Dietary patterns and exploratory gut microbiota profiles associated with diabetic retinopathy and cognitive impairment in type 2 diabetes

Abstract

Background

Diabetic retinopathy (DR) and cognitive impairment are closely related complications of type 2 diabetes mellitus (T2DM). Although dietary patterns and gut microbiota have each been linked to these conditions, their combined associations with co-occurring DR and cognitive impairment remain unclear. This study examined dietary patterns and exploratory gut microbiota profiles in relation to co-occurring DR and cognitive impairment in patients with T2DM.

Methods

In this cross-sectional study, 306 patients with T2DM were classified into four groups: no DR with normal cognition (DMCN), no DR with cognitive impairment (DMCI), DR with normal cognition (DRCN), and DR with cognitive impairment (DRCI). Dietary patterns were derived using principal component analysis. Gut microbiota composition was assessed using 16 S rRNA sequencing in a subset of 108 participants. Multinomial logistic regression was used to examine associations between dietary patterns and group classification, and microbiome analyses included diversity, taxonomic composition, exploratory differential abundance, and diet–microbiota correlations.

Results

Four dietary patterns were identified. In fully adjusted models, DP-I and DP-II were associated with higher odds of DMCI and DRCI, respectively, whereas DP-III was associated with lower odds of both DMCI and DRCN. DP-IV showed no significant association. Gut microbiota analyses showed modest but statistically significant group-related differences in community structure, with partial overlap across groups. Exploratory LEfSe analysis identified group-associated taxa, including higher relative abundances of Bifidobacterium, Streptococcus, and Dubosiella in DMCN and of Pseudomonas, Bilophila, and Sarcina in DRCI. However, these genus-level differences were not significant after covariate-adjusted MaAsLin2 analysis with false discovery rate (FDR) correction. Nominal diet–microbiota correlations were observed but were not statistically robust after FDR correction.

Conclusion

Dietary patterns were associated with clinical group classification based on DR and cognitive impairment in patients with T2DM. Gut microbiota analyses suggested modest, exploratory group-related differences, but diet–microbiota correlations were not statistically robust after FDR correction. These cross-sectional findings should be interpreted as hypothesis-generating and require validation in larger longitudinal studies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12986-026-01167-4.

Introduction

Diabetes mellitus (DM) is a major and growing global public health challenge, with its prevalence continuing to rise worldwide [1]. Type 2 diabetes mellitus (T2DM), accounting for approximately 90% of all diabetes cases, is associated with a wide range of chronic complications. Among these, diabetic retinopathy (DR) is one of the most common microvascular complications, affecting over one-fifth of patients with T2DM and representing a leading cause of visual impairment [2]. In addition, cognitive impairment has emerged as an important comorbidity in T2DM, characterized by deficits across multiple cognitive domains and associated with adverse clinical outcomes [3, 4]. Several meta-analyses have further reported a close association between DR and cognitive impairment among patients with diabetes [5–7]. Given the substantial burden of DR and its close relationship with cognitive impairment, early identification and proactive management of their co-occurrence have become important clinical and public health priorities.

The association between DR and cognitive impairment may be partly explained by shared pathological processes, including microvascular dysfunction, chronic inflammation, and oxidative stress [8]. Beyond these conventional mechanisms, the gut–retina–brain axis has been proposed as a conceptual framework linking metabolic, inflammatory, and neurovascular pathways [9, 10]. Existing studies have reported altered gut microbiota composition and community structure in T2DM patients with DR, suggesting that gut microbial changes may be related to retinal microvascular complications [11–13]. In parallel, studies of T2DM patients with cognitive impairment have also identified differences in gut microbial composition, although changes in overall microbial diversity appear less consistent [14, 15]. However, evidence remains limited on whether and how gut microbiota differ among T2DM patients with co-occurring DR and cognitive impairment.

Dietary patterns are a key modifiable factor that may be associated with both gut microbiota and diabetes-related retinal and cognitive outcomes. Accumulating evidence links dietary quality and dietary inflammatory potential to cognitive function and DR risk, with higher-quality dietary patterns generally associated with more favorable outcomes, while pro-inflammatory or highly processed patterns are associated with higher risk [16–18]. More recently, studies have begun to explore how diet-related variation in the gut microbiota may be associated with DR and cognitive outcomes. For example, in a cross-sectional study of 1,285 patients with diabetes, higher Dietary Index for Gut Microbiota (DI-GM) scores were associated with lower DR risk, suggesting a potential connection between microbiota-related dietary characteristics and retinal microvascular complications [19]. Similarly, regarding cognitive outcomes, another study reported that better diet quality was associated with a lower risk of cognitive impairment, with gut microbiota composition and miRNA expression proposed as potential intermediate factors [20]. However, few studies have examined the combined roles of dietary patterns and gut microbiota in the co-occurrence of DR and cognitive impairment, underscoring the need for further investigation.

To address this gap, this study examined the associations between PCA-derived dietary patterns and co-occurring DR and cognitive impairment in patients with T2DM. We further characterized exploratory gut microbiota profiles across clinical groups and assessed their associations with dietary patterns. By integrating dietary and microbiome-related analyses, this study aimed to provide exploratory evidence for understanding diabetes-related retinal and cognitive comorbidities.

