White matter microstructure differences in obstructive sleep apnea severity groups assessed by diffusion tensor metrics and biophysical modeling

Abstract

Obstructive sleep apnea (OSA) is a complex condition characterized by repeated episodes of upper airway collapse during sleep, leading to chronic intermittent hypoxia. Diffusion magnetic resonance imaging (dMRI) techniques offer sensitivity to white matter (WM) microstructure changes. 150 individuals from a community-based study underwent one-night nocturnal polysomnography (NPSG), cognitive assessments, and brain structural MRI. Gaussian and non-Gaussian diffusion signal changes in WM tracts were quantified with diffusion tensor metrics as (DTI) and Diffusion Kurtosis Imaging (DKI), respectively. While changes in WM microstructure were assessed in terms of Standard Model metrics. The genu of the corpus callosum (GCC) demonstrated negative correlations between AHI3A and FA (p < 0.01), AD (p < 0.05), f (SMI-based axonal water fraction) (p < 0.05), and p² (p < 0.05) (SMI-based extra-axonal water), alongside positive correlations with RD (p < 0.05). The right cingulum showed negative associations with FA (p < 0.01), RK (p < 0.01), f (p < 0.01) and (p < 0.01). Subjects without OSA showed higher values in FA (p = 0.001), AD (p = 0.01), f (p = 0.03), and (p = 0.0006) in the GCC and cingulum. The strongest differences between severity groups were observed between AHI3A(0–5/h) and AHI3A(> 30/h), particularly in the GCC FA (p = 0.001), RD (p = 0.008), and RK (p = 0.02), and the cingulum f (p ≤ 0.01). Decreases in AD, RK, and FA, and increased RD with increasing OSA severity suggest demyelination and axonal loss. This contrasts with f, a direct measurement for axonal density, which was lower in the OSA group, demonstrating that OSA affects the WM microstructure. Future studies should include longitudinal evaluations to assess the effects of disease duration, and the clinical significance of changes in dMRI metrics over time.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39162-7.

Introduction

Obstructive sleep apnea (OSA) is a complex condition characterized by repeated episodes of upper airway collapse during sleep, leading to sleep fragmentation and cognitive impairment. One of the hallmarks of OSA is the presence of chronic intermittent hypoxia, which starts a cascade of molecular and cellular events, resulting in damage to cell bodies and axons³ and contributing to their degeneration¹. In recent years, white matter integrity measured by diffusion magnetic resonance imaging (dMRI) techniques, such as Diffusion Tensor Imaging (DTI) and Diffusion Kurtosis Imaging (DKI), has emerged as a potential biomarker for neurodegeneration², early cerebrovascular disease³, and brain injury related to sleep disturbance⁴.

dMRI is a clinically helpful MRI (magnetic resonance imaging) method to probe brain microstructure by leveraging the sensitivity of water diffusion at the micrometer level to the restrictions provided by biological tissue⁵. The most commonly applied dMRI technique is diffusion tensor imaging (DTI)⁶, which quantifies the Gaussian properties of the dMRI signal. Furthermore, DKI is a clinically feasible extension of DTI that also enables the quantification of non-Gaussian diffusion through the estimation of the diffusional kurtosis, a quantitative measure of the non-Gaussianity of the diffusion process^7,8, providing information about the tissue’s complexity^9,10.

While the empirical DTI and DKI metrics capture diffusion signal changes in pathology, the relation to changes in tissue microstructure can be established using biophysical modeling⁵. The Standard Model (SM) of diffusion^5,11 has been suggested for the white matter as a modeling framework encompassing previous approaches^9,12,13 with anisotropic Gaussian compartments. SM assumes non-exchanging intra- and extra-axonal compartments with an arbitrary fiber orientation dispersion function (ODF) and yields the corresponding parameters (fractions and diffusivities) of the intra-axonal and extra-axonal spaces within a WM (White Matter) voxel, as well as the fiber ODF and its invariants^5,11. The SM has been histologically validated in animal models^14–19. The open-source SM parameter estimation package Standard Model Imaging, or SMI²⁰, employs machine learning methodology²¹ of connecting SM rotational invariants¹¹ to compartmental and ODF parameters.

Previous studies examining the impact of OSA in small samples of cognitively normal and mild cognitively impaired middle-aged individuals showed a negative association between all-brain white matter diffusion (Mean Diffusivity) and OSA severity²². This study found one of the strongest associations in the corpus callosum²². This finding opened the door to evaluating possible regional effects of hypoxia due to OSA^2,22. However, to our knowledge, even though previous studies have assessed some DTI and DKI metrics at a regional-based level^2,22–25, no other studies have examined the association between OSA severity and diffusion using DTI, DKI, and SMI metrics simultaneously in cognitively normal elderly. This may provide further information into the neurological underpinnings of OSA by adding insights regarding specific microstructural features of the intra- and extra-axonal space.

Our study aims to comprehensively evaluate the effects of OSA on white matter tracts using DTI, DKI, and SMI metrics in a cognitively normal older adult sample participating in sleep, memory, and healthy aging studies. Based on previous imaging studies on OSA^23,26–28, we hypothesize that OSA will be associated with decreased white matter integrity, which may impact cognitive functions^22,29–34, and that these changes will appear before the onset of clinically significant deficits (e.g., mild cognitive impairment).

Materials and Methods

Participants

This study, conducted according to the Helsinki Declaration, was approved by the New York University Grossman School of Medicine Institutional Review Board. Participants gave their written informed consent and were compensated for their participation. This article follows the STROBE guideline for observational studies³⁵.

