Distinct ocular microvascular alterations in Alzheimer's disease and cerebral small vessel disease Authors: Ziyi Zhang, Le Cao, Peng Liu, Shiyun Yang, Yuying Yan, William Robert Kwapong, Ying He, Mingyuan Yang, Hang Wang, Youjie Wang, Zijuan Feng, Jingyu Zhou, Ming Liu, Peng Lei, Bo Wu, Shuting Zhang Published: 2026 ABSTRACT AbstractINTRODUCTIONWhile both Alzheimer's disease (AD) and cerebral small vessel disease (CSVD) involve vascular dysfunction, the ocular microvascular differences between them remain largely underinvestigated.METHODSThis cross‐sectional study included 650 participants (149 AD, 276 CSVD, and 225 cognitively unimpaired individuals). Optical coherence tomography angiography (OCTA) metrics were analyzed to characterize ocular microvascular alterations.RESULTSCSVD patients exhibited a higher burden of retinal ischemic perivascular lesions (RIPLs) than AD patients (all p < 0.05), whereas AD patients showed reduced choriocapillaris (CC) density (p = 0.031). In the CSVD group, RIPLs were associated with total CSVD score (p = 0.049), while reduced CC density was associated with elevated phosphorylated tau 181 in the AD group (p = 0.001). Machine learning models integrating OCTA metrics effectively differentiated AD from CSVD (area under the curve = 0.86).DISCUSSIONAD and CSVD exhibit distinct ocular microvascular alterations, supporting OCTA‐derived measures as potential non‐invasive biomarkers. Trial registration number: ChiCTR2000041386Keywords: Alzheimer's disease, blood biomarkers, cerebral small vessel disease, choriocapillaris, retinal ischemic perivascular lesions, retinal microvasculature INTRODUCTION The most common causes of dementia are Alzheimer's disease (AD) and cerebral small vessel disease (CSVD). 1 CSVD affects the small arteries and veins in the brain, whereas AD is characterized by progressive neurodegeneration driven by amyloid beta (Aβ) plaques and tauopathy. 2 , 3 However, vascular injury—particularly microvascular dysfunction—is increasingly considered a pivotal contributor to AD pathophysiology. 4 , 5 Technical limitations in directly imaging the living human cerebral microvasculature have hindered the precise characterization of microvascular changes in these diseases. The retina, an extension of the central nervous system, provides a unique window into the brain. 6 Retinal and cerebral microvasculature share significant structural similarities, and alterations in retinal microvessels may mirror those in the cerebral microcirculation. 7 Retinal imaging studies have identified microvascular changes in both preclinical and symptomatic AD and CSVD cohorts, linking retinal vascular metrics to cognitive decline. 8 , 9 , 10 Swept‐source optical coherence tomography angiography (OCTA) is a non‐invasive tool that overcomes the retinal pigment epithelium barrier, enabling high‐resolution visualization of the retinal and choroidal microvasculature. 11 Although OCTA has documented microvascular alterations in AD and CSVD independently, no study has directly and comprehensively compared ocular microvascular injury signatures between them. Concurrently, blood‐based biomarkers, especially phosphorylated tau and Aβ, have emerged as robust tools for detecting AD pathology. 12 , 13 These biomarkers correlate with positron emission tomography (PET)‐confirmed Aβ and tau pathology, predict cognitive decline, and offer a scalable alternative to neuroimaging. 14 However, whether these core AD biomarkers correlate with in vivo microvascular measures—and whether such relationships are specific to AD or also occur in vascular disorders such as CSVD—remains unexplored. We have previously reported an association between choroidal metrics and phosphorylated tau at threonine 181 (p‐tau181) in a cohort of early‐onset AD 15 ; however, its relevance to sporadic AD and its specificity relative to CSVD are yet to be determined. In this study, we aim to (1) characterize ocular microvascular differences between AD and CSVD, (2) use machine learning models integrating OCTA metrics to differentiate AD from CSVD and cognitively unimpaired (CU) individuals, and (3) explore the associations of these ocular microvascular metrics with AD blood biomarkers and CSVD markers. METHODS Figure 1 illustrates the study design overview. This cross‐sectional study included participants derived from two concurrent prospective observational cohorts recruited from the Neurology Clinic of West China Hospital, Sichuan University, China (Figure S1 in supporting information). Individuals presenting with memory decline were enrolled in the Dementia and Aging Study, 15 whereas patients presenting with non‐acute cerebrovascular symptoms, including headache, dizziness, and memory complaints, were enrolled in the Cerebral Small Vessel Disease Study (NASCIR). 16 CU individuals were concurrently recruited from community‐dwelling older adults. Detailed inclusion criteria for each cohort have been previously described. 