Materials and methods

Study design and participants

This cross-sectional study was conducted between August and December 2025 in the ophthalmology and endocrinology departments of two tertiary hospitals in Shandong Province, China. The inclusion, exclusion, and analytical exclusion criteria are presented in Table 1. Based on DR diagnosis and education-adjusted MoCA (Montreal Cognitive Assessment) classification after exclusion of severe dementia using the Mini-Mental State Examination (MMSE), participants were classified into four groups: no DR with normal cognition (DMCN), no DR with cognitive impairment (DMCI), DR with normal cognition (DRCN), and DR with cognitive impairment (DRCI).

Table 1.

Inclusion, exclusion, and analysis exclusion criteria

Category	Criteria
Inclusion criteria	(1) Age ≥ 45 years
	(2) Confirmed diagnosis of T2DM by an endocrinologist
	(3) Best-corrected visual acuity (BCVA) ≥ 0.3 in at least one eye
	(4) Ability to complete questionnaires independently or with assistance
Exclusion criteria	(1) Inability to complete cognitive assessment due to severe visual impairment
	(2) History of major psychiatric or neurological disorders affecting cognitive function (e.g., schizophrenia, major depression, stroke, or Parkinson’s disease)
	(3) Malignant tumors or severe systemic diseases (e.g., advanced cardiac, hepatic, or renal dysfunction)
	(4) Use of medications known to significantly affect cognitive function (e.g., sedatives, antipsychotics, antidepressants, or anticholinergic agents)
Exclusion criteria during analysis	(1) Severe dementia, defined as an MMSE score ≤ 9
	(2) Implausible energy intake (< 500 or > 3500 kcal/day for women; <800 or > 4000 kcal/day for men)
Additional eligibility criteria for fecal sample collection	(1) No antibiotics, probiotics, or prebiotics within the past 4 weeks
	(2) No gastrointestinal diseases or acute infections that may affect gut microbiota
	(3) Fecal samples were collected before the use of laxatives, bowel preparation agents, or other gut microbiota–modifying interventions

A total of 306 participants were included in the final analysis. For microbiome analyses, eligible participants were invited to provide fecal samples after completing the questionnaire and clinical data collection. Initially, 236 participants met the additional eligibility criteria for fecal sample collection. However, because the inpatient stay was relatively short, not all eligible participants were able to provide qualified fecal samples within the required collection window. Ultimately, 108 participants with qualified fecal samples and complete microbiome data were included in the microbiome analyses.

Dietary assessment and derivation of dietary patterns

Dietary intake was assessed using a 25-item semi-quantitative food frequency questionnaire (FFQ) to capture habitual intake over the previous year, which has been previously applied in Chinese middle-aged and older populations [21]. For each item, participants reported their usual consumption frequency and portion size using either local weight units (e.g., 1 Liang = 50 g) or natural units (e.g., cups). Consumption frequency was recorded using predefined categories ranging from “never or occasionally” to “≥3 times per day” and subsequently converted into average daily intake values. Based on the Chinese Food Composition Table and local dietary habits, individual food items were grouped into predefined food categories for analyses, including refined grains, whole grains and tubers, green leafy vegetables, non-leafy vegetables, red meat, poultry, processed and organ meats, fish and shrimp, dairy and eggs, fruits, nuts and soybeans, and snacks and desserts.

Dietary patterns were derived using principal component analysis (PCA) based on standardized food-group intake variables. Food-group intakes were transformed into Z-scores before PCA to reduce the influence of differences in measurement scale. Dairy products and eggs were standardized separately before being combined into a single dairy-and-eggs variable. The suitability of the data for PCA was assessed using the Kaiser–Meyer–Olkin measure and Bartlett’s test of sphericity. Components were retained based on eigenvalues, inspection of the scree plot, cumulative explained variance, and interpretability of the resulting dietary patterns. Varimax rotation with Kaiser normalization was applied to improve interpretability. Food groups with absolute factor loadings ≥ 0.30 were considered meaningful contributors to each dietary pattern. Individual dietary pattern scores were calculated using SPSS regression, with higher scores indicating greater adherence to the corresponding pattern.

Clinical data collection and covariates

Sociodemographic, clinical, and laboratory characteristics

Sociodemographic and lifestyle characteristics, including age, sex, residence, marital status, family history of diabetes, education level, cigarette smoking, alcohol drinking, and exercise frequency, were collected using standardized questionnaires administered by trained investigators. Clinical characteristics, including body mass index (BMI), diabetes duration, hypertension, coronary heart disease, diabetic kidney disease, and insulin therapy, were obtained from electronic medical records. Laboratory indicators were also extracted from electronic medical records, including glycated hemoglobin (HbA1c), serum creatinine, estimated glomerular filtration rate (eGFR), total cholesterol, low-density lipoprotein cholesterol (LDL-C), and triglycerides (TG).

Assessment of diabetic retinopathy

All participants underwent standardized ophthalmologic examinations. DR was assessed by trained ophthalmologists using fundus photography according to established clinical criteria. Participants were classified as having DR if it was present in at least one eye.