A total of 150 cognitively normal older adults (aged 60 to 71 years) were selected for the study from a larger cohort of 448 participants recruited from the New York City area as part of studies on sleep and normal brain aging. The inclusion criteria for this subset required participants to have a Clinical Dementia Rating (CD³⁶) of 0 and must have completed dMRI scans. Most participants did not have a prior diagnosis of OSA before joining the study and consented to undergo in-lab nocturnal polysomnography (NPSG). The severity of OSA was determined using the AHI3A (Apnea–Hypopnea Index 3% Arousal), defined as the sum of all apneas and hypopneas with ≥ 3% desaturation, or EEG arousal, divided by the total sleep time in hours^37,38. Exclusion criteria consisted of intellectual disability, a history of medical conditions impacting brain structure or function (such as stroke), uncontrolled diabetes or hypertension, a history of traumatic brain injury, neurodegenerative diseases, current depression (measured by the Geriatric Depression Scale and clinical evaluation), normal pressure hydrocephalus, or MRI evidence of intracranial masses or infarcts. Participants completed a questionnaire to gather demographic information and assess their medical and treatment history. Sleep-related complaints and habits were evaluated during clinical history interviews, and excessive daytime sleepiness was measured using the Epworth Sleepiness Scale (ESS)³⁹.

Nocturnal polysomnography: sleep assessment

Each subject underwent a full in-laboratory NPSG performed according to AASM (American Academy of Sleep Medicine) guidelines⁴⁰ at the Mount Sinai Integrative Sleep Center using Compumedics E-series and Grael 2 systems (Melbourne, VIC, Australia). The signal acquisition included electroencephalogram (EEG) (minimum F3, F4, C3, C4, O1, O2, M1, and M2), left and right electrooculography (sampled at 256 Hz and referenced to a ground electrode), submental bipolar electromyography (EMG), bilateral tibialis anterior bipolar EMG, respiration measured by a pressure transducer and PAP (Positive Airway Pressure) device interface, breathing effort measured by rib/abdomen impedance plethysmography, single-channel ECG, and SpO2^37,38. PSGs were scored in 30-s epochs according to the standard sleep and EEG arousal^40,41criteria. Total sleep time and percent time spent in the wake, non-REM (Rapid Eye Movement) stage 1 (NREM 1), non-REM stage 2 (NREM 2), and non-REM stage 3 (NREM 3) were determined, as well as total sleep time, sleep latency, and sleep efficiency⁴⁰. Respiratory events were scored from the airflow signal using the AASM criteria, and the stage-specific (Total, REM, and NREM) apnea indices were calculated^37,38. Apneas are defined as the absence of airflow for ≥ 10 s. Hypopnea 4% is defined as a reduction in breathing amplitude by 30% or more for ≥ 10 s with a ≥ 4% decline in blood oxygen saturation, irrespective of arousal³⁸. Hypopnea 3% or arousal (3A) is defined as a reduction in breathing amplitude by 30% or more for ≥ 10 s accompanied by ≥ 3% decline in blood oxygen saturation or EEG arousal^37,38. EEG arousals are neurophysiologic events reflecting transient shifts toward wakefulness and are considered physiologic evidence of sleep fragmentation caused by breathing disturbances. Importantly, EEG arousals indicate that the hypopnea was disruptive enough to fragment sleep, even if oxygen desaturation was minimal or absent. Therefore, EEG arousals are not “equivalent” to hypopneas but rather represent one of two criteria that can define a hypopnea event^37,38. The AHI3A is the sum of all apneas and hypopneas with ≥ 3% desaturation, or EEG arousal, divided by the total sleep time in hours^37,38.

Neurocognitive assessment

Clinical Dementia Rating (CDR)³⁶ was applied by the study physician, excluding any participant with scores above 0. The Montreal Cognitive Assessment (MoCA), was also used. Subjects with scores below 26, which is the established cut-off score for Mild Cognitive Impairment (MCI)⁴², were excluded.

MRI acquisition

Subjects were scanned on a 3 Tesla integrated PET-MRI system (Siemens Biograph mMR, VB20) using a 12-channel phased array RF coil for reception and Siemens product diffusion sequence with bipolar gradients dMRI was acquired with the following parameters: b-value = 0 s/mm² images, b-value = 250 s/mm²—6 directions, b-value = 1000 s/mm²—20 diffusion gradient directions, b-value = 1500 s/mm²—20 diffusion gradient directions, b-value = 2000s/mm²—30 diffusion gradient directions, b-value = 2500 s/mm²—60 diffusion gradient directions, TE = 96 ms, TR = 7.9 s, 42 slices, resolution = 2.5 × 2.5 × 3mm³, 6/8 partial Fourier. One b-value = 0 was acquired with reverse phase-encoding direction for EPI distortion correction^43,44. The mean time between the sleep study and the PET-MR corresponded to 1 month (0–3 months).

MRI processing: dMRI metrics

dMRI preprocessing and tensor estimation were completed by raters blinded to sleep indices from the NYU Center for Biomedical Imaging (CBI). Preprocessing was performed using the DESIGNER pipeline^45,46, which uses MPPCA (Marchenko-Pastur Principal Component Analysis) adaptive-patch denoising^47–49, RPG (Removal of Partial Fourier-induced Gibbs ringing) Gibbs correction⁵⁰, EPI (susceptibility-induced distortion) distortion correction, eddy current and motion correction⁵¹, and Rician bias correction⁵². Diffusion and kurtosis tensors were estimated using weighted linear least squares⁵², and MD (mean diffusion), RD (radial diffusion), AD (axial diffusion), FA (fractional anisotropy), MK (mean kurtosis), RK (radial kurtosis), and AK (axial kurtosis) parametric maps were estimated from the tensors. These metrics were derived using DKI, which characterizes non-Gaussian diffusion properties, as opposed to the purely Gaussian part of the signal captured by DTI^53–55. SM metrics^5,11 (extra-axonal diffusivity along the axon (), extra-axonal diffusivity perpendicular to the axon (), axial diffusivity in the intra-axonal space (IAS) (Da), axonal water fraction (f) calculated by IAS/(EAS + IAS), and the anisotropy invariant of fiber ODF) allow us to be specific to cellular changes by using an appropriate biophysical model for diffusion in white matter, and were also estimated via SMI^20,56. SMI parameter estimation, we sampled training parameters from a uniform distribution to minimize the influence of prior assumptions¹⁷. The lower and upper bounds were set to [0.05, 1, 1, 0.1, 0.05] and [0.95, 3, 3, 1.2, 0.99], respectively, for the full parameter set⁵⁷. In the context of two-shell dMRI protocols, the free-water compartment is typically omitted because such acquisitions generally lack sufficient information to estimate all three compartments reliably²⁰ DESIGNER preprocessing and SMI estimation were applied using codes available at https://github.com/NYU-DiffusionMRI/DESIGNER-v2.

