ONH-based parameters included optic disc analysis and peripapillary retinal nerve fiber layer (pRNFL) thickness. Optic disc analysis included optic disc area/volume, cup area/volume, rim area/volume, and cup-to-disc (area, horizontal, and vertical) ratios. The pRNFL thickness was measured in hemisphere S (S-Hemi) and hemisphere I (I-Hemi) in Figure 3A and the average pRNFL thickness (González-García et al., 2009). Additionally, other pRNFL regions included quadrants-based Superior (S), Inferior (I), Temporal (T), Nasal (N), in Figure 3B, as well as supertemporal (ST), superonasal (SN), inferotemporal (IT), inferonasal (IN), nasal upper (NU), nasal lower (NL), temporal upper (TU), and temporal lower (TL) in Figure 3C.

GCC-based parameters evaluated retinal thickness (μm) in Inner_Retina [ILM to 10 μm below OPL (Hanumunthadu et al., 2021)], Full_Retina [ILM to RPE/BRM complex (Hanumunthadu et al., 2021)], and Outer_Retina [10 μm below OPL to RPE/BRM complex (Venkatesh et al., 2019; Ye et al., 2020)]. GCC-parameters, following Figure 4, included average, S, I, and S-I thicknesses for Inner_Retina, Full_Retina, and Outer_Retina layers definitions (Hanumunthadu et al., 2021; Venkatesh et al., 2019; Ye et al., 2020; Rao et al., 2010). RTVue software also provided global loss volume (GLV), focal loss volume (FLV), and root mean square (RMS) (Rao et al., 2010).

The retina map measured retinal OCT thickness (μm) and volume (mm³) around the fovea using 1 mm, 3 mm, and 5 mm circles. In Figure 5, foveal thickness/volume was denoted by Fovea, while quadrants of (S₃, I₃, T₃, N₃) and (S₅, I₅, T₅, N₅) corresponded to parafovea and perifovea, respectively. Also, in Figure 5, hemispheres S_Hemi₃ and I_Hemi₃ represented 3 mm rings (Para), while S_Hemi₅ and I_Hemi₅ indicated 5 mm rings (Peri). Additionally, parafovea was the combined S_Hemi₃ and I_Hemi₃, while perifovea was the combined S_Hemi₅ and I_Hemi₅. The retina map (μm/mm³) was calculated in Inner_Retina and Full_Retina layers, similarly to GCC parameters, as well as in RPE thickness (Elevation).

Retina3DFlowDensity investigated retinal VD changes using 3 mm quadrants and hemispheres (S and I) partitioning. The analysis also included Whole_Image, differing from Early Treatment of Diabetic Retinopathy Study (ETDRS) by including the fovea region (Savastano et al., 2021). Whole_Image was further split into S_Hemi₃ and I_Hemi₃. Additionally, VD changes in the 3 × 3 grid G and foveal density (FD_300) were also examined as part of Retina3DFlowDensity parameters, as illustrated in Figures 6A–D. FD_300 zone was used to compute FD_300_Area_Density and FD_300_Length_Density. Other Retina3DFlowDensity parameters include Fovea, FAZ_Area, FAZ_Perim (Mo et al., 2016), and AcirIndx (Tam et al., 2011). Importantly, VD was computed for various retinal layers (SVC, DVC, IVC) defined in Figure 2.

Macula_3mm parameters evaluated retinal thickness/volume changes in various quadrants and layers (ILM to IPL, ILM to RPE, ILM to BRM, and RPE to BRM). Following Figures 6A,B, thickness/volume analyses for distinct layers definitions included Center_1 (1 mm foveal ring), quadrants (S, I, T, N) in the 3 mm ring excluding 1 mm (1minus3), S and I hemispheres in 1minus3, and combined hemispheres (All_1minus3). Additionally, following Figure 4, thickness/volume analyses for distinct layers definitions involved S and I hemispheres field, where “field” indicated the inclusiveness of Center_1 region.