Assessment of cognitive function

Cognitive function was assessed using a two-step approach. Participants were first screened using the MMSE to exclude severe dementia, defined as an MMSE score ≤ 9 [22]. This screening step was applied before MoCA-based classification to minimize unreliable recall in FFQ-based dietary assessment and to ensure reliable questionnaire completion. Global cognitive function was subsequently evaluated using the Chinese version of the MoCA, which assesses multiple domains, including orientation, visuospatial and executive function, language, memory, attention, naming, and abstraction. Total MoCA scores range from 0 to 30, with higher scores indicating better cognitive performance. One point was added for participants with ≤ 12 years of education. Cognitive impairment was defined as an education-adjusted MoCA score < 26, whereas normal cognition was defined as an education-adjusted MoCA score ≥ 26 [23].

Fecal sample collection and DNA extraction

Fecal samples were collected using sterile tubes, with participants instructed to avoid contaminating the sample with urine and to provide at least 0.5 g of stool prior to any interventions affecting gut microbiota (e.g., antibiotics, probiotics, or laxatives). Samples were transported to the laboratory within 2 h under cold-chain conditions and stored at − 80 °C until DNA extraction. Microbial genomic DNA was extracted using the TGuide S96 magnetic bead–based stool DNA extraction kit (Tiangen Biotech, Beijing, China; DP812) following the manufacturer’s instructions. DNA concentration and quality were assessed by fluorescence-based quantification, and integrity was verified before downstream sequencing.

16S rRNA sequencing and bioinformatics processing

The V3–V4 hypervariable regions of the bacterial 16S rRNA gene were amplified using universal primers 338 F (5′-ACTCCTACGGGAGGCAGCA-3′) and 806R (5′-GGACTACHVGGGTWTCTAAT-3′). PCR products were purified, indexed, and sequenced on the Illumina NovaSeq 6000 platform (Illumina, USA) using paired-end 250 bp reads. Raw reads were filtered for quality using Trimmomatic (v0.33) [24], and primers were removed with Cutadapt (v1.8.3) [25]. High-quality reads were denoised, merged, and chimera-checked using DADA2 to generate amplicon sequence variants (ASVs) [26]. Taxonomic classification of ASVs was performed in QIIME2 (v2020.6) using the SILVA 138 database with a confidence threshold of 0.7 [27, 28].

Microbiome data analysis

Microbiome data were analyzed using established bioinformatics workflows. To minimize the influence of sequencing depth, samples were rarefied to a uniform sequencing depth before downstream diversity analyses. ASVs shared across the four clinical groups were identified, and their taxonomic composition was summarized at the genus level to describe shared microbial features.

Alpha diversity was assessed using the ACE and Chao1 indices for species richness, and the Shannon and Simpson indices for community diversity and evenness. Group differences in alpha diversity were first examined using unadjusted non-parametric tests. To account for potential confounding, adjusted linear models were further performed for alpha diversity indices, controlling for age, sex, BMI, and diabetes duration.

Beta diversity was calculated based on Bray–Curtis dissimilarity and visualized using principal coordinates analysis (PCoA). Differences in community structure among groups were evaluated using analysis of similarities (ANOSIM). In addition, PERMANOVA based on Bray–Curtis dissimilarity was performed as a covariate-adjusted sensitivity analysis, controlling for age, sex, BMI, and diabetes duration.

Taxonomic composition was summarized at multiple taxonomic levels, with a focus on the phylum and genus levels. Differential taxa among groups were first assessed using non-parametric statistical tests. Linear discriminant analysis effect size (LEfSe) was applied as an exploratory approach to identify group-associated microbial features, with a logarithmic LDA score threshold of 2.0 [29]. To further evaluate the robustness of LEfSe-identified genera under covariate adjustment, MaAsLin2 was performed as a multivariable sensitivity analysis at the genus level [30], with group as the main variable and adjustment for age, sex, BMI, and diabetes duration. Multiple testing was controlled using the Benjamini–Hochberg false discovery rate (FDR). In addition, the Bacteroidota-to-Bacillota (B/F) ratio was calculated to characterize broad compositional differences in the gut microbiota.

Statistical analysis

Statistical analyses were performed using SPSS version 26.0 (IBM Corp., Armonk, NY, USA) and R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were expressed as mean ± SD, and categorical variables were presented as n (%). Group differences were assessed using one-way analysis of variance (ANOVA) or the Kruskal–Wallis test for continuous variables, and the chi-square test or Fisher’s exact test for categorical variables, as appropriate.

Multinomial logistic regression was used to examine associations between dietary pattern quartiles and clinical group classification, with the DMCN group as reference. Dietary pattern scores were categorized into quartiles and entered as ordinal variables to estimate odds ratios for a per-quartile increase. Two main models were constructed. Model 1 was adjusted for age and diabetes duration. Model 2 was further adjusted for sex, BMI, education level, residence, cigarette smoking, alcohol drinking, exercise frequency, hypertension, and coronary heart disease. Three sensitivity analyses were conducted to examine the robustness of the findings: (1) additional adjustment for glycemic control and insulin therapy; (2) additional adjustment for kidney function and lipid profile; and (3) quartile-specific analyses in which dietary pattern quartiles were entered as categorical variables, with the lowest quartile as the reference. To assess the robustness of the PCA factor-retention decision, an alternative five-component PCA solution was also examined as an additional sensitivity analysis.