For the Region of interest (ROI) extraction in DTI/DKI maps, outlier voxels across all DTI/DKI parametric maps were defined based on their physically possible lower and upper bounds: diffusivity D (MD, AD, RD) > 0 (upper bound of D > 3 is rarely violated, so it was not included here), fractional anisotropy 0 < FA < 1, and kurtosis⁵⁸ K (MK, AK, RK) > − 2 (upper bound was not included as max K observed within our ROIs was ~ 4). Outliers are values outside of these bounds. Additionally, outliers in each SMI map were defined as voxels with values more than two standard deviations from the mean. For quantitative measures, to exclude boundary voxels affected by the partial volume effect, while also avoiding any changes we would see in the subject space, JHU white matter atlas ROIs⁵⁹ were each clipped in atlas space to exclude 5% of the lowest FA. They were then warped to each participant’s FA map by first applying a linear registration using FSL’s FLIRT (FMRIB’s Linear Image Registration Tool)^43,60, which was used to initialize the nonlinear warp using FNIRT^61,62.

Based on the literature of white matter tracts previously linked to OSA, we selected the following ROIs for analysis: Corpus Callosum (Genu (GCC) and Body (BCC))^4,23,63,64, cingulum^23,65, fornix^23,65, Anterior Limb of the Internal Capsule¹², and External Capsule^38–40.

Statistical analysis

The package MatchIt was used to create an age- and sex-matched dataset to compare OSA vs. non-OSA groups. Propensity score matching was performed using the MatchIt package (MatchIt: Nonparametric Preprocessing for Parametric Causal Inference.” Journal of Statistical Software, 42(8), 1–28)⁴¹ in R (version 2025.05.0 + 496 (2025.05.0 + 496). We used optimal matching (method = “optimal”) on sex and age to ensure perfect covariate balance between OSA severity groups. Exact matching creates subclasses of participants with identical values on matching variables, thereby eliminating confounding from these covariates. After matching, 58 patients with severe OSA (AHI ≥ 30) were matched to 58 controls (AHI < 30), achieving balance (standardized mean difference = 0) for all covariates. Descriptive variables were summarized as the number of participants (n [%]), mean, median, or standard deviation (SD). The groups, OSA vs no OSA, were determined based on AHIA3%, where we consider “NO OSA” as AHI3A% < 15 and “OSA” as AHI3A% > 15. The comparison was made as the mean dMRI metric value in the No OSA—the mean dMRI metric value OSA. The severity groups were assigned AHI3A < 5 (No OSA), AHI3A 5–15 (Mild), AHI3A 15–30 (Moderate), and AHI3A > 30 (Severe).

To verify the distribution, a histogram was performed for all the dMRI variables in the primary analyses. All variables presented a normal distribution with residuals close to 0. After normality was assured, a correlation matrix was done to evaluate the potential interaction between the OSA measure (AHI3A%) and mean DTI, DKI, and SMI measures in each ROI. The ROIs were determined using JHU-atlas⁶⁶ (For more details, please refer to Section “MRI processing: dMRI metrics” MRI Processing—Diffusion metrics). Then, multiple individual linear regression models were used to assess the association between AHI3A% and the diffusion variables (dMRI metric in WM tracts) obtained from the matrix, adjusted by age, sex, hypertension, smoking history, diabetes mellitus type 2 (DM2), and ApoE4 status. The choice of variables beyond age and sex was based on previous reports highlighting their association with WMT microstructure¹. Also, to verify potential bias, we included a summary table of the statistical weights for each predictor (Supplementary Table 2). Showing that significance was achieved only in a handful of the variables analyzed.

The regenerated graph shows AHI34% adjusted by residuals based on the model using age as a potential co-founder. To study sex differences, an age-matched comparison was followed by a Pearson correlation test. Comparisons between age- and sex-matched groups were performed using ANCOVA with age as a covariate. Mann–Whitney U test, or chi-squared tests, were used to determine the differences in descriptive characteristics between the OSA vs. non-OSA groups and between OSA severity groups. FDR correction via statistical significance was defined by p < 0.05. The size effect was calculated using Cohen’s D Test. Statistical analyses were performed in the commercial software RStudio Team (2021). RStudio: http://www.rstudio.com.

Results

Cohort characteristics

150 participants were included, with a mean age of 66 ± 5.4 years (range: 61–70; 49 male, 101 female). The racial composition includes White (60%), Black or African American (37%), and Asian (3%) participants. The ethnicity distribution was 40% Hispanic (N = 60) and 60% non-Hispanic (N = 90). On average, participants had 16 years of education (range: 15–18 years), and the mean body mass index (BMI) was 29.8 (range: 25.8–34.1).

In our sample, participants were diagnosed with hypertension (43%), diabetes (11%), and thyroid disease (15%), and 35% were Apoe4 carriers (Table 1). The mean total sleep time was 6.4 h (range [5.6–7.3]), the mean AHI3A (indicating sleep apnea severity) was 16 (0- 73.8), and all participants had a CDR of 0. The mean MoCA score was 28 ± 1.32 (Table 1 and Supplementary Table 1). Supplementary Table 1 shows the groups after matching by age, sex, and race (OSA vs. No OSA).

Table 1.

Demographics.