The R mice package (multivariate imputation by chained equations) was used to deal with missing values (van Buuren and Groothuis-Oudshoorn, 2011). A maximum threshold value of 50 percent was adopted as the limit for data eligibility for multiple imputations (Mishra and Khare, 2014; Mera-Gaona et al., 2021), with the number of iterations and multiple imputations set at 5. Interestingly, the number of missing values was ≤16% for certain parameters. The multiple imputation method was set as “default” using predictive mean matching for numeric data, logistic regression imputation for binary data with a factor of 2 levels, polytomous regression imputation for unordered categorical data (factor >2 levels), and a proportional odds model for (ordered, factor >2 levels) (Mishra and Khare, 2014). Once the En-Face Angio scans were selected for all classes, then these Angio scans were used to select the corresponding OCT/OCTA extracted parameters’ values. Basically, each Angio scan involves the patient’s ID and visit date/time, which were used to extract machine parameters associated with such conditions.

The Shapiro–Wilk test was used for normality (Shapiro and Wilk, 1965). Non-normal parameters were tested with Kruskal–Wallis (Kruskal, 1952) to find significance among AD, MCI, and HCs. Normal parameters underwent Bartlett’s test for homogeneity of variance (HOV) (Bartlett, 1937). If Bartlett’s p-value >0.05, one-way ANOVA was used, assuming equal variances (Bartlett, 1937). Both Bartlett and Shapiro–Wilk ensured HOV and normality, respectively, as conjoined conditions for the ANOVA test. If Bartlett’s p-value ≤0.05, indicating unequal variances, Welch’s ANOVA was used (Delacre et al., 2019).

After finding certain parameters to be significant (p_value <0.05), further examination should be carried out to investigate differences between the groups (Allen, 2017). Post-hoc analysis was used to determine which two groups correlate most with significant parameters (Allen, 2017). The Mann–Whitney U test was used for non-normal parameters across group combinations (Dem vs. MCI, Dem vs. HCs, MCI vs. HCs) (Sainani, 2012; Field, 2017). For normally distributed parameters, Dunnett’s test was used for failed HOV (Ruxton and Beauchamp, 2008; Hoffman et al., 2008), while the Tukey-HSD test was used for satisfied HOV (Abdi and Williams, 2010). The Mann–Whitney test results were adjusted for multiple comparisons, consistent with the adjustments made for Dunnet and Tukey’s tests.

To explore correlations between parameters and age, statistical analysis was performed using the HCs group. Pearson correlation was used for parameters with normal and linear distributions (Sedgwick, 2012), while the Spearman non-parametric test was used for non-linear and non-normal parameters. Shapiro–Wilk and Kolmogorov–Smirnov were used to test for normality (Shapiro and Wilk, 1965) and linearity (Hong-Zhi and Bing, 1991), respectively. Parameters with Pearson or Spearman of p_value ≤0.05 were considered significantly correlated with age. This method was repeated for all the parameters.

Sample size calculation for 90% power was based on the effect size (Δ or Cohen’s d) from the systematic review and meta-analysis conducted by Thomson et al. (2015), which revealed a weighted mean difference (WMD) of 12.44 μm in pRNFL thickness between AD patients and HCs. We can proceed with a sample size calculation described by Equation 1:

where Z_α/2 = 1.96 for a two-tailed test at a 5% significance level (α = 0.05), Δ = 12.44 μm (effect size), σ = 15 μm (assumed standard deviation). Additionally, to calculate the sample size with 90% power instead of 80%, the Z-value for 90% power (Z_β) should be 1.28 instead of 0.84 (Whitley and Ball, 2002; Flahault et al., 2005). Substituting the previous values into Equation 1, we get n = 15.28. Therefore, approximately 16 subjects per group were required to achieve 90% power to detect a weighted mean difference (WMD) of 12.44 μm in pRNFL thickness, assuming a standard deviation of 15 μm.

CupVolume and pRNFL thickness in the SN1 sector showed statistical significance (p = 0.048 each). ONH CupVolume decreased, while pRNFL thickness increased in the Dem and MCI groups compared to HCs. Other ONH, GCC, and Macula_3mm parameters were not significant (p-value >0.05) when comparing the three groups.

The Inner_Retina layers in the retina map showed significant thinning in parafovea, S_Hemi₃, N₃, perifovea, I_Hemi₅ sections for Dem and MCI, with no significant changes in other parafovea and perifovea quadrants. Moreover, significant Inner_Retina volume reduction was observed in parafovea, S_Hemi₃, I_Hemi₃, S₃, I₃, T₃, perifovea, S_Hemi₅, I_Hemi₅, S₅, and N₅ partitions for Dem and MCI groups; however, other quadrants’ volume changes were not significant.

