Predicting cognitive impairment in chronic kidney disease patients using structural and functional brain network: An application study of artificial intelligence.

ABSTRACT

OBJECTIVE: To develop and validate artificial intelligence models for the prediction of cognitive impairment in chronic kidney disease (CKD) patients using structural and functional brain network. METHODS: This study retrospectively recruited 621 CKD patients and 625 healthy controls in Jinling hospital and 57 CKD patients in Hainan hospital. These CKD patients were divided into cognitive function impairment (CFI) group and non-CFI group based on diagnostic criteria. All patients underwent brain MRI scan, neuropsychological test and laboratory exam. A deep learning model (Attention MLP) based on structural and functional sub-network (determined by the comparison between the patients and healthy controls) topological properties was developed to generate the MRI signature for the discrimination of CFI. Finally, a clinical-topological logistic regression model was built by combining MRI signature and clinical features. The area under curve (AUC), sensitivity and specificity were calculated to evaluate the model performance. Delong test was used to examine the difference of AUCs between models. The integrated discrimination improvement (IDI) and net reclassification index (NRI) between models were calculated. RESULTS: Attention MLP model performed well in both internal test set and external test set (AUC = 0.744 and 0.763, respectively). After combining with the clinical features, the model performance was further improved both in the internal (AUC 0.748) and external test sets (AUC 0.774), while both IDI and NRI were significant (all p < 0.05) in the external test set. According to the comprehensive comparison, the AUC of the Attention MLP model was significantly or marginal significantly higher than that of traditional machine learning models (logistic regression AUC = 0.634; support vector machine AUC = 0.613; decision tree AUC = 0.539; XGBoost AUC = 0.639) in internal test set. The results showed that the model built on the combining of structural and functional networks data outperformed those on the single network, as well as the connection matrix. CONCLUSION: The result indicated that the integration of the clinical information and the MRI signature generated by artificial intelligence model based on structural and functional network topological properties could help to predict the CFI of CKD patients effectively. Our results provided a set of quantifiable imaging biomarkers for CFI which may be beneficial to CKD patients.