BDNF markers linked to cognitive impairment in kidney disease

Higher blood levels of brain-derived neurotrophic factor (BDNF) were associated with a lower likelihood of prevalent cognitive impairment in a cross-sectional study of 240 patients with chronic kidney disease, while a ratio combining BDNF with an estimate of kidney function produced the strongest exploratory machine-learning classification. Patients classified as cognitively impaired had a mean serum BDNF concentration of 8.95 ng/mL, compared with 10.71 ng/mL among those without impairment, a statistically significant difference. After adjustment for age, sex, stroke history, hypertension, diabetes and estimated glomerular filtration rate, each 1-ng/mL increase in BDNF was associated with an odds ratio of 0.72 for prevalent cognitive impairment, with a 95% confidence interval of 0.63–0.82.

The findings came from a retrospective analysis of routinely collected hospital records at Taizhou Fourth People’s Hospital in Jiangsu, China. Patients were identified between January 2024 and January 2026, and the final analytic cohort included 240 people with chronic kidney disease who had complete clinical information, cognitive assessments and blood samples. Cognitive status was evaluated with the validated Chinese version of the Montreal Cognitive Assessment, or MoCA, a 30-point screening instrument covering visuospatial and executive function, attention, memory, language and orientation. The prespecified impairment threshold was a score below 26. Of the 240 participants, 129 met that definition and 111 did not. Because the design was cross-sectional, the study examined associations with cognitive impairment present at the same clinical evaluation rather than predicting future decline.

Participants’ average estimated glomerular filtration rate was 9.01 mL/min/1.73 m², and the cohort included 169 men and 71 women. Stroke history was more common among patients with cognitive impairment, occurring in 26 of 129 participants, or 20.2%, compared with 5 of 111, or 4.5%, among those without impairment. Those classified as impaired were also older, with a mean age of 53.43 years versus 45.77 years. No significant between-group differences were reported for sex, hypertension or diabetes. Blood samples were collected during the same clinical evaluation period as the cognitive assessments, with serum stored at −80 °C until BDNF measurement by ELISA. The investigators analyzed BDNF both as a continuous concentration and in quartiles.

Researchers also examined the BDNF-to-eGFR ratio, an exploratory composite intended to represent circulating neurotrophic activity relative to renal filtration capacity. The ratio was not presented as an established clinical biomarker, and the investigators did not formally test whether it supplied additional information beyond BDNF alone. In the fully adjusted model, higher BDNF/eGFR values were associated with lower odds of cognitive impairment, with an odds ratio of 0.62 and a 95% confidence interval of 0.45–0.82. The analysis did not identify a significant correlation between BDNF and eGFR, with a Spearman coefficient of −0.039 and P = 0.552. Variance-inflation factors in a joint model were low, indicating no substantial multicollinearity.

The quartile analyses showed inverse associations for the markers overall, although the fully adjusted BDNF/eGFR estimates were not uniformly stepwise. Compared with the lowest BDNF quartile, participants in the highest quartile had an adjusted odds ratio of 0.09 for cognitive impairment, with a 95% confidence interval of 0.03–0.23. The corresponding comparison for the highest versus lowest BDNF/eGFR quartile was 0.27, with a confidence interval of 0.11–0.64. The fully adjusted BDNF/eGFR estimate for the second quartile versus the first was 1.68, with a 95% confidence interval of 0.69–4.12, and was therefore neither inverse nor statistically significant. Both quartile analyses had P values below 0.001 for trend. Restricted cubic spline analyses indicated an inverse dose-response relationship for BDNF with no significant evidence of nonlinearity. The BDNF/eGFR analysis also showed an inverse relationship but produced evidence of nonlinearity, with a reported P = 0.024.

Subgroup analyses indicated that the inverse associations were visible across several clinical categories while also revealing evidence of effect modification. For BDNF, significant associations were reported among participants younger than 60 years and those aged 60 or older, among women and men, and among patients with diabetes. A statistical interaction was detected for sex, with P = 0.004. The BDNF odds ratios were 0.84 for women and 0.49 for men, while the interaction indicates that the associations differed by sex. For the BDNF/eGFR ratio, inverse associations were likewise present in multiple age, sex and stroke-history strata. Significant interactions were found for sex, with P = 0.037, and stroke history, with P = 0.028, while age, hypertension and diabetes did not significantly modify the association. These subgroup estimates were exploratory, and the limited sample reduced their precision and generalizability.