Only T₃ quadrant of Full_Retina showed a significant volume reduction in Dem and MCI groups compared to HCs, with negligible volume/thickness parameters in other Full_Retina quadrants. The RPE thickness (Elevation) was significantly thinner for the Dem and MCI groups only in I₃ quadrant, with negligible changes in other quadrants. Figure 7 illustrates an example of significant parameters in the inner retina, with the green-shaded areas indicating statistical significance (p < 0.05), computed following methods outlined in section 2.5.2 across the three groups.

The significant parameters included SVC VD decline in Whole_Image, Whole Image in S_Hemi₃ and I_Hemi₃ for Dem and MCI groups. Significant Retina3DFlowDensity parameters within the 3 mm ETDRS grid included SVC VD decrease for Dem and MCI patients in parafovea/inner ring across all quadrants (S₃, I₃, T₃, N₃), parafovea S_Hemi₃ and I_Hemi₃ parts, and the full SVC parafovea. Furthermore, the SVC VD of Whole 3 mm ETDRS was significantly reduced for the Dem and MCI groups. However, SVC_L1_Fovea, representing Fovea VD, was insignificant in group comparisons.

The other significant parameters of Retina3DFlowDensity following the 3 × 3 grid were SVC VD diminution of G11 throughout G33 parts for Dem and MCI patients. Notably, most significant parameters in SVC layers were also significant in DVC layers; however, the DVC VD of Fovea, parafovea S_Hemi₃, and S₃ failed to reach statistical significance. Other negligible Retina3DFlowDensity parameters in DVC following the 3 × 3 grid were G12, G13, G22, and G23. Both FD_300_Area_Density and FD_300_Length_Density were significantly lowered for Dem and MCI patients against HCs, while variations in FAZ_Area, FAZ_Perim, and AcirIndx lacked statistical significance. Figure 8 illustrates significant parameters with green shades indicating statistical significance (p < 0.05)

Despite the initial lack of significance, certain parameters may show notable significance in specific comparisons. Post-hoc analysis revealed significant GCC parameters, including average thickness reduction at Full_Retina in S for MCI vs. HCs, and thickness decrease at Inner_Retina in S-I quadrant for Dem vs. MCI. Other significant ONH parameters included an increase of rim area and optic disc volume for Dem vs. HCs.

Contrary to initial results, certain Macula_3mm parameters showed statistical significance, including volume reductions ( mm 3 ) from ILM to RPE layers at S _ Hemi 3 and N parts, both under 1 minus 3 definition, when comparing Dem vs. HCs. Significant Retina3DFlowDensity parameters emerged from post-hoc analysis, including DVC VD reduction at S _ Hemi 3 , G13, and G23 sections for Dem vs. HCs.

The post-hoc analysis also identified significant retina map parameters such as volumetric reduction in S _ Hemi 3 , S _ Hemi 5 , and T 5 parts of Full_Retina layers, and shrinkage in I 5 and T 5 quadrants of Inner_Retina for Dem vs. HCs. Moreover, other retina map parameters, including notable thickness reduction of Inner_Retina layers in I _ Hemi 3 , S 3 , I 3 , S _ Hemi 5 , and N 5 parts for Dem vs. HCs. Lastly, only N 5 section of RPE thickness (Elevation) was significantly reduced in Dem vs. MCI.

Age was heavily correlated with OCT/OCTA parameters, and hence, it was used as an offset variable to correct age effects on predictors (Field, 2017). The TW-GLM identified many parameters correlated with the MMSE score, and we shall highlight key associations only.

In terms of GCC-based parameters, a lower MMSE score was correlated with inferior thinning in Inner,Full,Outer _ Retina , and superior thinning in Inner,Full _ Retina . Other significant associations with lower MMSE included average S-I thickness in Full,Outer _ Retina and FLV, GLV, RMS, and average Inner_Retina thickness.

ONH-based parameters were correlated with MMSE decline, except for vertical cup-to-disc ratio as well as rim and disc volumes. The pRNFL thickness parameters were correlated with MMSE, except for these sections: T, S, NL, TL, ST2, NL1, NL2, IN2, IN1, IT2, and TL1.