15 , 16 , 17 All participants underwent a comprehensive assessment protocol, including OCTA, 3T brain magnetic resonance imaging (MRI), cognitive evaluation, and blood sampling. Additionally, participants from the Dementia and Aging Study underwent 18F‐florbetapir amyloid PET (Aβ PET) to confirm Aβ pathology. Demographic and clinical characteristics were systematically collected for all participants, encompassing age, sex, education level, vascular risk factors (hypertension, diabetes mellitus, dyslipidemia, smoking status, alcohol use), and past medical history. Study design diagram. The workflow illustrates participant classification, multimodal data acquisition, and subsequent analysis. Aβ, amyloid beta; Aβ PET, 18F‐florbetapir amyloid positron emission tomography; AD, Alzheimer's disease; APOE, apolipoprotein E; CC, choriocapillaris; CSV, choroidal stroma volume; CSVD, cerebral small vessel disease; CU, cognitively unimpaired; CVV, choroidal vessel volume; DVC, deep vascular complex; DWMH, deep white matter hyperintensities; FAZ, foveal avascular zone; HIV, human immunodeficiency virus; ML, machine learning; MR, magnetic resonance; O CTA, optical coherence tomography angiography; p‐tau181, phosphorylated tau at threonine 181; PVS‐BG, perivascular spaces in the basal ganglia; PVS‐CS, perivascular spaces in the centrum semiovale; PWMH, periventricular white matter hyperintensities; RIPLs, retinal ischemic perivascular lesions; SVC, superficial vascular complex; t‐tau, total tau; TP, treponema pallidum. RESEARCH IN CONTEXT Systematic review: We searched PubMed for optical coherence tomography angiography (OCTA) studies comparing Alzheimer's disease (AD) and cerebral small vessel disease (CSVD). While microvascular alterations have been documented separately in each condition, direct comparisons to identify distinctive differences are lacking. Systematic review: We searched PubMed for optical coherence tomography angiography (OCTA) studies comparing Alzheimer's disease (AD) and cerebral small vessel disease (CSVD). While microvascular alterations have been documented separately in each condition, direct comparisons to identify distinctive differences are lacking. Interpretation: Our findings indicate that AD and CSVD exhibit divergent ocular microvascular alterations. CSVD presents with focal retinal ischemic lesions linked to total CSVD score, whereas AD shows diffuse choriocapillaris reduction associated with elevated phosphorylated tau 181. A combined machine learning model integrating these metrics effectively differentiated the two conditions. These results suggest OCTA as a potential non‐invasive tool for clinical differential diagnosis and triage. Interpretation: Our findings indicate that AD and CSVD exhibit divergent ocular microvascular alterations. CSVD presents with focal retinal ischemic lesions linked to total CSVD score, whereas AD shows diffuse choriocapillaris reduction associated with elevated phosphorylated tau 181. A combined machine learning model integrating these metrics effectively differentiated the two conditions. These results suggest OCTA as a potential non‐invasive tool for clinical differential diagnosis and triage. Future directions: Subsequent research should address: (a) validating the longitudinal associations of these metrics with amyloid positron emission tomography and quantitative white matter damage; and (b) elucidating the mechanisms underlying choroidal vulnerability in AD through basic experimental models. Future directions: Subsequent research should address: (a) validating the longitudinal associations of these metrics with amyloid positron emission tomography and quantitative white matter damage; and (b) elucidating the mechanisms underlying choroidal vulnerability in AD through basic experimental models. This study was approved by the Biomedical Research Ethics Committee on Human Research of West China Hospital, Sichuan University (2021–917), and conducted in accordance with the tenets of the Declaration of Helsinki. Written informed consent was obtained from all participants or their legal guardians. We followed the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines. Participants All patients with AD included in our study met the National Institute on Aging–Alzheimer's Association biological definition. 18 Inclusion criteria required: (1) a clinical syndrome of dementia, Aβ PET positivity, and a Montreal Cognitive Assessment (MoCA) score < 23; (2) complete clinical information; (3) normal appearance of the optic nerve head and retina upon examination; and (4) the ability to cooperate with MRI, OCTA, and blood draws. Exclusion criteria were as follows: individuals with (1) severe white matter hyperintensity (WMH) burden (Fazekas score > 2), history of stroke, cortical or subcortical infarcts, multiple lobar cerebral microbleeds (CMBs), cortical superficial siderosis, or other vascular lesions, to reduce significant vascular contribution; (2) history of other neurological or major psychiatric disorders (e.g., traumatic brain injury, brain tumor, intracranial infection, major depression, or schizophrenia); (3) evidence suggestive of alternative neurodegenerative dementia etiologies; (4) significant ocular disease (e.g., diabetic retinopathy, severe glaucoma, or other conditions/inflammatory conditions affecting the optic fundus) or intraocular surgery within the preceding 6 months; (5) uncontrolled systemic diseases (e.g., severe hypertension, systemic inflammatory diseases); (6) history of any malignancy; (7) history of alcohol or substance abuse. Although dementia with Lewy bodies (DLB), frontotemporal dementia, and other dementia subtypes were considered clinically less likely based on comprehensive neurological evaluations, coexisting