPS, ccbALPS, ccsALPS), other diffusion metrics, rCPV, and rWMHV. To assess temporal changes, differences in ALPS index and related diffusion parameters between the 4 PM and 9 AM MRI sessions (Δ values) were calculated. These Δ values were then analyzed in relation to PSG-derived sleep parameters. Descriptive statistics for ALPS index values at both time points and for ΔALPS (9 AM–4 PM) were summarized using the mean ± standard deviation and the median with interquartile range, and are reported in the main Results section. Pearson correlation was used, assuming approximate normality of imaging variables. Statistical significance was set at p < 0.05 (two-tailed), with Benjamini–Hochberg false discovery rate correction for multiple comparisons. Correlation strength was categorized as very weak (0–0.19), weak (0.20–0.39), moderate (0.40–0.59), strong (0.60–0.79), or very strong (0.80–1.00). To support transparency and reproducibility, p value heatmaps were also generated and are presented in the Supplementary Appendix (Fig. S1). These were intended to complement the r-value heatmaps presented in the main text by providing a statistical significance reference for each correlation.

Correlation analysis of diffusion parameters with rCPV and WHMVEffect of ALPS index and variant ALPS indices on rCPV and rWHMV

To examine how the ALPS index and its variants correlate with glymphatic indicators, we performed correlation and multiple regression analyses using the full dataset, including MRI scans at 4 PM and 9 AM. Correlation coefficients and p values were calculated between the ALPS index, its variants, rCPV, and rWMHV, with FDR correction applied.

Multiple regression was then used to assess the relationship between diffusion indices and rCPV or rWMHV, adjusting for age and psychiatric conditions using 0/1 dummy variables. To account for potential confounding effects from psychiatric comorbidities and psychotropic medications, we included psychiatric diagnoses as covariates in the models. Although these factors may influence brain structure and function, this adjustment aimed to mitigate their impact within our statistical analyses. All continuous variables were standardized (mean = 0, SD = 1). Standardized regression coefficients and 95% confidence intervals were computed, and variance inflation factors were checked to rule out multicollinearity. Forest plots visualized these coefficients and their intervals. Statistical significance was defined as a 95% CI not crossing zero (≈ p < 0.05). Analyses were done separately for rCPV and rWMHV. To confirm the robustness of the findings, we additionally performed two sensitivity analyses: a subgroup analysis limited to OSA participants without psychiatric comorbidities (n = 8, 16 scans), and an analysis excluding the 9 participants who had received prior CPAP treatment. These complementary analyses allowed us to examine whether the main associations held true under more homogeneous clinical conditions.

Effects of diffusion parameters other than the ALPS index on rCPV and rWMHV

To investigate whether diffusivity metrics not used in the ALPS index formulation are also associated with glymphatic function, we performed additional multiple regression analyses using individual diffusion parameters obtained from the ALPS-ROI and corpus callosum ROI. These models included projection (projDyy, projDzz), association (assocDyy, assocDzz), and corpus callosum (ccbDxx, ccbDyy, ccbDzz) diffusivity values as predictors, adjusting for age and psychiatric status. As with the primary models, all variables were standardized and statistical significance was assessed based on the 95% confidence interval not crossing zero.

Evaluation of the effects of ALPS-ROI and corpus callosum diffusion parameters on the ALPS index

To assess whether diffusivity components orthogonal to the perivascular direction influence the ALPS index, mediation analyses were conducted using directional diffusivity measures within the ALPS-ROI and corpus callosum. In each model, the diffusivity aligned with the perivascular direction (e.g., assocDxx or projDxx) was set as the independent variable, and orthogonal diffusivity components (e.g., assocDyy, assocDzz, projDyy, projDzz) were tested as potential mediators. Similarly, to examine the influence of commissural fibers, mediation analyses were also performed using corpus callosum diffusivity measures (e.g., ccbDyy, ccbDzz) as mediators of the relationship between ALPS-ROI diffusivity and the ALPS index. All models were adjusted for age. Average causal mediation effects (ACME) were estimated using nonparametric bootstrapping, and statistical significance was defined by a 95% confidence interval not crossing zero. Both full and partial mediation patterns were considered.

The results of all 10 mediation models are summarized in Table 2, which presents the standardized coefficients (β) and 95% confidence intervals (CI) for the average causal mediation effect (ACME), average direct effect (ADE), and total effect derived from bootstrap simulations (1000 iterations). Mediation pathways are denoted using the format X (Independent Variable) → M (Mediator) → Y (Dependent Variable) to clarify the directionality of influence. The table is structured in alignment with Fig. 5, beginning with intra-regional mediation within the ALPS-ROI, followed by corpus callosum-related mediations, allowing for direct comparison between graphical and quantitative results.

Evaluation of the impact of the direction of medullary vessel on diffusivity

To assess whether the orientation of medullary vessels relative to white matter fibers affects diffusion characteristics, variant ALPS indices were calculated in three sub-regions of the corpus callosum: the genu, body, and splenium. These regions were selected because the fiber orientation is relatively homogeneous, and the course of medullary vessels in each region is anatomically distinct. Each variant index (ccgALPS, ccbALPS, and ccsALPS) was defined as the ratio of diffusivity along the presumed vascular direction to that along the orthogonal direction. One-sample t tests were planned to compare the mean values of each variant ALPS index against a reference value of 1.0, which represents isotropic diffusion.