N	150
Sex [n (%)]
Male	49 (33%)
Female	101 (67%)
Age [Median (Q25—Q75)	66 (61–70)
Race [n (%)]
White	90 (60%)
Black or African American	56 (37%)
Asian	4 (3%)
Ethnicity [n (%)]
Hispanic	60 (40%)
Non-hispanic	90 (60%)
Years of education	16 (15–18)
Hypertension [n (%)]	64 (43%)
Cardiovascular Disease [n (%)]	14 (9.3%)
Diabetes [n (%)]	16 (11%)
Thyroid disease [n (%)]	22 (15%)
Body mass index (BMI) (Range)	29.8 (25.8, 34.1)
Total sleep time (hours) [Mean (CI)]	6.4 (5.6–7.3)
Epworth sleepiness score [Mean (CI)]	5 (3–7)
AHI3A [Mean (Range)]	16. (0–73.8)
AHI3A [Median (Q25—Q75)	13.3 (5.2–19.9)
APOe4 carriers [n (%)]	52 (35%)
MoCA score [Mean (SD)]	28 (1.32)
CDR [Median]	0 (0)

Sleep parameters between groups

Individuals with AHI3A greater than 30 per hour had shorter total sleep time than those with AHI3A less than 5 per hour (p = 0.04) (Supplementary Table 4). No differences between groups were observed in sleep efficiency, latency, or wake-after-sleep onset (WASO). Only NREM1% (p < 0.001), NREM2% (p = 0.001), and NREM3% (p = 0.04) showed differences between groups, with shorter NREM1% in the no OSA and mild groups compared to the severe group and higher percentages of NREM3 in the no OSA group (Supplementary Table 4).

dMRI and AHI3A associations

A correlation matrix of the unadjusted model (Model 1) was generated to evaluate the association between AHI3A and the diffusion tensor metrics (FA, MD, AD, RD, AK, MK, RK), and SMI (, , f, , ) metrics (Fig. 1—Supplementary Tables 5–6).

Fig. 1.

White matter ROIs unadjusted correlation with AHI3A%. Color-matched by the level of statistical significance based on p-value. FA: Fractional Anisotropy, AD: Axial Diffusivity, : Extra-axonal diffusivity along axon, RD: Radial Diffusivity, RK: Radial Kurtosis, LC: Left Cingulum, RC: Right Cingulum, GCC: Genu Corpus Callosum, EC: External Capsule, f: Axonal Water Fraction p2: Fiber orientation anisotropy.

The right cingulum demonstrated negative associations with FA (r = − 0.263, p < 0.01) and RK (r = − 0.240, p < 0.01), alongside positive correlations with RD (r = 0.253, p < 0.01). Similar but weaker patterns were observed in the left cingulum (FA: r = − 0.182, p < 0.05; RD: r = − 0.196, p < 0.05; RK: r = − 0.196, p < 0.05). Both hemispheres showed significant negative correlations with f (left: r = − 0.234, p < 0.01; right: r = − 0.235, p < 0.01) and p² (left: r = − 0.205, p < 0.05; right: r = − 0.267, p < 0.01).

In the GCC, we found negative correlations with FA (r = − 0.257, p < 0.01), AD (r = − 0.168, p < 0.05), f (r = − 0.166, p < 0.05), and p² (r = − 0.182, p < 0.05), while positive correlations were observed with RD (r = 0.216, p < 0.05) and AK (r = 0.167, p < 0.05).

The external capsule exhibited hemispheric differences, with the right showing stronger effects. There are negative correlations with FA (r = − 0.221, p < 0.01) and RK (r = − 0.258, p < 0.01) in the right external capsule. In contrast, the left showed weaker but significant correlations with FA (r = − 0.201, p < 0.05). The left anterior limb of the internal capsule showed correlations with FA (r = − 0.177, p < 0.05) and RK (r = − 0.228, p < 0.01). The uncinate fasciculus showed limited significant correlations, the most notable being RK in the right (r = − 0.220, p < 0.01).

In Table 2, we can see the linear regression model, adjusted for age, sex, diabetes mellitus type II (DM2), smoking history, hypertension, and ApoE4 status, showed a negative association between AHI3A and the GCC FA (p = 0.003) (Fig. 2A) and a positive association with RD (p = 0.01) (Fig. 2B), RK and f also demonstrated negative associations (Fig. 2C, p-value < 0.004 and Fig. 2D, p-value < 0.04, respectively). The cingulum demonstrated bilateral effects, with the left showing negative associations with FA (p < 0.001) (Fig. 2E) and positive associations with RD (p = 0.003) (Fig. 2F). The right cingulum exhibited similar patterns with FA (negative association, p = 0.002) (Fig. 2G) and RD (positive association, p = 0.009) (Fig. 2H). The right uncinate fasciculus displayed negative associations with both axial diffusivity (p = 0.006) (Fig. 2I) and p² (p = 0.003) (Fig. 2). Significant negative associations were also found in the left anterior limb of the internal capsule AD (p = 0.002) and RK (p = 0.003); in the left cingulum AD (p = 0.03), RK (p = 0.009), f (p = 0.009), and p² (p = 0.001); right cingulum RK (p = 0.009); and in the right uncinate fasciculus RK (p = 0.007) and (p = 0.03). The strongest associations are shown in Fig. 1, while these and other linear regression results are available in Supplementary Table 7.

Table 2.

Multivariable linear regression models adjusted for age, sex, hypertension, diabetes, smoking history, and APOe4 carrier status.