pathology could not be completely ruled out. Patients with CSVD were included if they presented with subjective cognitive complaints and a MoCA score < 26, coupled with moderate‐to‐severe WMH on MRI. 19 Exclusion criteria included: (1) any cause of stroke other than CSVD, cortical infarct, subcortical infarct > 1.5 cm in diameter, multiple CMBs, or cortical superficial siderosis; (2) abnormally elevated serum p‐tau181 levels; (3) ocular diseases affecting vascular metrics; (4) history of other major neurological disorders; and (5) any history of malignancy. However, concomitant AD pathology could not be entirely excluded. Thus, this was a clinically defined, CSVD‐enriched population with possible mixed pathology. CU individuals were recruited from a community‐dwelling older adult population. The inclusion criteria were as follows: (1) age ≥ 50 years, (2) absence of subjective cognitive complaints, (3) normal objective cognitive performance (Mini‐Mental State Examination [MMSE] ≥ 26), and (4) the capacity to complete all study procedures. Exclusion criteria aligned with those of the AD group, excluding conditions that could confound cognitive or vascular assessments. MRI protocol and CSVD markers assessment The brain imaging protocol has been described in detail in our previous reports. 20 Briefly, all participants underwent standardized MRI scanning on a 3.0‐T system (Magnetom Trio, Siemens Medical Systems), including 3D T1‐weighted, T2‐weighted, fluid‐attenuated inversion recovery (FLAIR), diffusion‐weighted imaging (DWI), and susceptibility‐weighted imaging (SWI) sequences. CSVD markers, including lacunes, CMBs, periventricular white matter hyperintensities (PWMH), deep white matter hyperintensities (DWMH), and perivascular spaces (PVS), were evaluated according to the Standards for Reporting Vascular Changes on Neuroimaging‐2 (STRIVE‐2) criteria. 21 Specifically, lacunes were defined as rounded or ovoid subcortical cavities measuring 3 to 15 mm in diameter with cerebrospinal fluid (CSF)‐like signal intensity on T2‐weighted and FLAIR images. 21 CMBs were defined as homogeneous, rounded hypointense lesions measuring 2 to 10 mm on SWI. 21 Lacunes and CMBs were dichotomized as absent/present variables for statistical analyses. WMH severity was visually assessed on FLAIR images using the Fazekas scale, 22 with PWMH and DWMH rated separately on a 0 to 3 scale. PVS was defined on T2‐weighted image as small (< 3 mm) CSF‐like spaces in the basal ganglia or centrum semiovale and was rated using a validated semi‐quantitative 0 to 4 scale. 23 A total CSVD score ranging from 0 to 4 was calculated as previously described. 24 Representative examples of CSVD markers are shown in Figure 1. MRI markers were primarily evaluated by an experienced rater (C.L.) blinded to clinical and OCTA data. To assess inter‐rater reliability, a second independent rater (Z.Z.) evaluated a randomly selected subset of 30 participants. Inter‐rater agreement was assessed using Cohen kappa for binary variables (lacunes and CMBs) and weighted kappa for ordinal variables (PWMH, DWMH, and PVS). Any discrepancies between raters were resolved through consensus discussion. Inter‐rater agreement was high for lacunes (kappa = 0.85), CMBs (kappa = 0.83), PVS in the centrum semiovale (weighted kappa = 0.75), PVS in the basal ganglia (weighted kappa = 0.86), PWMH (weighted kappa = 0.89), and DWMH (weighted kappa = 0.90). AD blood biomarkers Venous blood was collected in serum vacutainer tubes without additives and transported to the laboratory within 30 minutes. Serum was separated by centrifugation at 845 × g for 15 minutes at 4°C, aliquoted, and immediately stored at –80°C until batch analysis. Concentrations of serum Aβ40, Aβ42, and total tau (t‐tau) were quantified using the Quanterix Simoa Human Neurology 3‐Plex A Advantage Kit (Cat No. 101995). Serum p‐tau181 was measured using the Quanterix Simoa p‐tau 181 Advantage V2 Kit (Cat No. 103714). All assays were run on a Quanterix SR‐X analyzer platform strictly following the manufacturer's protocols, as previously described. 15 , 25 To minimize technical variance, all samples were analyzed in a single batch. OCTA imaging Swept‐Source OCTA (VG200S; SVision Imaging; version 2.0.106) was used to capture retinal and choroid images. Details of the OCTA equipment are available in our prior reports. 15 , 26 , 27 OCTA images were acquired using a 512 B‐scan raster scanning protocol centered on the fovea (6×6 mm2 area), with each B‐scan consisting of 512 A‐scans and repeated four times at identical locations to generate the OCTA image. The built‐in OCTA software generated angiograms of the superficial vascular complex (SVC), deep vascular complex (DVC), and choriocapillaris (CC). The boundary of the SVC and DVC was defined as the inner two‐thirds and outer one‐third border of ganglion cell‐inner plexiform layer. The CC was defined as the microvasculature extending 20 µm beneath retinal pigment epithelium (RPE)–Bruch's membrane complex. Vessel density (%), calculated as the percentage of blood flow area relative to the total measured area in binarized images, quantified the microvasculature of the SVC, DVC, and CC. The foveal avascular zone (FAZ) area was measured in mm2. Choroidal thickness was defined as the vertical distance from the RPE‐Bruch's membrane complex to the choroidoscleral interface. Choroidal vascular volume (CVV) and choroidal stromal volume (CSV) represented the volumes of large‐to‐medium choroidal vessels and non‐vascular tissue respectively, spanning from beneath the RPE‐Bruch's membrane complex to the choroidoscleral junction. The choroidal vascular index (CVI) was calculated as the ratio of CVV to total choroidal volume. Retinal ischemic perivascular lesions (RIPLs) were identified on OCT B‐scans as focal contraction of the inner nuclear layer accompanied by compensatory upward expansion of the outer nuclear layer into the expected plane of the outer plexiform layer (OPL; Figure 1). 