Spearman correlation analysis was first conducted to assess unadjusted associations between dietary patterns and selected differential gut genera, followed by FDR correction for multiple comparisons. As sensitivity analyses, covariate-adjusted linear models were further performed, with log-transformed genus-level relative abundance as the outcome, PCA-derived dietary pattern scores as the main predictors, and age, sex, BMI, and diabetes duration as covariates. FDR correction was also applied to these sensitivity analyses. Two-sided p-values < 0.05 were considered nominally significant. For analyses involving multiple comparisons, FDR-adjusted q-values < 0.05 were considered statistically significant.

Results

Baseline characteristics

A total of 306 patients with T2DM were included: 85 (27.8%) in the DMCN group, 66 (21.6%) in the DMCI group, 83 (27.1%) in the DRCN group, and 72 (23.5%) in the DRCI group. The mean age was 60.71 ± 11.27 years, the mean BMI was 25.29 ± 3.18 kg/m², and the mean diabetes duration was 12.13 ± 7.69 years. Across the four groups, significant differences were observed in age, diabetes duration, and HbA1c levels, whereas most other baseline characteristics did not differ statistically (p > 0.05). The mean education-adjusted MoCA scores were 26.99 ± 2.30, 20.74 ± 3.00, 26.69 ± 2.49, and 20.76 ± 2.72 in the DMCN, DMCI, DRCN, and DRCI groups, respectively, with detailed MoCA score distributions presented in Supplementary Table 1. Detailed baseline characteristics are presented in Table 2.

Table 2.

Sociodemographic and clinical characteristics of participants by clinical group (n = 306)

Variable		DMCN (n = 85)	DMCI (n = 66)	DRCN (n = 83)	DRCI (n = 72)	p-value
Age, year		57.75 ± 11.09	66.44 ± 12.15	57.29 ± 9.94	62.09 ± 9.57	< 0.001
BMI, kg/m2		25.22 ± 3.48	25.35 ± 3.27	25.21 ± 2.91	25.41 ± 3.10	0.974
Diabetes duration, years		8.53 ± 6.32	11.31 ± 7.75	13.84 ± 7.60	15.14 ± 7.52	< 0.001
Sex	Male	48(56.5)	31(47.0)	49(59.0)	36(50.0)	0.428
	Female	37(43.5)	35(53.0)	34(41.0)	36(50.0)
Residence	Rural	41(48.2)	36(54.5)	40(48.2)	41(56.9)	0.610
	Urban	44(51.8)	30(45.5)	43(51.8)	31(43.1)
Marital status	Married	77(90.6)	56(84.8)	77(92.8)	66(91.7)	0.400
	Others	8(9.4)	10(15.2)	6(7.2)	6(8.3)
Family history of diabetes	Yes	15(17.6)	16(24.2)	19(22.9)	14(19.4)	0.725
	No	67(78.8)	45(68.2)	60(72.3)	52(72.2)
	Unclear	3(3.5)	5(7.6)	4(4.8)	6(8.3)
Education level, years	≤ 6	28(32.9)	29(43.9)	29(34.9)	28(38.9)	0.400
	7–9	28(32.9)	18(27.3)	21(25.3)	26(36.1)
	≥ 10	29(34.1)	19(28.8)	33(39.8)	18(25.0)
Cigarette smoking	Yes	22(25.9)	24(36.4)	21(25.3)	20(27.8)	0.436
	No	63(74.1)	42(63.6)	62(74.7)	52(72.2)
Alcohol drinking	Yes	24(28.2)	27(40.9)	23(28.0)	19(26.4)	0.219
	No	61(71.8)	39(59.1)	59(72.0)	53(73.6)
Hypertension	Yes	49(57.6)	35(53.0)	54(65.1)	37(51.4)	0.313
	No	36(42.4)	31(47.0)	29(34.9)	35(48.6)
Coronary heart disease	Yes	4(4.7)	10(15.2)	9(10.8)	9(12.5)	0.180
	No	81(95.3)	56(84.8)	74(89.2)	63(87.5)
Diabetic kidney disease	Yes	5(5.9)	5(7.6)	9(10.8)	11(15.3)	0.222
	No	80(94.1)	61(92.4)	74(89.2)	61(84.7)
Exercise frequency	None	45(52.9)	36(54.5)	56(67.5)	50(69.4)	0.192
	≤ 3 times/wk	34(40.0)	27(40.9)	24(28.9)	21(29.2)
	> 3 times/wk	6(7.1)	3(4.5)	3(3.6)	1(1.4)
Insulin therapy	No	45(52.9)	31(47.0)	36(43.4)	30(41.7)	0.489
	Any	40(47.1)	35(53.0)	47(56.6)	42(58.3)
HbA1c, %		7.08 ± 1.22	7.17 ± 1.34	7.62 ± 2.14	7.75 ± 1.61	0.024
Serum creatinine, µmol/L		72.50 ± 16.84	70.77 ± 17.65	75.96 ± 30.97	79.29 ± 28.57	0.164
eGFR, mL/min/1.73 m²		93.01 ± 23.34	94.11 ± 18.50	92.11 ± 21.26	87.18 ± 18.06	0.194
Total cholesterol, mmol/L		4.44 ± 1.23	4.50 ± 1.39	4.57 ± 1.34	4.34 ± 1.34	0.695
LDL-C, mmol/L		2.56 ± 1.07	2.79 ± 1.31	2.75 ± 1.02	2.56 ± 1.07	0.336
Triglycerides, mmol/L		1.75 ± 1.14	1.59 ± 1.03	1.66 ± 1.03	1.57 ± 1.16	0.787

