Hippocampal subfield radiomics reveals the link between inflammatory status and mild cognitive impairment in patients with non-dialysis chronic kidney disease.

OBJECTIVE: To investigate the associations among inflammatory status, hippocampal subfield radiomics phenotypes, and mild cognitive impairment (MCI) in patients with non-dialysis chronic kidney disease (CKD), and to construct an interpretable individualized risk prediction model. METHODS: Patients with non-dialysis CKD stages 3-5 were included from a primary cohort of 212 patients, including 86 with CKD-MCI and 126 cognitively normal individuals, and an external validation cohort of 98 patients. Clinical data, inflammatory markers, and Montreal Cognitive Assessment (MoCA) data were collected. Based on 24 bilateral hippocampal subfields derived from high-resolution structural MRI, radiomics features were extracted and dimensionally reduced using mRMR and LASSO regression to construct a hippocampal radiomics score (Hip-Radscore). Mediation analysis was used to explore the association pathway among inflammation, Hip-Radscore, and MCI. A combined prediction model was constructed using the XGBoost algorithm and externally validated, with SHAP applied to interpret the model decision-making mechanism. RESULTS: Patients with CKD-MCI exhibited an inflammatory status characterized by elevated hsCRP and SII levels. The Hip-Radscore constructed from eight core features, mainly texture features enriched in subfields such as CA1 and the subiculum, showed good performance in identifying MCI (AUC = 0.81), significantly outperforming the traditional hippocampal volume model (AUC = 0.67, p = 0.002). Mediation analysis showed that Hip-Radscore partially mediated the association between elevated hsCRP and lower MoCA scores; the indirect pathway through Hip-Radscore accounted for 39.4% of the total hsCRP-MoCA association. External validation confirmed that the XGBoost combined model integrating clinical inflammatory variables and Hip-Radscore achieved the best discriminative performance (AUC = 0.86), with good calibration and net benefit. SHAP analysis clearly showed that higher Hip-Radscore and hsCRP levels were the core features driving the model to predict a high risk of MCI. CONCLUSION: Elevated hsCRP was associated with hippocampal subfield radiomics abnormalities and cognitive decline in patients with non-dialysis CKD. Hip-Radscore may sensitively capture hippocampal tissue heterogeneity and may serve as a potential imaging intermediate phenotype linking inflammatory status to cognitive dysfunction. The interpretable XGBoost model integrating systemic inflammatory indicators and local radiomics features may provide a useful tool for identifying patients at high risk of CKD-related MCI.