28 En face images at the level of the OPL, reconstructed from sequential B‐scans of the three‐dimensional volumetric OCT acquisition, were used to confirm each lesion and its perivascular location. Two ophthalmologists, blinded to participants’ clinical information, independently evaluated all images. Interobserver agreement was high (kappa = 0.90, p < 0.001), and discrepant or equivocal cases were resolved by consensus. RIPLs were counted separately in each eye, and their presence, quadrant distribution (defined as the number of affected quadrants), and total lesion count were recorded. RIPL burden was categorized using a four‐level ordinal scale: 0, no lesions; 1, one lesion; 2, two to three lesions; and 3, more than three lesions. Our study excluded eyes with ophthalmological disorders including severe cataracts, diabetic retinopathy, age macular degeneration, glaucoma, and vascular occlusion. Images with a signal strength < 8 were also excluded. All the OCTA data presented adhered to the OSCAR‐IB quality criteria and Advised Protocol for OCT Study Terminology and Elements recommendation. 29 , 30 Statistical analysis Continuous data were summarized as the mean ± standard deviation (SD) or median (interquartile range [IQR]) based on distribution, and categorical data as frequencies and percentages. Baseline characteristics were compared using one‐way analysis of variance (ANOVA), the Kruskal–Wallis test, chi‐squared test, or Fisher exact test, as appropriate. Generalized estimating equation (GEE) models were used to compare OCTA metrics among the AD, CSVD, and CU groups while accounting for inter‐eye correlation. Continuous OCTA parameters were analyzed using Gaussian GEE models. The presence of RIPLs was evaluated using Poisson GEE models, whereas RIPL burden was analyzed using ordinal logistic GEE models. Post hoc pairwise comparisons were subsequently performed with Bonferroni correction for multiple testing. Within‐group associations between OCTA metrics, RIPLs, blood biomarkers, and CSVD markers were assessed using linear or Poisson GEE models, as appropriate. All regression models were adjusted for age, sex, education, hypertension, and diabetes mellitus. To evaluate the diagnostic performance of OCTA parameters in differentiating CU, AD, and CSVD participants, we constructed two sets of classification models: a clinical baseline model and a combined model incorporating OCTA metrics. The clinical baseline model included demographic and vascular risk factors (age, sex, education, hypertension, and diabetes mellitus) and was developed using least absolute shrinkage and selection operator (LASSO) logistic regression. The combined model integrated both clinical variables and OCTA features and was evaluated using five machine learning algorithms, including LASSO, random forest (RF), support vector machine (SVM), multilayer perceptron (MLP), and eXtreme Gradient Boosting (XGBoost). To ensure robust evaluation and prevent data leakage, a strict hold‐out validation framework was used. Initially, the dataset was randomly partitioned into a training set (80%) and an independent internal test set (20%). Within the 80% training set, we used 5‐fold cross‐validation combined with a grid search to fine‐tune the hyperparameters for each algorithm. Specifically, for LASSO, the penalty parameter (lambda) was tuned; for RF, the number of variables randomly sampled as candidates at each split and the maximum depth of the trees were adjusted; in SVM, the cost parameter (C) and the kernel parameter (sigma) were optimized; for the MLP, the size of the hidden layers, the activation function, and the weight decay parameter were tuned; and in XGBoost, the learning rate, the number of estimators, and the maximum depth of the trees were adjusted. For all models, the combination of hyperparameters that maximized the mean cross‐validated area under the receiver operating characteristic curve (AUC) was selected. The final predictive performance of the optimized models was exclusively assessed on the 20% hold‐out test set, which remained completely unseen during the training and tuning phases. Performance metrics, including the AUC, accuracy, sensitivity, and specificity, were derived solely from this independent test set. p values < 0.05 were considered statistically significant. All analyses were conducted in R 4.4.2. The “caret” package was used for cross‐validation. “glmnet,” “ranger,” “e1071,” “nnet,” and “xgboost” were for LASSO, RF, SVM, MLP, XGBoost construction respectively. RESULTS Participant characteristics We initially screened a total of 710 participants: 178 with AD, 300 with CSVD, and 232 CU individuals. Based on predefined criteria, we excluded 29 AD, 24 CSVD, and 7 CU participants (Figure