Data are presented as mean ± SD or n (%). Group differences were assessed using one-way ANOVA, chi-square test, or Fisher’s exact test, as appropriate. Abbreviations: BMI, body mass index; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; LDL-C, low-density lipoprotein cholesterol; WK, week

Dietary pattern analysis in participants

Four dietary patterns were identified using PCA. The data showed moderate suitability for PCA, with a Kaiser–Meyer–Olkin value of 0.603 and a significant Bartlett’s test of sphericity (χ² = 493.522, df = 66, p < 0.001). Based on the scree plot, cumulative explained variance, and interpretability of the extracted components, a four-component solution was retained. Together, these four dietary patterns explained 54.77% of the total variance in dietary intake. The full rotated factor loading matrix, eigenvalues, explained variance, and cumulative variance are presented in Fig. 1, and the scree plot is provided in Supplementary Fig. 1. An alternative five-component solution was examined in sensitivity analyses.

Fig. 1.

Full rotated factor loading matrix of PCA-derived dietary patterns (n = 306) Note: Extraction method: principal component analysis; rotation method: varimax with Kaiser normalization. Absolute factor loadings ≥ 0.30 were considered meaningful. Eigenvalues, explained variance, and cumulative variance are shown for each retained dietary pattern

Dietary pattern I (DP-I), explaining 17.57% of the variance, was characterized by higher loadings for non-leafy vegetables, green leafy vegetables, whole grains and tubers, and red meat, together with a negative loading for refined grains. Dietary pattern II (DP-II), explaining 15.38% of the variance, was characterized by higher loadings for snacks and desserts, processed and organ meats, nuts and soybeans, refined grains, red meat, and fruits. Dietary pattern III (DP-III), explaining 11.47% of the variance, was characterized by higher loadings for dairy and eggs, fish and shrimp, fruits, and poultry. Dietary pattern IV (DP-IV), explaining 10.34% of the variance, was characterized by higher loadings for nuts and soybeans, poultry, and fish and shrimp, together with a negative loading for refined grains.

Dietary patterns and group classification

Multinomial logistic regression analyses showed that several PCA-derived dietary pattern scores were associated with clinical group classification in the fully adjusted model (Fig. 2, model 2). A per quartile increase in DP-I score was associated with higher odds of DMCI relative to DMCN (OR = 1.49, 95% CI: 1.07–2.07; p for trend = 0.018). DP-II was associated with higher odds of DRCI (OR = 2.10, 95% CI: 1.47–3.00; p for trend < 0.001), whereas its associations with DMCI and DRCN did not reach statistical significance. DP-III was associated with lower odds of both DMCI (OR = 0.58, 95% CI: 0.41–0.81; p for trend = 0.001) and DRCN (OR = 0.64, 95% CI: 0.47–0.86; p for trend = 0.004). DP-IV was not significantly associated with group classification. Similar results were observed in sensitivity analyses with additional adjustment for glycaemic control and insulin therapy, kidney function and lipid profile, and in quartile-specific models using the lowest quartile as the reference (Supplementary Figs. 2 and 3). An alternative five-component PCA solution was examined as a sensitivity analysis. Although the factor structure and some associations differed from the primary four-component solution, the main findings were not substantially altered (Supplementary Material 1; Supplementary Fig. 8; Supplementary Table 8).

Fig. 2.

Associations of dietary pattern quartiles with clinical group classification (n = 306) Note: ORs and 95% CIs were estimated per quartile increase in dietary pattern scores using multinomial logistic regression, with DMCN as the reference group. Model 1 was adjusted for age and diabetes duration. Model 2 was additionally adjusted for sex, BMI, education level, residence, cigarette smoking, alcohol drinking, exercise frequency, hypertension, and coronary heart disease

Gut microbiome characteristics

16 S rRNA sequencing was performed in 108 participants with qualified fecal samples. Compared with participants without fecal samples, those included in the microbiome analysis were significantly younger, whereas other baseline characteristics did not differ significantly (p > 0.05; Supplementary Table 2). Within the microbiome subset, differences were observed in age, diabetes duration, insulin therapy, HbA1c, eGFR, total cholesterol, LDL-C, and triglycerides. Baseline characteristics of the microbiome subset across the four clinical groups are summarized in Supplementary Table 3.