Macula_3mm parameters correlated with MMSE decline, with a few exceptions described in Supplementary material D. VD, changes around the fovea were significantly associated with MMSE decline, except for FAZ area changes that will be discussed further.

Retina map age-corrected parameters that were correlated with MMSE decline included Inner_Retina thickness thinning ( T 3 , T 5 , I 5 ), Inner_Retina volume reductions ( N 3 , I 5 ), Full_Retina thickness decrease (ParaFovea, PeriFovea, S _ Hemi 3 , T 3 , S 3 , T 5 , N 5 , I 5 ), and Full_Retina volume reduction ( S _ Hemi 3 , I _ Hemi 3 , S 3 , I 3 , T 3 , N 3 ). These findings are interesting since the initial statistical results correlated only Full_Retina volume decrease with cognitive decline among the groups.

The OL-GLM demonstrated no correlation of GCC-related parameters against class variables, which aligned with our initial statistical analysis results.

Interestingly, our OL-GLM showed a significant correlation between class variables against many of ONH-based parameters, which is surprising since our initial analysis only found an association between class variable against CupVolume and SN1 pRNFL thickness. These inconsistencies will be discussed in section 4.3.

Parameters based on Macula_3mm were correlated with class variables. Some of these parameters following ILM to IPL layer definition (Tao et al., 2019; Zhang et al., 2019; Yang et al., 2022) include significant thickness reduction in the S hemisphere (1minus3 defined in section 2.4.5) and in the whole field (without partitions). Additional Macula_3mm parameters (ILM to IPL layers) include volumetric reduction of N and I-Hemi sections (1minus3) and the S and I hemispheres field (defined in Figure 4). On the contrary, the class variables studied by OL-GLM was associated with other Macula_3mm parameters (ILM to BRM layer definition) including a significant thickness thinning in Center _ 1 , T and I (1minus3 definition), S and I hemispheres (1minus3), S and I hemispheres field (defined in Figure 4) sections, as well as in the whole field (without partitions). Moreover, following ILM to BRM layer definitions, the class variables analyzed by OL-GLM was also correlated with a notable volumetric shrinkage in Center _ 1 , I (1minus3), S-Hemi (1minus3), S-Hemi field (defined in Figure 4), as well as in the whole field (without partitions).

In terms of parameters related to retina map with Full_Retina layers definition, the OL-GLM linked class variables with significant thickness/volume reduction in fovea , I _ Hemi 3 , I 3 , and whole perifovea sections. Moreover, the class variable was also associated with a notable thinning in whole parafovea, S _ Hemi 5 , I _ Hemi 5 , S 5 , and T 5 sections. Additionally, the OL-GLM associated class variables with a prominent volumetric decrease in S 3 , T 3 , N 3 , and N 5 sections. On the other hand, the OL-GLM correlated class variables with retina map parameters in Inner_Retina layers definition. Specifically, a significant thinning in fovea , parafovea, S _ Hemi 3 , N 3 , T 3 , and I 3 sections were linked with class variables. Moreover, class variables was associated with prominent Inner_Retina volumetric reduction observed in fovea , parafovea, S _ Hemi 3 , I _ Hemi 3 , N 3 , T 3 , perifovea, S _ Hemi 5 , I _ Hemi 5 , S 5 , I 5 , T 5 , and N 5 partitions.

Following RPE thickness (Elevation) assessment, the OL-GLM related class variables with significant thinning in I 3 section, which is in line with our initial analysis. However, OL-GLM also found an association between class variables with notable thinning in T 3 , T 5 , S 5 , and I 5 sections.

Next, the class variables by OL-GLM was linked to most parameters based on Retina3DflowDensity. This indicates the importance of VD changes in SVC/DVC layers even after adjusting for confounding factors like age, gender, and eye pathologies. It is fair to say that the OL-GLM results are more reliable than the initial statistical results due to the necessary adjustments that were applied.

Moving our attention to Retina3DflowDensity parameters, the OL-GLM correlated these parameters with class variables. Some of these correlations agreed with our initial statistical analysis; however, OL-GLM discovered more parameters. A detailed description of the discrepancies between both methods (initial statistical analysis against OL-GLM) can be found in discussion section 4.1.3.