S1). Exclusions were primarily due to other neurological disorders, abnormal MRI findings, or ungradable retinal conditions. The final analytical cohort comprised 149 AD, 276 CSVD, and 225 CU participants. Baseline demographic, clinical, imaging, and biomarker characteristics are summarized in Table 1. The CSVD group exhibited a significantly higher prevalence of vascular risk factors, including hypertension, diabetes mellitus, dyslipidemia, smoking, and alcohol use, compared to the AD group (all p < 0.05). Both disease groups had lower MoCA and MMSE scores than CU controls (all p < 0.05). TABLE 1Demographic and clinical information.VariablesCUCSVDAD p valueNumber225276149Age, years57.46 (6.63)62.62 (8.91)65.29 (9.62)<0.001 a , b , c Male, n (%)87 (38.7)191 (69.2)54 (36.2)<0.001 b , c Hypertension, n (%)45 (20.0)106 (38.4)41 (27.5)<0.001 b , c Diabetes, n (%)13 (5.8)109 (39.5)20 (13.4)<0.001 a , b , c Dyslipidemia, n (%)56 (24.9)39 (14.1)37 (24.8)0.003 b , c Smoking, n (%)26 (11.6)134 (48.6)19 (12.8)<0.001 b , c Drinking, n (%)37 (16.4)103 (37.3)14 (9.4)<0.001 b , c Education, years11.29 (3.71)8.30 (4.28)9.14 (4.32)<0.001 a , b MoCA score26.0 (24.0–28.0)20.0 (14.8–24.0)14.0 (10.0–19.0)<0.001 a , b , c MMSE score29 (28.0–29.0)26 (22.8–28.0)22 (19.0–26.0)<0.001 a , b , c DWMH, n (%)<0.001 a , b , c 070 (31.1)0 (0.0)27 (18.1)1154 (68.4)128 (46.4)97 (65.1)21 (0.4)112 (40.5)25 (16.8)30 (0.0)36 (13.1)0 (0.0)PWMH, n (%)<0.001 a , b , c 0104 (46.2)0 (0.0)27 (18.1)1121 (53.8)95 (34.6)81 (54.4)20 (0.0)114 (41.2)41 (27.5)30 (0.0)67 (24.2)0 (0.0)PVS‐BG, n (%)<0.001 a , b , c 059 (26.2)2 (0.7)9 (6.0)1134 (59.6)146 (52.9)103 (69.1)226 (11.6)78 (28.1)32 (21.5)36 (2.7)41 (15.0)5 (3.4)40 (0.0)9 (3.3)0 (0.0)PVS‐CS, n (%)<0.001 a , b 018 (8.0)2 (0.7)0 (0.0)198 (43.6)83 (30.1)37 (24.8)275 (33.3)99 (36.0)66 (44.3)331 (13.8)74 (26.8)41 (27.5)43 (1.3)18 (6.5)5 (3.4)Lacunes, n (%)9 (4.0)123 (44.5)15 (10.1)<0.001 a , b , c CMBs, n (%)27 (12.0)130 (47.1)50 (33.6)<0.001 a , b , c Total CSVD score0 (0–1)2 (1–3)1 (0–1)<0.001 a , b , c t‐tau, pg/ml0.60 (0.15–1.21)0.97 (0.70–1.45)1.32 (0.62–2.09)<0.001 a , b , c p‐tau181, pg/ml0.60 (0.28–1.06)1.04 (0.54–1.66)2.39 (1.53–3.42)<0.001 a , b , c Aβ 42/40, pg/ml0.06 (0.04–0.09)0.04 (0.02–0.07)0.04 (0.02–0.07)0.011 a , b Abbreviations: AD, Alzheimer's disease; Aβ, amyloid beta; CMBs, cerebral microbleeds; CSVD, cerebral small vessel disease; CU, cognitively unimpaired; DWMH, deep white matter hyperintensity; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; p‐tau181, phosphorylated tau at threonine 181; PVS‐BG, perivascular spaces in basal ganglia; PVS‐CS, perivascular spaces in centrum semiovale; PWMH, periventricular white matter hyperintensity; t‐tau, total tau. a p < 0.05 for the comparison between CU and AD patients. b p < 0.05 for the comparison between CU and CSVD patients. c p < 0.05 for the comparison between AD and CSVD patients Demographic and clinical information. Abbreviations: AD, Alzheimer's disease; Aβ, amyloid beta; CMBs, cerebral microbleeds; CSVD, cerebral small vessel disease; CU, cognitively unimpaired; DWMH, deep white matter hyperintensity; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; p‐tau181, phosphorylated tau at threonine 181; PVS‐BG, perivascular spaces in basal ganglia; PVS‐CS, perivascular spaces in centrum semiovale; PWMH, periventricular white matter hyperintensity; t‐tau, total tau. p < 0.05 for the comparison between CU and AD patients. p < 0.05 for the comparison between CU and CSVD patients. p < 0.05 for the comparison between AD and CSVD patients Neuroimaging confirmed higher burdens of CSVD markers, including PWMH and DWMH, PVS in the basal ganglia, CMBs, and lacunes, in the CSVD group compared to the AD group (all p < 0.05, Table 1). We tested AD blood biomarkers in our study. The AD group exhibited elevated levels of p‐tau181 (median 2.39, IQR 1.53–3.42 pg/ml) and t‐tau (median 1.32, IQR 0.62–2.09 pg/ml), with reduced Aβ42/Aβ40 ratio (median 0.04, IQR 0.02–0.07) compared to the CU group (p‐tau181: median 0.60, IQR 0.28–1.06 pg/ml; t‐tau: median 0.60, IQR 0.15–1.21 pg/ml; Aβ42/Aβ40 ratio: median 0.06, IQR 0.04–0.09; all p < 0.05, Table 1 and Figure 2). On the other hand, the CSVD group also exhibited a significantly lower Aβ42/Aβ40 ratio (median 0.04, IQR 0.02–0.07) and higher p‐tau181 levels (median 1.04, IQR 0.54–1.66 pg/ml) over the CU individuals (all p < 0.05). Compared to the CSVD group, the AD group exhibited significantly higher p‐tau181 (median 2.39, IQR 1.53–3.42 pg/ml vs. median 1.04 pg/ml, IQR 0.54–1.66 pg/ml, p < 0.001) and t‐tau (median 1.32, IQR 0.62–2.09 pg/ml vs. median 0.97 pg/ml, IQR 0.70‐1.45 pg/ml, p = 0.006) levels, whereas Aβ42/Aβ40 ratios were comparable between groups (Table 1 and Figure 2). Blood biomarkers and RIPLs across three groups. A, Comparisons of log2‐transformed blood biomarkers among the CU, CSVD, and AD groups. B, Comparisons of RIPLs among the three groups, adjusted for age, sex, education, diabetes, and hypertension (Bonferroni‐corrected p values). Aβ, amyloid beta; AD, Alzheimer's disease; CSVD, cerebral small vessel disease; CU, cognitively unimpaired; p‐tau181, phosphorylated tau at threonine 181; RIPLs, retinal ischemic perivascular lesions; t‐tau, total tau. Ocular microvascular differences between AD and CSVD OCTA analysis revealed distinct ocular microvascular signatures between the