Microbiota diversity and community structure

A total of 82,554 ASVs were identified after sequence processing using the DADA2 pipeline. Among these, 1,055 ASVs were shared across all four groups, while additional ASVs were group-specific (Fig. 3A). The dominant genera among these shared ASVs included Bacteroides, Escherichia-Shigella, Blautia, Bifidobacterium, and Klebsiella, with detailed taxonomic composition provided in Supplementary Table 4. Rarefaction and Shannon rarefaction curves approached a plateau, indicating adequate sequencing depth and coverage (Supplementary Fig. 4).

Fig. 3.

Gut microbiota diversity and community structure across clinical groups (n = 108) Note: (A) Venn diagram showing shared and unique amplicon sequence variants (ASVs) among the four groups. (B) Comparisons of α-diversity indices, including ACE, Chao1, Shannon, and Simpson indices. (C) Principal coordinates analysis (PCoA) based on Bray–Curtis distance showing the distribution of microbial community structure across groups. (D) Comparison of the Bacteroidota-to-Bacillota (B/F) ratio across groups. Group differences were assessed using appropriate parametric or non-parametric tests

For α-diversity, no significant differences were observed across groups in ACE, Chao1, or Shannon indices (all p > 0.05). The Simpson index differed between the DRCI and DMCI groups in unadjusted comparison (p = 0.021; Fig. 3B). In adjusted linear models controlling for age, sex, BMI, and diabetes duration, ACE, Chao1, and Shannon indices remained non-significant, whereas the overall group difference in the Simpson index remained statistically significant (overall p = 0.021; Supplementary Table 5).

For β-diversity, Bray-Curtis PCoA showed partial separation among the four groups, with considerable overlap (Fig. 3C). The first three principal coordinates explained 8.18%, 6.28%, and 5.01% of the total variance, respectively. ANOSIM indicated a statistically significant but modest difference in microbial community structure across groups (R = 0.105, p = 0.001). In adjusted PERMANOVA, group classification remained associated with Bray–Curtis community dissimilarity after controlling for age, sex, BMI, and diabetes duration (R² = 0.0414, F = 1.5065, p = 0.001; Supplementary Table 6). In addition, the B/F ratio differed across groups (Kruskal–Wallis p < 0.001), with higher median values in the DRCN and DRCI groups than in the DMCN and DMCI groups (Fig. 3D). The median (IQR) B/F ratios were 0.158 (0.058–0.564), 0.202 (0.019–0.388), 0.806 (0.317–1.197), and 0.577 (0.245–0.843) in the DMCN, DMCI, DRCN, and DRCI groups, respectively.

Microbiota composition and differential taxa

At the phylum level, the gut microbiota was mainly composed of Bacillota, Bacteroidota, and Pseudomonadota across all groups (Fig. 4A). Additional phylum-level comparisons showed nominal between-group differences in Bacteroidota, Thermodesulfobacteriota, and Actinomycetota (Supplementary Fig. 5). At the genus level, the most abundant genera included Bacteroides, Escherichia-Shigella, Bifidobacterium, Blautia, and Ruminococcus (Fig. 4B). Several genera showed nominal between-group differences, including Parabacteroides, Bacteroides, Coprococcus, Dorea, Alistipes, and Bifidobacterium (Fig. 4C; all nominal p < 0.05).

Fig. 4.

Gut microbiota composition, exploratory differential taxa, and associations with dietary patterns (n = 108) Note: (A) Relative abundance of dominant bacterial taxa at the phylum level across groups. (B) Relative abundance of dominant bacterial taxa at the genus level across groups. (C) Comparison of selected genera across groups with nominal statistical significance. (D) Cladogram of group-associated taxa identified by exploratory LEfSe analysis. (E) Linear discriminant analysis (LDA) scores of taxa identified by LEfSe. (F) Heatmap of Spearman correlations between dietary patterns and selected differential genera before multiple testing correction. *Nominal p < 0.05 before FDR correction

As an exploratory differential abundance analysis, LEfSe identified 18 taxa with LDA scores > 2.0 across groups (Fig. 4D-E). Several group-associated genera were observed in nominal LEfSe comparisons, including higher relative abundance of Bifidobacterium, Streptococcus, and Dubosiella in the DMCN group, and higher relative abundance of Pseudomonas, Bilophila, and Sarcina in the DRCI group. Some LEfSe-identified genera, including Bifidobacterium, Bilophila, Sarcina, Streptococcus, Pseudomonas, and Thomasclavelia, showed nominal associations in MaAsLin2 analysis. However, none remained statistically significant after FDR correction (q > 0.05 for all), indicating that the LEfSe-identified genus-level differences should be interpreted as exploratory signals rather than robust differential taxa (Supplementary Table 7).

Associations between dietary patterns and gut microbiota

Spearman correlation analyses between dietary patterns and selected differential gut genera showed several nominal associations before multiple testing correction (Fig. 4F). DP-I was positively correlated with Blautia (r = 0.216, p = 0.024), whereas DP-III was inversely correlated with Bacteroides (r = − 0.199, p = 0.040) and positively correlated with Blautia (r = 0.209, p = 0.030). However, none of these associations remained significant after FDR correction (Supplementary Fig. 6). In covariate-adjusted sensitivity analyses controlling for age, sex, BMI, and diabetes duration, several nominal associations were observed, but none remained statistically significant after FDR correction (Supplementary Fig. 7).