Region	Metric	β	SE	R2	p-value	Significance
Left Anterior limb of internal capsule	Axial Diffusion (AD)	− 0.001	0	0.2	0.002	**
	Radial Kurtosis (RK)	− 0.001	0.001	0.1	0.033	*
		− 0.002	0.001	0.1	0.003	**
	f	0	0	0.2	0.024	*
Left Cingulum	Axial Diffusion (AD)	− 0.001	0	0.1	0.029	*
	Radial Diffusion (RD)	0.001	0	0.2	0.003	**
	Fractional Anisotropy (FA)	− 0.001	0	0.3	< 0.001	***
	Radial Kurtosis (RK)	− 0.002	0.001	0.2	0.009	**
	f	− 0.001	0	0.2	0.009	**
		− 0.001	0	0.2	0.001	**
Right Cingulum	Radial Diffusion (RD)	0.001	0	0.3	0.009	**
	Fractional Anisotropy (FA)	− 0.001	0	0.2	0.002	**
	Radial Kurtosis (RK)	− 0.002	0.001	0.2	0.009	**
	f	0	0	0.2	0.022	*
		− 0.001	0	0.2	0.003	**
Left External Capsule	Fractional Anisotropy (FA)	0	0	0.2	0.015	*
Right External Capsule	Fractional Anisotropy (FA)	0	0	0.2	0.019	*
Genu of corpus callosum	Radial Kurtosis (RK)	− 0.001	0	0.1	0.024	*
	Axial Diffusion (AD)	− 0.001	0	0.1	0.038	*
	Radial Diffusion (RD)	0.001	0	0.2	0.013	*
	Fractional Anisotropy (FA)	− 0.001	0	0.2	0.003	**
	Axial Kurtosis (AK)	0.001	0	0.1	0.018	*
	Radial Kurtosis (RK)	− 0.002	0.001	0.1	0.04	*
	f	0	0	0.1	0.04	*
		− 0.001	0	0.1	0.008	**
Left Uncinate fasciculus		0.001	0.001	0.1	0.044	*
Right Uncinate fasciculus	Axial Diffusion (AD)	− 0.001	0	0.2	0.006	**
	Fractional Anisotropy (FA)	− 0.001	0	0.1	0.025	*
	Radial Kurtosis (RK)	− 0.002	0.001	0.1	0.007	**
		− 0.001	0.001	0.1	0.041	*
	f	− 0.001	0.01	0.1	0.03	*

Fig. 2.

Scatterplot and linear regression after adjusting by age, sex, hypertension, diabetes, and Apoe4 Status. Association between dMRI parameter values in white matter tracts and AHI3A%. (A) Genu of the Corpus Callosum Fractional Anisotropy (FA), negative association, p-value 0.003 (B) Genu of the Corpus Callosum Radial Diffusivity (RD), positive association, p-value 0.01 (C) Genu of the Corpus Callosum RK, negative association, p-value < 0.04 (D) Genu of the Corpus Callosum f negative association, p-value < 0.04 (E) Left Cingulum FA, negative association, p-value, < 0.001 (F) Left Cingulum RD, positive association, p-value 0.003 (G) Right Cingulum FA, negative association, p-value 0.002 (H) Right Cingulum RD, positive association, p-value 0.009 (I) Right Uncinate Fasciculus AD, negative association, p-value 0.006 J Right Uncinate Fasciculus p2, negative association, p-value 0.003.

Sex differences

Figure 3 examines differences by sex. The correlations between AHI3A and RD in the left cingulum differed significantly by sex (Fig. 3E, Z-diff 3.7). No significant differences were found in GCC and left cingulum in , and (Fig. 3A–D), with nearly significant differences for the left cingulum , (p = 0.07, Fig. 3A) and GCC (Fig. 3D, p = 0.07).

Fig. 3.

Pearson correlation test and Z-fisher for differences between sex. (A) Left Cingulum , p-value 0.07 (B) Genu of the Corpus Callosum , p-value 0.4 (C) Left Cingulum , p-value 0.2 (D) Genu of the Corpus Callosum , p-value 0.07 (E) Left Cingulum RD, p-value < 0.01 (F) Genu of the Corpus Callosum RD, p-value 0.5.

OSA versus No OSA

In Fig. 4, we performed the ANCOVA analysis between subjects with OSA (n = 58) and without OSA (n = 58), matched by age and sex and adjusted for age as a covariate, indicating differences across multiple dMRI variables. The genu of the corpus callosum showed the most differences, with significant differences in FA (mean difference [MeanDiff] = 0.04, p = 0.001) (Fig. 4A), RD (MeanDiff = − 0.03, p = 0.01) (Fig. 4B), AD (MeanDiff = 0.03, p = 0.01) (Fig. 4C), RK (MeanDiff = 0.05, p = 0.03) (Fig. 4D), AK (MeanDiff = − 0.03, p = 0.01) (Fig. 4E), f (MeanDiff = 0.01, p = 0.03) (Supplementary Table 8), and p² (MeanDiff = 0.03, p = 0.0006) (Fig. 4F).

Fig. 4.

ANCOVA comparison between subjects with OSA and without OSA matched by age and sex and adjusted by age as a covariate. Genu of Corpus Callosum, Fractional Anisotropy (FA), p-value 0.001. (B) Genu of the Corpus Callosum, Radial Diffusivity (RD), p-value 0.01. (C) Genu of Corpus Callosum, Axial Diffusivity (AD), p-value 0.01 (D) Genu of Corpus Callosum, Radial Kurtosis (RK), p-value 0.03 (E) Genu of Corpus Callosum, Axial Kurtosis (AK), p-value 0.01 (F) Genu of Corpus Callosum, p2, p-value 0.0006. (G) Right Cingulum FA, p-value < 0.05 (H) Right Cingulum, p2, p-value 0.001 (I) Fornix , p-value 0.05. (J) Left Cingulum Da, p-value 0.03 (K) Left Cingulum f, p-value 0.05.

The cingulum demonstrated bilateral effects, with the right cingulum showing differences in FA (MeanDiff = 0.01, p = 0.04) (Fig. 4G), f (MeanDiff = 0.01, p = 0.05) (Supplementary Table 8), and p² (MeanDiff = 0.02, p = 0.001) (Fig. 4H). The left cingulum exhibited differences in Da (MeanDiff = 0.02, p = 0.03) (Fig. 4J) and f (MeanDiff = 0.01, p = 0.05) (Fig. 4K). Additionally, significant differences were observed in the fornix (MeanDiff = 0.32, p = 0.05) (Fig. 4I) and p2 (MeanDiff = 0.1, p = 0.05) (Supplementary Table 8).