two groups (Figures 2 and 3). In the retinal microvasculature, patients with CSVD demonstrated a significantly higher prevalence (p = 0.015) and burden (p = 0.021) of RIPLs than those with AD (Figure 2 and Table S1 in supporting information), whereas the average vessel densities (SVC and DVC) did not significantly differ between the two groups (p > 0.05, Figure 3 and Table S2 in supporting information). Ocular microvascular metrics and diagnostic performance. A, Comparisons of OCTA parameters among the CU, CSVD, and AD groups, adjusted for age, sex, education, diabetes, and hypertension (Bonferroni‐corrected p values). B, ROC curves illustrating the diagnostic performance of the baseline clinical model (developed via LASSO regression), the combined clinical‐OCTA LASSO model, and the optimal‐performing combined machine learning models. AD, Alzheimer's disease; AUC, area under the curve; CC, choriocapillaris; CSV, choroidal stroma volume; CSVD, cerebral small vessel disease; CT, choroidal thickness; CVV, choroidal vessel volume; CU, cognitively unimpaired; CVI, choroidal vascularity index; DVC, deep vascular complex; FAZ, foveal avascular zone; LASSO, least absolute shrinkage and selection operator; OCTA, optical coherence tomography angiography; RF, random forest; ROC, receiver operating characteristic; SVC, superficial vascular complex; SVM, support vector machine. However, in the choroidal microvasculature, patients with AD showed significantly lower CC density than those with CSVD (p = 0.031, Figure 3) after adjustment for age, sex, education, hypertension, and diabetes. Hypertension‐stratified analyses and the interaction test for CC density are consistent with this finding (Tables S3 and S4 in supporting information). Furthermore, choroidal thickness and CVV were also lower in patients with AD (all p < 0.05, Figure 3), collectively indicating a marked impairment of choroidal microvasculature in AD. Machine learning model performance for differential diagnosis Machine learning models integrating OCTA metrics with clinical variables generally demonstrated improved discrimination compared to the clinical baseline model, with performance evaluated on the independent test set (Figure 3 and Figure S2 and Table S5 in supporting information). For differentiation between AD and CSVD, the RF and SVM models achieved the highest AUCs of 0.86 (95% confidence interval [CI]: 0.80–0.92). In distinguishing disease groups from CU individuals, the SVM model yielded the highest AUC for CSVD classification (0.93, 95% CI: 0.90–0.97), whereas the RF model achieved the best performance for AD classification (0.85, 95% CI: 0.79–0.92). Additional performance metrics, including accuracy, sensitivity, and specificity, are summarized in Table S5. Association between ocular microvascular metrics and CSVD burden The RIPLs were correlated with higher total CSVD score in the CSVD group (relative risk 1.21, 95% CI, 1.00–1.45, p = 0.049) but not in the AD group (Table 2). No significant associations were observed between retinal and choroidal microvascular densities and total CSVD score in either the AD or CSVD groups (all p > 0.05, Table S6 in supporting information). Notably, lower MoCA scores were associated with greater PWMH, DWMH, and total WMH, as well as with lacunes, CMBs, and total CSVD score in patients with CSVD (all p < 0.05, Table S7 in supporting information), whereas lower MoCA scores were associated only with greater PWMH in the AD group (p = 0.010, Table S7 in supporting information). TABLE 2Associations between RIPLs and CSVD markers.VariablesRIPLsCSVD (n = 276)AD (n = 149)RR (95% CI) p valueRR (95% CI) p valuePWMH1.01 (0.83 to 1.23)0.8861.01 (0.77 to 1.34)0.928DWMH1.13 (0.94 to 1.35)0.1891.22 (0.96 to 1.56)0.108TWMH1.08 (0.89 to 1.31)0.4181.15 (0.89 to 1.49)0.282PVS‐BG1.09 (0.91 to 1.3)0.3641.05 (0.8 to 1.36)0.740PVS‐CS1.05 (0.88 to 1.26)0.5821.13 (0.86 to 1.47)0.386Lacunes1.15 (0.96 to 1.38)0.1401.08 (0.86 to 1.37)0.496CMBs1.11 (0.92 to 1.34)0.2810.94 (0.71 to 1.24)0.650Total CSVD score1.21 (1.00 to 1.45)0.0491.04 (0.8 to 1.35)0.786 Notes: Associations were evaluated using Poisson generalized estimating equation (GEE) models to account for inter‐eye correlation, with adjustment for age, sex, education, diabetes, and hypertension.Abbreviations: AD, Alzheimer's disease; CMBs, cerebral microbleeds; CI, confidence interval; CSVD, cerebral small vessel disease; DWMH, deep white matter hyperintensity; PVS‐BG, perivascular spaces in basal ganglia; PVS‐CS, perivascular spaces in centrum semiovale; PWMH, periventricular white matter hyperintensity; RIPLs, retinal ischemic perivascular lesions; RR, relative risk; TWMH, total white matter hyperintensity. Associations between RIPLs and CSVD markers. Notes: Associations were evaluated using Poisson generalized estimating equation (GEE) models to account for inter‐eye correlation, with adjustment for age, sex, education, diabetes, and hypertension. Abbreviations: AD, Alzheimer's disease; CMBs, cerebral microbleeds; CI, confidence interval; CSVD, cerebral small vessel disease; DWMH, deep white matter hyperintensity; PVS‐BG, perivascular spaces in basal ganglia; PVS‐CS, perivascular spaces in centrum semiovale; PWMH, periventricular white matter hyperintensity; RIPLs, retinal ischemic perivascular lesions; RR, relative risk; TWMH, total white matter hyperintensity. Association