Discussion

In this cross-sectional study, PCA-derived dietary patterns were associated with clinical group classification based on DR and cognitive function among patients with T2DM. DP-I and DP-II were associated with higher odds of DMCI and DRCI, respectively, whereas DP-III was associated with lower odds of both DMCI and DRCN. In the microbiome subset, gut microbiota analyses showed modest but statistically significant group-related differences in community structure and taxonomic composition, with considerable overlap across groups. Exploratory LEfSe analysis identified several group-associated taxa; however, these genus-level differences were not statistically significant after covariate-adjusted MaAsLin2 analysis with FDR correction. In addition, although several nominal correlations between dietary patterns and selected gut genera were observed, none remained significant after FDR correction. Overall, these findings suggest that dietary patterns were associated with DR and cognitive function in patients with T2DM, whereas the microbiome findings provide exploratory evidence of modest group-related differences.

In this study, PCA-derived empirical dietary patterns showed different associations with clinical group classification. DP-I, a mixed vegetable–whole grain–red meat pattern, was associated with higher odds of DMCI. Although this pattern included several nutrient-rich food groups, its relatively high loading for red meat may be relevant, as higher red meat intake has been linked to poorer cognitive outcomes in previous studies [31]. DP-II, mainly characterized by higher loadings for snacks and desserts, processed and organ meats, refined grains, and red meat, was associated with higher odds of DRCI. Compared with DP-I, DP-II contained more dietary features commonly related to inflammatory and metabolic burden. This may reflect pathways involving glycaemic dysregulation, oxidative stress, systemic inflammation, and microvascular injury, which are relevant to both DR and cognitive impairment [32, 33]. This finding is consistent with recent multi-dataset evidence linking higher dietary inflammatory burden to DR and poorer cognitive outcomes in older adults with diabetes [34]. In contrast, DP-III, characterized by higher loadings for dairy and eggs, fish and shrimp, fruits, and poultry, was associated with lower odds of both DMCI and DRCN. This pattern may reflect a nutrient-dense dietary profile, consistent with evidence linking diets rich in fish, fruits, and higher-quality protein sources to better cognitive performance and lower risk of diabetic microvascular complications [35, 36]. Overall, these findings suggest that dietary pattern composition may be relevant to DR and cognitive function in patients with T2DM. The differences among DP-I, DP-II, and DP-III highlight the potential importance of limiting processed, refined, sweet, and red meat–dominant dietary components while considering dietary profiles rich in higher-quality protein sources and fruits. Longitudinal studies are needed to clarify temporality and clinical relevance.

To our knowledge, this is the first study to examine differences in gut microbiota among T2DM patients grouped by DR and cognitive impairment. In the microbiome subset, analyses revealed modest but statistically significant group-related differences, with α diversity largely similar across groups and beta diversity showing partial separation. These findings align with previous studies reporting β-diversity differences in DR [37, 38], whereas studies of T2DM patients with cognitive impairment have generally shown less consistent alterations in overall microbial diversity [14]. We also observed a higher B/F ratio in the DRCN and DRCI groups than in the DMCN and DMCI groups, consistent with a previous study in DR [11], whereas evidence in cognitive impairment is inconsistent [39]. Overall, gut microbiota differences in this study appeared more evident in groups with DR, whereas differences associated with cognitive impairment alone were less apparent, particularly for overall microbial diversity.

Exploratory LEfSe analysis identified several genus-level taxa that differed nominally across clinical groups; however, none of these genus-level differences remained statistically significant after covariate-adjusted MaAsLin2 analysis with FDR correction. In nominal comparisons, the DMCN group showed a higher relative abundance of Bifidobacterium. Previous studies in T2DM and DR have reported lower Bifidobacterium abundance associated with altered gut barrier, metabolic, or inflammatory profiles [13, 14]. In the DRCI group, Pseudomonas showed a higher relative abundance. This genus has been reported in proliferative DR cohorts and is predicted to participate in microbial metabolic pathways related to glucose and nicotinate degradation [40]. Previous studies have also reported heterogeneous genus-level alterations in DR and cognitive impairment, although the specific taxa identified vary across study populations, sequencing methods, and analytical approaches [12, 37, 38, 41]. These findings suggest that gut microbiota composition may differ across DR and cognitive impairment groups in T2DM, but the identified taxa should be interpreted as exploratory signals. Further studies with functional metagenomics and metabolomic profiling are needed to clarify their biological relevance.

