Serum BDNF and cognitive risk in maintenance hemodialysis: a machine learning study

Cognitive impairment is common in maintenance hemodialysis, but practical risk-assessment tools remain limited. We evaluated the association between baseline serum brain-derived neurotrophic factor (BDNF) and 1-year screening-defined incident cognitive impairment and examined whether BDNF added predictive information beyond routine clinical, laboratory, and baseline cognitive variables. This single-center prospective cohort included 130 maintenance hemodialysis patients with baseline Montreal Cognitive Assessment (MoCA) scores ≥26. The primary outcome was screening-defined incident cognitive impairment, defined as follow-up MoCA <26. Serum BDNF was measured by enzyme-linked immunosorbent assay. Five models were assessed using repeated 5-fold cross-validation; repeated nested cross-validation was added as a sensitivity analysis. During follow-up, 70 patients (53.8%) met the screening-defined outcome. Each 1-SD higher BDNF level was associated with lower odds of the outcome [adjusted OR, 0.52 (95% CI, 0.34-0.80); p = 0.003]. In the primary analysis, random forest and extreme gradient boosting (XGBoost) had AUCs of 0.798 and 0.797, respectively. Adding BDNF to the XGBoost base model changed AUC from 0.770 to 0.797 (DeLong p = 0.236), PR-AUC from 0.746 to 0.779, and Brier score from 0.190 to 0.181. In nested validation, XGBoost AUC was 0.782 (SD 0.019), while paired Base and Base + BDNF AUCs were 0.761 and 0.782, respectively. Lower BDNF was associated with the screening-defined outcome, but its added predictive value was small and statistically uncertain. Serum BDNF may therefore have value as an adjunctive marker for refining risk estimates rather than as a stand-alone predictor; independent multicenter external validation is required before clinical use.