between ocular microvascular metrics and core AD biomarkers Association analyses revealed that CC impairment was linked to core AD biomarkers. In the AD group, reduced CC density was associated with higher p‐tau181 and t‐tau levels (all p < 0.05, Figure 4 and Table S8 in supporting information). However, these associations were absent in the CSVD group (Table S9 in supporting information). Associations between OCTA metrics and AD blood biomarkers. Blood biomarkers were log2‐transformed prior to analysis. A, Associations between CC and AD blood biomarkers. B, Associations between OCTA metrics and blood biomarkers in the AD group. C, Associations between OCTA metrics and blood biomarkers in the CSVD group. Associations were evaluated using generalized estimating equation models, accounting for inter‐eye correlation and adjusting for age, sex, education, diabetes and hypertension. *p < 0.05, **p < 0.01. Aβ, amyloid beta; AD, Alzheimer's disease; CC, choriocapillaris; CSV, choroidal stroma volume; CSVD, cerebral small vessel disease; CT, choroidal thickness; CVI, choroidal vascularity index; CVV, choroidal vessel volume; DVC, deep vascular complex; FAZ, foveal avascular zone; OCTA, optical coherence tomography angiography; p‐tau181, phosphorylated tau at threonine 181; SVC, superficial vascular complex; t‐tau, total tau. DISCUSSION In this study, we reported distinct ocular microvascular alterations in AD and CSVD. CSVD may be characterized by a higher burden of RIPLs, and RIPLs were associated with total CSVD score within the group. On the other hand, CC reduction was more prominent in AD, consistent with our previous study. 15 Notably, CC impairment in AD was associated with the core AD blood biomarker p‐tau181, whereas this association was not observed in the CSVD group. Machine learning models integrating clinical and OCTA features demonstrated high accuracy in distinguishing AD from both CSVD and CU, highlighting the potential of OCTA as a non‐invasive diagnostic approach. Consistent with established AD biomarker biology, 2 our study showed a decreased Aβ42/Aβ40 ratio in the AD group. We also found that patients with CSVD had a lower Aβ42/Aβ40 ratio than CU individuals. This finding is consistent with recent plasma and CSF biomarker studies, which reported that greater CSVD severity, particularly greater WMH burden and the presence of lacunes, was associated with a lower Aβ42/Aβ40 ratio, mainly driven by increased Aβ40 levels. 31 , 32 , 33 However, the blood Aβ42/Aβ40 ratio is influenced by multiple factors, 34 , 35 and its diagnostic performance for AD remains limited despite Aβ being a core pathological feature of AD. By comparison, blood p‐tau181 has shown higher diagnostic accuracy and strong associations with cerebral amyloid burden. 14 , 25 , 36 The retina provides a unique window for in vivo assessment of cerebral microvasculature. 6 Our study identified distinct retinal microvascular patterns in CSVD and AD. CSVD was associated with a higher burden of RIPLs, suggesting more focal ischemic retinal lesions. RIPLs are considered markers of previous retinal ischemic infarcts and have been associated with ischemic stroke and cardiovascular disease. 28 , 37 , 38 In our cohort, RIPLs were associated with total CSVD score in CSVD, whereas this association was not observed in AD. Consistent with our findings, a previous study also reported an association between RIPLs and WMH burden in patients with a single subcortical infarction. 20 Together, these findings suggest that retinal microvascular alterations may reflect cerebral vascular injury. Further studies incorporating quantitative neuroimaging markers and longitudinal follow‐up are needed to validate these associations. Although CSVD is commonly associated with vascular risk factors, including hypertension and diabetes, which may affect the CC, we unexpectedly observed lower CC density in AD than in CSVD. To our knowledge, this difference has not previously been reported. Moreover, it remained significant after adjustment for age, sex, education, hypertension, and diabetes, indicating that the measured demographic characteristics and vascular risk factors did not fully account for this finding. Because the CC depends on trophic support from the RPE, its greater impairment in AD may reflect both direct Aβ deposition within CC and indirect effects mediated by Aβ‐induced RPE injury. Animal studies and human post mortem investigations have identified Aβ deposits within the walls of choroidal vessels, particularly arteries, where they may exert direct cytotoxic effects on the choroidal vasculature. 39 , 40 , 41 , 42 In parallel, studies in aged mice and older human eyes have shown that Aβ‐related deposits beneath the RPE can activate complement and other inflammatory responses, resulting in RPE injury. 43 , 44 Because soluble vascular endothelial growth factor (VEGF) secreted by the RPE is essential for maintaining the CC, RPE dysfunction may reduce VEGF availability and thereby contribute to CC atrophy or loss. 43 , 44 These proposed mechanisms, however, require experimental validation. Our findings further suggest that CC density may have potential utility as an ocular biomarker associated with AD. 45 In this larger cohort of patients with sporadic AD, we replicated our previous findings in early‐onset AD that lower CC density was associated with serum p‐tau181. 