Differences between OSA severity groups and dMRI metrics

ANCOVA with age as a covariate and pairwise comparison analysis between subjects in different severity groups showed significant differences between AHI3A (0–5) compared to other groups (AHI3A 5–15, AHI3A 15–30, and AHI3A > 30), with the strongest differences between AHI3A (0–5) versus AHI3A (> 30) (Supplementary Table 9). In Table 3, and Fig. 5, we summarized the main differences between AHI3A (0–5) and AHI3A (> 30), which were seen in the genu of the corpus callosum FA (p = 0.001) (Fig. 5A), RD (p = 0.008) (Fig. 5B), RK (p = 0.02) (Fig. 5C), and p² (p = 0.02) (Fig. 5D).

Table 3.

ANCOVA comparison between subjects with OSA and without OSA, matched by age and sex, and adjusted by age as a covariate*.

Region	Metric	Comparison	Estimate	SE	T ratio	P-value	Significance
Genu of corpus callosum	Radial Diffusion (RD)	No OSA—Mild	− 0.045	0.015	-2.974	0.018	*
		No OSA—Moderate	− 0.046	0.016	-2.868	0.024	*
		No OSA—Severe	− 0.059	0.018	-3.237	0.008	**
	Fractional Anisotropy (FA)	No OSA—Mild	0.04	0.013	3.174	0.01	*
		No OSA—Moderate	0.045	0.013	3.345	0.006	**
		No OSA—Severe	0.057	0.015	3.761	0.001	**
	Radial Kurtosis (RK)	No OSA—Mild	0.086	0.029	2.995	0.017	*
		No OSA—Moderate	0.083	0.031	2.709	0.038	*
		No OSA—Severe	0.1	0.035	2.897	0.023	*
		No OSA—Moderate	0.035	0.012	3.022	0.016	*
	p2	No OSA—Severe	0.039	0.013	2.909	0.022	*
Left Anterior limb of internal capsule	Radial Kurtosis (RK)	Moderate—Severe	0.081	0.028	2.928	0.021	*
	f	Moderate—Severe	0.023	0.009	2.72	0.037	*
Left Cingulum	Radial Kurtosis (RK)	No OSA—Severe	0.092	0.035	2.623	0.047	*
	f	No OSA—Severe	0.033	0.01	3.369	0.005	**
Right Cingulum	Radial Diffusion (RD)	No OSA—Severe	− 0.031	0.011	-2.689	0.04	*
	Fractional Anisotropy (FA)	No OSA—Severe	0.035	0.011	3.147	0.011	*
	Radial Kurtosis (RK)	No OSA—Severe	0.093	0.03	3.129	0.011	*
	f	No OSA—Severe	0.03	0.009	3.235	0.008	**
		Mild—Severe	0.026	0.009	3.033	0.015	*
	p2	No OSA—Severe	0.038	0.011	3.406	0.005	**
		Mild—Severe	0.029	0.01	2.809	0.029	*
Right External capsule	Fractional Anisotropy (FA)	No OSA—Severe	0.019	0.007	2.699	0.039	*
	Radial Kurtosis (RK)	No OSA—Severe	0.054	0.017	3.097	0.013	*
		Mild—Severe	0.054	0.016	3.325	0.006	**
		Moderate—Severe	0.069	0.017	4.059	0.001	***
	f	Mild—Severe	0.015	0.006	2.735	0.035	*
		Moderate—Severe	0.023	0.006	3.846	0.001	**

*Only statistically significant results are shown. More results are available in the Supplementary Data.

Fig. 5.

Pairwise comparison between OSA severity groups adjusted by age. (A) Genu of Corpus Callosum, Fractional Anisotropy (FA), p-value 0.001. (B) Genu of Corpus Callosum, Radial Diffusivity RD, p-value 0.008. (C) Genu of Corpus Callosum, Radial Kurtosis RK, p-value 0.02. (D) Genu of Corpus Callosum, p2, p-value 0.02, Left Cingulum (FA), p-value < 0.05. (E) Right Cingulum FA, p-value 0.01 (F) Right Cingulum f, p-value 0.008 (G) Left Cingulum f, p-value 0.005.

The cingulum demonstrated bilateral alterations, with the right displaying differences in FA (p = 0.01) (Fig. 5E), RK (p = 0.01) (Supplementary Table 9), p² (p = 0.005) (Supplementary Table 9), and f (p = 0.008) (Fig. 5F). At the same time, the left cingulum showed differences in p² (p = 0.05) (Supplementary Table 9) and f (p = 0.005) (Fig. 5G). The right external capsule also showed differences in FA between No OSA and Severe OSA (p = 0.04), while RK (p = 0.006) and f (p = 0.001) showed significant differences between mild OSA and Severe OSA (Supplementary Table 9). To complement the quantitative data, a representative map for dMRI metrics for both groups (no OSA and OSA) was included (Figs. 6 and 7).

Fig. 6.

Distribution map for diffusion metrics FA, MD, AD, RD, MK, AK, and RK.

Fig. 7.

Distribution map for SMI metrics , , , p², and f.

Discussion

This study evaluated the associations between OSA severity (based on AHI3A) and white matter microstructure in cognitively unimpaired older adults using DTI, DKI, and SMI metrics. The strongest associations between AHI3A and dMRI parameters were found in the genu of the corpus callosum, which exhibited negative associations with FA, AD, f, and p₂ and positive associations with RD and AK. Similar differences were found in the cingulum and external capsule, with effects predominantly in the right hemisphere. Notably, some of these associations showed sex-specific differences, particularly in the left cingulum, with women showing higher RD and values.