Research conducted by Chua et al. (2020), Bulut et al. (2018), Sadda et al. (2019), and Salobrar-Garcia et al. (2020) studied VD alterations in various layers’ definitions; however, only Bulut et al. (2018) used a similar SVC layer definition to ours. Our findings demonstrated substantial VD decline of Whole_Image and parafovea in SVC layers for Dem vs. HCs, aligned with Bulut et al. (2018). However, Bulut et al. (2018) also found a significant fovea VD decrease in SVC layers for Alzheimer’s type dementia; unlike our initial and post-hoc analyses. Bulut et al. (2018) excluded patients with macular pathologies, whereas our cohort included them due to old-aged participants, with the majority of them having eye pathologies.

Although our study followed definitions of SVC/DVC layers, it was worth discussing other works in the literature that adopted slightly different definitions of retinal layers. Importantly, Chua et al. (2020) showed a significant VD loss in superficial capillary plexus (SCP) deep capillary plexus (DCP) layers for AD against HCs. Notably, SCP and DCP layers were defined by Chua et al. (2020) and Zhang et al. (2019) from ILM to IPL and INL to OPL, respectively. The other study by Zhang et al. (2019) revealed significant VD reduction in SCP layers at parafovea, matching our results. Despite different SCP layer definitions by Chua et al. (2020) and Zhang et al. (2019), the findings emphasize parafovea VD changes in superficial layers and their correlation with cognitive decline.

Our initial findings agreed with Yan et al. (2021) that significant VD reductions in SVC (parafovea, I _ Hemi 3 , T 3 , N 3 ) and in DVC ( I _ Hemi 3 , T 3 ) were associated with AD compared to HCs. Our results further highlighted a notable VD decrease in SVC layers, specifically in S _ Hemi 3 , S 3 , and I 3 . Additionally, a prominent VD decline was observed in DVC layers within parafovea, N 3 , and I 3 . On the contrary, these significant changes in SVC and DVC layers were not significant according to Yan et al. (2021). Intriguingly, both studies (our and Yan et al., 2021) agreed on minor VD changes in fovea at SVC/DVC and in S at DVC between AD and HCs. These results doubt the reliability of VD reductions around the fovea. However, VD changes in specific SVC/DVC sections were reliably correlated with AD.

Wu et al. (2020) explored VD variations in SCP (3 μm below ILM to 15 μm below IPL) and DCP (from 15–70 μm below IPL till OPL) layers, using similar Avanti SD-OCT. Our findings match Wu et al. (2020), showing VD decrease in superficial layers (only S 3 ), and in deep layers (parafovea, I 3 , T 3 , N 3 ) linked to AD compared to HCs. Wu et al. (2020) associated VD decline in S 3 of superficial layers with AD; however, we observed VD decline in superficial layers (parafovea, S _ Hemi 3 , I _ Hemi 3 , I 3 , T 3 , and N 3 ) and I _ Hemi 3 of deep layers. Wu et al. (2020) also found VD decline in S 3 of deep layers linked to AD, contradicting our results.

Our research found noteworthy VD decline in all superficial layers’ sections, except for fovea , for MCI against HCs; however, Wu et al. (2020) did not report similar significant changes. Additionally, both studies (ours and Wu et al., 2020) observed VD decline in I 3 and T 3 sections in deep layers for MCI, unlike HCs. Moreover, both studies showed some inconsistencies: (a) prominent deep layers VD decrease in S 3 was linked with MCI in Wu et al. (2020) but not observed in our results; (b) significant deep layers VD decrease in N 3 was correlated with MCI in our analysis but not significantly in Wu’s et al. (2020) study. These inconsistencies may arise from different layer definitions.

Yang et al. (2022) demonstrated a significant VD decrease in parafovea, S 3 , I 3 , T 3 , N 3 sections in SCP/DCP layers for MCI compared with HCs, which is consistent with our findings in superficial layers only.