15 Notably, this association was not observed in patients with CSVD. Chan et al. similarly reported an association between CC flow deficits and the CSF Aβ42/tau ratio in CU individuals with AD pathology. 46 However, they did not detect an association with CSF p‐tau181 levels, possibly because their relatively small sample size limited statistical power. These findings may suggest that CC impairment is more closely associated with AD‐related biological changes than with non‐specific vascular injury. Differentiating AD from CSVD remains clinically challenging due to overlapping phenotypic and neuroimaging profiles, 47 while definitive diagnostic modalities—such as Aβ PET and CSF assays—are constrained by high costs and invasiveness. Consequently, integrating OCTA metrics with machine learning algorithms provides a pragmatic strategy for clinical translation. 48 By leveraging distinct microvascular signatures alongside clinical baseline data, our predictive models achieved favorable discriminatory performance between AD and CSVD. This underscores the potential of machine learning–augmented OCTA as a scalable, non‐invasive triage modality. In clinical practice, such an approach may help stratify patients for confirmatory testing, thereby optimizing resource allocation and streamlining diagnostic workflows. The present study possesses several notable strengths. First, it provides a comprehensive comparison of both retinal and choroidal vascular parameters between patients with AD and CSVD. Second, the assessment of AD blood biomarkers across all three groups, together with Aβ PET confirmation in the AD group, strengthened the biological characterization of the participant groups and helped reduce potential confounding from unrecognized AD pathology in the CSVD and CU groups. Finally, the inclusion of a large, well‐characterized sample enhanced the statistical robustness and reliability of our findings. Several limitations of the present study warrant consideration. First, although AD blood biomarkers were assessed across all three groups, Aβ PET was not universally performed in the CSVD and CU groups, leaving a risk of unrecognized AD pathology. Second, given the absence of post mortem neuropathological validation, mixed or coexisting pathologies cannot be fully excluded in this in vivo cohort, including potential contributions from DLB, limbic‐predominant age‐related TDP‐43 encephalopathy, cerebral amyloid angiopathy, or other age‐related neurodegenerative processes. Third, AD biomarkers were quantified in serum. The absence of anticoagulants during serum collection may increase the susceptibility of proteins such as Aβ to coagulation‐related degradation and matrix entrapment. 14 , 34 , 35 Therefore, future studies using plasma‐based assays are warranted to further validate these findings. Fourth, RIPL assessment was manual and semiquantitative. While interobserver agreement was high, potential observer dependence cannot be entirely excluded, and the validation of automated quantitative methods is warranted in future studies. Finally, the single‐center design may limit the generalizability of our results, and external validation in larger multicenter cohorts is needed. AUTHOR CONTRIBUTIONS Shuting Zhang had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Conceptualization: Ziyi Zhang, Le Cao, Bo Wu, Shuting Zhang; methodology: Ziyi Zhang, Le Cao, Peng Lei; acquisition, analysis, or interpretation of data: all authors; writing—original draft: Ziyi Zhang, Le Cao, Peng, Liu, Shuting Zhang; critical revision of the manuscript for important intellectual content: Peng Lei, Ziyi Zhang, Le Cao; data curation and formal analysis: Ziyi Zhang, Le Cao, Peng Liu; obtained funding: Bo Wu, Shuting Zhang, Le Cao, Ziyi Zhang; administrative, technical, or material support: Peng Lei; supervision: Peng Lei, Bo Wu, Shuting Zhang. All authors read and approved the final manuscript. CONFLICT OF INTEREST STATEMENT The authors declare that they have no conflicts of interest. The sponsor had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. All data were accessible to all authors. The corresponding author had final responsibility for the decision to submit the manuscript for publication. Author disclosures are available in the Supporting Information. CONSENT STATEMENT Informed consent was obtained from all participants or their guardians. Supporting information Supporting Information Supporting Information ACKNOWLEDGMENTS The authors thank all participants and their families for their contribution to this study. This study was supported by the National Natural Science Foundation of China (82371210, U24A20690, U25A2073), the Scientific Research Innovation Capability Support Project for Young Faculty (ZYGXQNJSKYCXNLZCXM‐H14), Changping Laboratory (2025B‐07‐11), the Sichuan Science and Technology Program (2026NSFSC1623), the National Key R&D Program of China (2023YFC2506603), the Qimingxing Research Fund for Young Talents (HXQMX0052), 1·3·5 project for disciplines of excellence, West China Hospital, Sichuan University (2024HXFH023, ZYYC23016). Contributor Information Peng Lei, Email: peng.lei@scu.edu.cn. Bo Wu, Email: dr.bowu@hotmail.com. Shuting Zhang, Email: shutingzhang@scu.edu.cn. REFERENCES Associated Data Supplementary Materials Supporting Information Supporting Information