When comparing OSA (AHI3A > 15/h) versus No OSA (AHI3A < 15/h) groups, subjects without OSA showed higher values in FA, AD, RK, f, p₂, Da, and , while demonstrating lower values in RD and AK. In pairwise comparisons between OSA severity groups, the most significant differences were found between AHI3A (0–5/h) and AHI3A (> 30/h) groups in the GCC, cingulum, and external capsule. All white matter tracts where we identified differences between the groups have been associated with memory, cognitive^67–69, and executive⁶⁷ functions.

While DTI and DKI metrics are sensitive to OSA severity, the observed changes in SMI metrics may provide specificity to the underlying cellular processes driving brain microstructure changes in OSA: Indeed, a decrease in f suggests chronic changes involving demyelination and axonal loss^14,70, as also supported by the increase in RD and decrease in RK. On the other hand, the decrease in Da suggests (sub)acute pathology of beading¹⁵, which is also supported by the increase in AK, and the increase in p₂ suggests a more isotropic environment potentially caused by increased cellularity caused by inflammatory processes and microglial activation^15,70.

OSA is characterized by intermittent episodes of hypoxia and reoxygenation, which generate a mismatch in oxygen supply and demand to brain tissue, resulting in a pro-inflammatory state²⁶. As a result, heterogeneous findings are expected in the dMRI, particularly in acute versus chronic conditions, in which there is an increase in edema followed by apoptosis over time^26,27,31,71. Our findings on OSA severity and DTI parameters are similar to those described by Chen et al.²⁶, who evaluated this process using a voxel-based approach. They showed that, in individuals with severe OSA, FA values were negatively correlated with early apoptosis in leukocytes—an indirect biomarker for systemic inflammation²⁶. Suggesting a possible link between inflammation response in OSA and white matter changes; however, this association focuses on systemic inflammatory biomarkers, not specific brain biomarkers.⁷².

While considering inflammation as a mechanistic explanation, an essential element is chronicity⁶⁵. Kumar et al.⁶⁵ showed that AD and RD were globally reduced in newly diagnosed sleep apnea compared to No OSA. Such differences might be explained by acute myelin swelling from newly diagnosed OSA slowing the water motion in both parallel and perpendicular directions; consequently, the acute hypoxic exposure (due to OSA) can lead to myelin and axonal swelling and decreased extracellular/extra-axonal space, showing an initial decrease in the DTI values^64,65. Thus, disease duration and the physiological effects of hypoxia in the myelin⁶⁵ could influence dMRI variables differently.

Myelin tends to be more susceptible to hypoxic injury than neural tissue (e.g., axons). Among all dMRI metrics, FA is a commonly utilized diffusion index that quantifies the degree of diffusion anisotropy in biological tissues⁷³. It measures the normalized variance among the diffusion tensor’s three eigenvalues⁷⁴. While the mean diffusivity (MD) reflects the average movement of water molecules, FA offers complementary insights regarding the directionality and coherence of the diffusion process⁷⁴. Higher FA values signify that one or more eigenvalues deviate from the mean, indicating strong orientation-dependent diffusivity⁷³. Decreases in FA have been noted in pathological conditions associated with demyelination and neurodegeneration⁷³, rendering FA an important metric for assessing white matter integrity; however, it is not advisable to refer to FA solely as a general index of white matter integrity⁷³.

In subacute to chronic injuries, for example, several reports suggest that demyelination processes and axonal loss lead to reduced FA, accompanied by increased RD and MD with stabilization of MD in more chronic stages^65,75. AD has also been reported as a sensitive measure of myelin integrity; however, this metric becomes less informative in more chronic stages due to debris clearance and edema reduction^22,65,75. Recently, Ning et al. evaluated patients with OSA with cognitive impairment through neurite orientation dispersion and density imaging (NODDI)^13,76. In their findings, FA, neurite density index (NDI) “—Analogous to f in SMI-, and the orientation dispersion index (ODI)—Analogous to p² in SMI” were lower in patients with OSA than in patients without OSA. The authors suggest that the reduction in the NDI represents a decrease in nerve fiber density in the white matter.

In contrast, the decrease in the ODI suggests damage to the myelin⁷⁶. In our results, most differences showed lower FA, AD, RK, f, p², Da, and higher RD and AK in subjects with OSA, with little or no difference in MD. These elements suggest, at least in our cohort, that most of the differences seen could be linked to a subacute setting (i.e., low or mild gliosis), possibly moving to a chronic inflammatory stage in some individuals. A different trend compared to other neurodegenerative diseases, such as depression or schizophrenia, was that diffusion tended to increase, opening the discussion to explore further mechanisms. However, establishing OSA chronicity in asymptomatic patients with OSA is one of the biggest challenges in several trials²⁸. In either case, the observations from this study suggest an increase in the complexity along the axon and the myelin^23,65. These elements also include a decrease in the fiber density (as f suggest), potentially with an inflammatory component⁷⁶.

In aging, some diffusion studies report accelerated age-related changes in men compared to women; however, these differences were only apparent with the multishell CHARMED procedure that separates intra- and extra-axonal diffusion components⁷⁷. In our evaluation of the differences between men and women, we found variations in the association between AHI3A and RD in the left cingulum and the GCC (Fig. 3), along with nearly significant differences in the left cingulum , and GCC (Fig. 3). Both RD and diffusion variables are linked to the restriction of water perpendicular to the axon and indicate variations in the extra-axonal component. However, while RD in women starts lower, the rate of decline is steeper than in men. Compared to men, women tend to develop OSA later in life^{22,65,77–80} but may seek medical attention sooner due to daytime sleepiness^60–63 and headaches⁸¹, among other issues^{22,65,77–80}. This pattern suggests that OSA in older women may present in a more subacute stage compared to older men, who often experience a more chronic disease trajectory. Other notable sex differences include the greater OSA severity in men, a higher frequency of REM-predominant OSA in women, and a distinct age-of-onset pattern^60–63, which relates to the relative chronicity of OSA in men compared to women. However, to clarify this, a solution would be to longitudinally evaluate a cohort with recently diagnosed sleep apnea, monitor progression, and test the long-term effects of OSA treatment with imaging.