The investigators compared nine machine-learning algorithms for classifying whether participants had cognitive impairment at the time of assessment. The candidate variables were age, sex, stroke history, hypertension, diabetes and either BDNF or the BDNF/eGFR ratio. The dataset was divided with a fixed random seed into a development cohort containing 70% of the data and a held-out testing cohort containing 30%, with five-fold cross-validation used for hyperparameter tuning in the development set. A random forest incorporating BDNF achieved the highest observed testing-set area under the receiver operating characteristic curve, 0.803, with a 95% confidence interval of 0.700–0.906, narrowly ahead of the support vector machine at 0.801. The authors presented these results as descriptive comparisons rather than evidence that random forest was statistically superior to the other algorithms.

Models incorporating BDNF/eGFR generally had higher observed discrimination than those using BDNF alone. The random forest with the ratio produced the largest reported testing-set area under the curve, 0.844, with a 95% confidence interval of 0.752–0.937. Its reported accuracy was 0.803, sensitivity 0.816, specificity 0.788 and F1 score 0.816, while a neural network achieved an area under the curve of 0.834 with a 95% confidence interval of 0.734–0.933. Calibration analyses descriptively compared estimated and observed probabilities, and decision-curve analyses estimated net benefit across decision thresholds. However, the held-out testing set was small, no external cohort was used, and no threshold was selected as clinically optimal. The results therefore show exploratory discrimination of contemporaneous cognitive status, not validated prediction of future cognitive decline or evidence of clinical utility.

To interpret the strongest-performing random forest, the team used SHapley Additive exPlanations, or SHAP, which assigns each feature a contribution to individual model outputs and overall predictions. In the BDNF-based model, BDNF was among the largest contributors, with higher values associated with negative SHAP values and less contribution toward classification as cognitively impaired. In the ratio-based model, BDNF/eGFR had the largest feature contribution, again with higher values generally linked to lower model-generated probabilities of impairment. SHAP plots showed that clinical characteristics also contributed to classifications. The method clarifies which variables drove the fitted models, but feature importance in a machine-learning system is not proof of biological importance, independent causality or a mechanism linking kidney dysfunction to cognition. The cross-sectional associations likewise cannot establish that BDNF changes precede impairment or that increasing BDNF would prevent it.

Several limitations constrain the conclusions. The retrospective, single-hospital design limits generalizability, and the investigators could not reconstruct exact numbers of screened or excluded records because no prospective screening log was maintained. Potentially relevant factors—including medication use, depressive symptoms, education, socioeconomic status, nutrition, physical activity, inflammatory markers, anemia-related measures, detailed vascular burden and stage-specific CKD characteristics—were not comprehensively available for adjustment, so residual confounding cannot be excluded. The MoCA was a screening instrument rather than a comprehensive neuropsychological assessment, making some misclassification possible. Blood samples were collected during hospitalization and at varying times of day, which may have introduced variation in serum BDNF, while dialysis status was recorded only at enrollment and no participants were receiving dialysis then. With only 240 patients, subgroup and machine-learning estimates were imprecise. The study’s results support the hypothesis that BDNF-related markers may be relevant to cognitive impairment in CKD, but longitudinal studies with independent validation are required to determine their temporal and prognostic significance.

Subject of Research: Associations between BDNF-related blood markers and prevalent cognitive impairment in chronic kidney disease

Article Title: BDNF-related biomarkers and cognitive impairment in chronic kidney disease: associations and machine learning-based discrimination analysis

Article References: Li, X., Qian, H., & Xiao, W. (2026). BDNF-related biomarkers and cognitive impairment in chronic kidney disease: associations and machine learning-based discrimination analysis. Frontiers in Medicine, 13, Article 1951634. https://doi.org/10.3389/fmed.2026.1951634

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1951634

Keywords: chronic kidney disease, cognitive impairment, brain-derived neurotrophic factor, BDNF, kidney function, blood biomarkers, machine learning, random forest, SHAP analysis, cross-sectional study, BDNF-related, biomarkers