We found negligible VD changes in all deep layers’ sections between MCI and HCs. Interestingly, the findings of Yang et al. (2022) and our study reported considerable VD decline in FD_300_Length_Density for MCI. Our initial findings of subtle changes in FAZ_Area, FAZ_Perim, and AcirIndx for MCI vs. HCs were also consistent with Yang et al. (2022). Other inconsistencies comparing MCI against HCs, our study showed minor VD changes in FD_300_Length_Density in our study but notable by Yang et al. (2022). Notably, SCP and DCP layers were defined by Yang et al. (2022) from ILM to IPL and from IPL to OPL, respectively. Despite different SCP and DCP layers definitions by Yang et al. (2022) and our SVC and DVC layers definitions, VD changes in superficial layers were similar.

The class variable studied by OL-GLM was correlated with Retina3DflowDensity parameters. Importantly, a few of these parameters were significant in our initial statistical analysis; however, they were not significant by OL-GLM. These parameters were VD changes in SVC (parafovea, S _ Hemi 3 , I _ Hemi 3 ) and in DVC ( Whole _ ETDRS ). We believe that these parameters were heavily influenced by confounding factors.

Interestingly, the VD of SVC and DVC following the 3 × 3 grid in all sections were significantly correlated with class variables by OL-GLM, unlike our initial statistical analysis that failed to link cognitive decline with DVC VD decrease in Fovea, parafovea S _ Hemi 3 , and S 3 sections. This indicates that the OL-GLM model, adjusted for confounders, confirmed the association of SVC/DVC VD decline in many sections and captured insights that were missed by traditional statistical analysis methods.

Similarly, the class variables studied by OL-GLM were correlated with notable VD decline in SVC quadrants ( S 3 , I 3 , T 3 , N 3 ), in DVC quadrants ( S 3 , I 3 , T 3 , N 3 ), as well as a significant decrease in FD_300_Area_Density and FD_300_Length_Density. The same parameters were correlated with MCI by Yang et al. (2022). On the contrary, another research by Yan et al. (2021) confirmed VD decline for the AD group against HCs only in T 3 and N 3 sections of SVC and in T 3 section of DVC. This indicates that the reliability of SVC/DVC VD in some sections is questionable for the AD group and more consistent for the MCI group.

The TW-GLM analysis, adjusted for age, revealed a strong correlation between FAZ area changes in IVC layers and MMSE scores. The age-related analysis failed to link old age to FAZ changes; however, this analysis was applied to HCs. Based on TW-GLM, FAZ changes might help detect cognitive decline, but more extensive future studies are still needed. The findings by Zhang et al. (2019), Peng et al. (2021), Robbins et al. (2022), Biscetti et al. (2021), and Yang et al. (2022) support our conclusion of minor FAZ area changes related to cognitive decline in initial or post-hoc analyses.

The age-related analysis linked some VD-based parameters with old age. This might explain inconsistencies between our cohort (Dem, MCI, HCs aged 82.20 ± 6.22, 80.70 ± 5.90, and 79.07 ± 5.90) and the cohort by Wu et al. (2020) (AD, MCI, HCs aged 69.94 ± 6.39, 67.81 ± 5.96, and 68.67 ± 5.85) and younger groups by Yang et al. (2022) (MCI, HCs aged 58.3 ± 8.3, and 51.0 ± 7.8). VD changes in deep retinal layers could be affected by age, impacting cognitive decline detectability. The TW-GLM, adjusted for age, found significant correlations between VD parameters and MMSE scores, confirming VD’s importance in detecting cognitive decline. Our findings highlighted the importance of significant VD changes in all G grid sections (G11 to G33); however, they were not linked with cognitive decline in the literature, requiring further investigation.

Contradicting our statistical and post-hoc analysis, the age-adjusted TW-GLM model showed a strong correlation between thickness/volume changes around the fovea with MMSE scores. Age-related analysis revealed significant correlations between many Macula_3mm parameters (thickness/volume) and old age. Hence, TW-GLM identified the potential usage of these parameters by relating them with MMSE scores.

The age-adjusted TW-GLM revealed significant correlations between many ONH-based and GCC-based parameters with MMSE scores, which is opposite to our initial statistical and post-hoc analyses. Contradicting our assumption, these predictors were not correlated with old age; however, the age analysis included only HCs without Dem/MCI groups. These results (GCC and ONH parameters) suggested that retinal structural changes (around the fovea and optic disc) might be linked to age and cognitive decline conjointly, requiring further research in larger cohorts.