The diet–microbiota correlation results require cautious interpretation. Although several nominal correlations between dietary patterns and selected gut microbial genera were observed, none remained statistically significant after FDR correction. This indicates that microbiota-mediated links were not statistically supported in this dataset, and alternative explanations for the observed diet–clinical associations should be considered. Glycaemic regulation may be one such pathway, as a recent study in patients with T2DM showed that higher HbA1c levels were associated with increased DR risk, particularly among individuals with dietary patterns characterized by higher animal-protein or processed-food intake, or lower vegetable intake [32]. Dietary inflammatory potential may represent another relevant pathway. A prospective cohort study involving 16,224 patients with diabetes reported that higher DII scores were associated with an increased risk of DR [17], while a meta-analysis of nine observational studies involving 19,379 participants found that higher DII scores were associated with higher odds of cognitive impairment [42]. Further, the absence of FDR-significant diet–microbiota correlations does not exclude a possible diet–microbiota–disease relationship, but indicates that this pathway was not statistically supported in the present dataset. Accordingly, future longitudinal studies incorporating functional metagenomics, metabolomics, inflammatory biomarkers, and more detailed lifestyle assessment are needed to clarify the pathways linking dietary patterns with DR and cognitive function in T2DM.

Limitations

This study has several limitations. First, the cross-sectional design precludes causal inference regarding the relationships among dietary patterns, gut microbiota, and the co-occurrence of DR and cognitive impairment; thus, temporal sequence, directionality, and mediation pathways cannot be established. Second, the microbiome analyses were based on a feasibility-based subset of participants with qualified fecal samples rather than the full study population. Participants included in the microbiome subset were younger than those without fecal samples, and the relatively small microbiome sample may have limited statistical power, particularly after FDR correction. Although age-adjusted and multivariable sensitivity analyses were performed for key microbiome outcomes, residual selection bias and insufficient power cannot be excluded. Moreover, nominal diet–microbiota correlations did not remain significant after FDR correction, limiting inference regarding a microbiota-mediated pathway in this dataset.

Third, cognitive impairment was defined using the MoCA rather than a comprehensive neuropsychological diagnosis. Severe dementia was excluded before MoCA-based classification to improve the reliability of FFQ-based dietary assessment and questionnaire completion, which may limit generalizability to T2DM patients with severe cognitive impairment. Residual misclassification related to education, age, or visual impairment also cannot be fully excluded. Fourth, dietary intake was assessed using an FFQ, which is subject to recall bias and measurement error and may not fully capture short-term dietary influences on gut microbiota. Total energy intake was not included as a covariate because the FFQ was designed to assess habitual food-group intake patterns rather than to provide a comprehensive estimate of total energy intake. Finally, despite adjustment for multiple sociodemographic, lifestyle, and clinical covariates, residual confounding remains possible, and recruitment from two tertiary hospitals in one province of China may limit generalizability. Future studies with larger, more diverse, and longitudinal samples are needed to validate these findings.

Conclusion

This study simultaneously examined dietary patterns and gut microbiota composition in relation to the co-occurrence of DR and cognitive impairment among patients with T2DM. PCA-derived dietary patterns showed different associations with clinical group classification, with DP-I and DP-II associated with higher odds of DMCI and DRCI, respectively, and DP-III inversely associated with both DMCI and DRCN. In the microbiome subset, modest but statistically significant group-related differences in gut microbiota composition and community structure were observed, although genus-level findings should be interpreted as exploratory. Importantly, correlations between dietary patterns and selected microbial genera were no longer significant after FDR correction, indicating that this study does not provide robust evidence for microbiota-mediated links between dietary patterns and DR or cognitive function. Overall, these findings support the relevance of considering dietary patterns in relation to DR and cognitive function in T2DM, while the microbiome findings remain exploratory and require validation in longitudinal, multi-omics studies.

Abbreviations

ACE

Abundance-based coverage estimator

ANOSIM

Analysis of similarities

ASV

Amplicon sequence variant

B/F ratio

Bacteroidota-to-bacillota ratio

BMI

Body mass index

DI-GM

Dietary index for gut microbiota

DII

Dietary inflammatory index

DP

Dietary pattern

DR

Diabetic retinopathy

eGFR

estimated glomerular filtration rate

FDR

False discovery rate

FFQ

Food frequency questionnaire

HbA1c

Glycated hemoglobin

IQR

Interquartile range

LDL-C

Low-density lipoprotein cholesterol

LEfSe

Linear discriminant analysis effect size

MaAsLin2

Microbiome multivariable association with linear models 2

MMSE

Mini-mental state examination

MoCA

Montreal cognitive assessment

T2DM

Type 2 diabetes mellitus

TG

Triglycerides

Author contributions

YS, HYC, DSL: Conceptualization and methodology; YS, XJS, BGH: Investigation; YS, CLL: Data curation and formal analysis; YS, CLL: Software and bioinformatic analysis; HYC, DSL: Supervision and project administration; DSL: Funding acquisition and resources; YS: Writing – original draft; YS, HYC, CLL, DSL: Writing – review & editing; DSL: Final approval and overall responsibility for the work.

Funding statement

This work was supported by the Joint Project of the National Administration of Traditional Chinese Medicine (Grant No. GZY-KJS-SD-2024-059) and the Research Project of the Shandong Association for Science and Technology (Grant Nos. 26010132009055 and 26010132009054).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to privacy and ethical restrictions but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

All study procedures were approved by the Ethics Committee of Shandong University (No. 2024-R-144; December 12, 2024) and were conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment.

Competing interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During manuscript preparation and revision, AI-assisted language tools were used only for language polishing and editorial refinement. All intellectual content, data interpretation, and final revisions were reviewed and approved by the authors.