Exploring predictors of low nutritional literacy in maintenance hemodialysis patients: a machine learning and network analysis approach

Background Patients undergoing maintenance hemodialysis (MHD) commonly experience substantial self-care challenges due to the complexity and highly restrictive nature of their treatment regimens. Nutritional literacy is a critical determinant of effective self-management in this population. However, the main predictors of low nutritional literacy and their relationships remain unclear, limiting early targeted interventions. This study aimed to identify important predictors of low nutritional literacy in MHD patients and to explore how these predictors relate to each other. Methods This study included 712 patients on maintenance hemodialysis (MHD). Latent profile analysis was used to identify individuals with low nutritional literacy, followed by machine learning to screen for key predictors. SHAP analysis was employed to visualize the importance of these predictors. Finally, network analysis was used to investigate the underlying relationships among the key predictors. Results The LightGBM model demonstrated the best predictive performance, with an AUC of 0.817 and an F1-score of 0.738. The LightGBM model identified nine key predictors: social support, depression, anxiety, self-efficacy, educational level, dialysis frequency, age, living arrangement (living with a spouse and children), and place of residence. Among these, social support had the highest SHAP value. Network analysis revealed that self-efficacy served as the central predictor(rs = 0.56), showing a significant positive correlation with social support (EW = 0.148) and significant negative correlations with depression and anxiety(EW = -0.209, EW = -0.158). The stability and accuracy analyses of the network demonstrated good overall stability, with a correlation stability (CS) coefficient for strength centrality of 0.67. In addition, the Bootstrap 95% confidence intervals were within an acceptable range. Conclusion In this study, the LightGBM model was used to identify nine key predictors of low nutritional literacy in MHD patients. Self-efficacy was confirmed as a core node in the predictor network for low nutritional literacy in this population. These findings provide reliable empirical evidence for the early clinical identification of MHD patients at high risk for low nutritional literacy. Furthermore, using self-efficacy as a core intervention target, healthcare professionals can implement targeted interventions for high-risk populations to effectively improve their nutritional literacy. This study offers important practical guidance for optimizing nutritional literacy and intervention strategies for MHD patients.