Regarding specific ROIs, the corpus callosum shows the most differences between groups in our results. Carvalho et al., in their study of 103 participants comparing severe OSA versus moderate and mild OSA (matched by age, sex, and N3%), found that subjects with severe OSA presented lower FA values in the genu of corpus callosum (GCC) (median [IQR] 0.57 [0.55–0.63]) compared to those with mild OSA (0.63 [0.58–0.65], p = 0.007)⁶⁴. Our pairwise comparison yielded similar results (Fig. 4A—Supplementary Table 9), showing lower FA values in the GCC for the severe group compared to mild, moderate, and no OSA groups.

Our cohort found the most noticeable differences in the dMRI metrics while comparing No OSA with severe OSA in the genu of the corpus callosum, the cingulum—a prominent tract connecting frontal, parietal, and medial temporal sites while linking subcortical nuclei to the cingulate gyrus⁶⁷, and the external capsule, a connection of the cholinergic pathway previously associated with cognitive decline in Alzheimer’s Disease^2,82,83. This association between cognitive impairment, OSA, and the risk for Alzheimer’s disease has been suggested by our group and others^{1,24,32,33,37,84–91}. An example is a cross-sectional study conducted by Koo et al. that showed that subjects with untreated OSA presented lower cognitive performance and tract-specific alterations, indicating impaired WM integrity compared to healthy subjects³¹. In our sample, we did not evaluate or control for the potential confounding effect of Alzheimer’s disease pathology measured with PET or biofluid biomarkers, and we only included cognitively unimpaired individuals.

Our study’s main strength is its use of data from a well-characterized cross-sectional sample to comprehensively examine the relationships between OSA severity and multi-shell dMRI data, which is comprehensively analyzed using DTI, DKI, and SMI. We also used gold-standard diagnostic and data analysis measurements established in the field.

Some main limitations are intrinsic to the cross-sectional design, which prevents any inference about causality or temporal relationships. Second, age and sex matching, although necessary to reduce confounding, resulted in a reduction of sample size and, consequently, statistical power. This effect was amplified by the underlying imbalance in the original dataset, in which the number of female participants substantially exceeded that of males. Future studies with larger and more balanced sex distributions—particularly with more male participants—are needed to assess potential sex-specific effects better. Additionally, some of the regions of interest, such as the fornix, are small and therefore susceptible to partial volume effects. The relatively coarse spatial resolution of the DTI acquisition (2.5 × 2.5 × 3 mm³) may further limit the accuracy of diffusion estimates in these structures. Another limitation is establishing the chronicity of OSA. OSA is a prevalent condition that is significantly underdiagnosed^92,93. This can be due to the participants living alone (without a proxy that can report snoring or gasping) and limited awareness of the condition or health social determinants^92,93. Thus, it is difficult for a cross-sectional study to establish the time from the onset of the illness to the time of the evaluation. In addition, our cohort is mainly composed of highly educated subjects, in whom cognitive reserve may mask early cognitive decline. Also, compartment fractions are weighted by their T2 values, but our dMRI data was acquired at a finite TE = 96 ms. This finite TE limitation makes it difficult to estimate free water fraction⁷⁰, so the free water compartment was not included in our standard model. Thus, including acquisitions at a range of TE can help estimate unweighted fractions and thus increase specificity.

Conclusions

This study compared cognitively normal subjects with varying degrees of OSA severity based on AHI3A and examined differences in white matter microstructure using dMRI metrics. The most significant differences among groups were observed in the genu of the corpus callosum, cingulum, and external capsule, where dMRI variables such as AD and FA decreased, which may indicate axonal beading and lower white matter integrity, respectively, as OSA severity increased. In contrast, RD was increased, suggesting both demyelination and axonal loss in the OSA group. Additionally, f, a direct measurement for axonal density, was lower in the OSA group. These findings, consistent with previous reports^{4,22,23,26,27,64,65,76}, suggest that OSA severity may impact white matter tract microstructure in regions commonly involved in memory and executive functions. Future studies should include longitudinal evaluations to assess the effects of disease duration, cognitive decline, possible links with other biomarkers, OSA treatments, and the clinical significance of changes in dMRI metrics over time.

Author contributions

L.F.F., J.C., and N.L.G. performed diffusion MRI analyses. D.S.N., and E.F. performed the diffusion MRI analyses and developed the imaging pipelines that were used. X.S., T.J., G.S.-A., and S.P. contributed to data acquisition, processing and quality control. M.G., J.R.-C., J.G., I.A., K.K., A.E.M., D.M.R., and A.P. assisted with clinical characterization, sleep data acquisition and analyses. R.S.O. and L.F.F. conceived and supervised the study. L.F.F. and R.S.O. drafted the manuscript. I.R., S.L.N., S.G.B., O.M.B., D.M.R., I.A., A.V., and E.B. provided critical input on study design, interpretation, and manuscript revision. All authors reviewed and approved the final version.

Funding

The NIH supported research reported in this manuscript under R01AG056031, R01AG056531, R01AG080609, the REC Scholar Program P30 AG066512, R21AG087904, R01NS088040, and P41EB017183. O. M. Bubu is supported by the NIH/NIA (K23AG068534, R01AG082278, RF1AG083975, P30AG059303 [Pilot], P30AG066512 [Pilot], L30-AG064670) and grants from the Alzheimer’s Association (AARG-D-21-848397), BrightFocus Foundation (ADR-A2022033S), and American Academy of Sleep Medicine Foundation (BS-231-20). The funders had no role in the conception or preparation of this manuscript.

Data availability

The data that support the findings of this study are not openly available due to reasons of sensitivity and are available from the corresponding author upon reasonable request. Data are located in controlled-access data storage at New York University Grossman School of Medicine.

Declarations

Competing interests

The authors declare no competing interests.