Deep Learning May Identify Cognitive Impairment in Kidney Disease

Beijing Institute of Technology Press Co., Ltd

Chronic kidney disease (CKD) is a major global health concern that affects not only kidney function but also multiple other organs, including the brain. Cognitive impairment (CI) is a common complication of CKD. It can affect executive function, attention, and memory, reducing patients' quality of life and potentially interfering with treatment adherence.

Although previous studies have reported associations between impaired kidney function and brain alterations, the neuroimaging characteristics associated with cognitive impairment across different stages of CKD remain incompletely understood. In particular, it is important to identify brain imaging features that are associated with both CKD progression and cognitive status.

To address this issue, the research team developed a neuroimaging feature-selection framework that considers both CKD stage and cognitive status. Based on the selected features, they further developed a deep learning model to identify cognitive impairment in patients with CKD.

The study included a discovery cohort of 342 patients with CKD, comprising 177 patients with cognitive impairment and 165 without cognitive impairment, as well as 58 age- and sex-matched healthy controls. The researchers collected multimodal magnetic resonance imaging (MRI) data to characterize brain alterations from different perspectives.

Three imaging techniques were used: T1-weighted imaging to assess brain structure, arterial spin labeling (ASL) to measure cerebral blood flow (CBF), and quantitative susceptibility mapping (QSM) to characterize tissue susceptibility.

The research process consisted of two main stages. First, the team analyzed neuroimaging changes across CKD stages 1 to 5 and compared imaging features between patients with and without cognitive impairment. Features were retained only when they showed both significant stage-related trends and significant differences associated with cognitive status. This strategy resulted in 24 selected neuroimaging features, including 17 structural features and 7 susceptibility features.

Second, the selected features were incorporated into a Dual-Branch Feature-Gated Network (DBFG-Net). The model uses two separate branches to process structural and susceptibility information, respectively, before integrating them through a gated fusion module. This design enables the model to combine complementary information from different imaging modalities for cognitive impairment identification.

The model was evaluated in the discovery cohort and two independent external validation cohorts to assess its performance and generalizability.

The analysis revealed distinct patterns of brain alterations across CKD stages. Structural measures, including cortical thickness and subcortical volume, showed significant stage-related trends, with more advanced CKD stages generally associated with reductions in these measures. Tissue susceptibility also exhibited significant stage-related changes, although the patterns varied across brain regions. In contrast, cerebral blood flow did not show a significant linear trend across CKD stages.

When comparing patients with and without cognitive impairment, significant differences were primarily observed in structural and susceptibility measures, whereas no significant differences in CBF were identified. These findings suggest that structural and susceptibility features may provide more relevant imaging information for distinguishing cognitive status in this cohort.

Based on the selected features, DBFG-Net achieved an area under the receiver operating characteristic curve (AUC) of 0.942 in the discovery cohort. Its performance was also maintained in the two external validation cohorts, with AUC values of 0.902 and 0.926, respectively. These results demonstrate the model's ability to distinguish patients with CKD-related cognitive impairment across different cohorts.

Ablation analyses evaluated the contribution of each branch and architectural component, suggesting that both structural and susceptibility features contributed to model performance. The findings also highlighted the potential relevance of specific cortical and subcortical regions to cognitive function in patients with CKD.

This study presents a stage- and cognition-informed neuroimaging feature-selection framework and integrates it with a dual-branch deep learning network for identifying cognitive impairment in patients with CKD. By selecting features associated with both disease stage and cognitive status, the approach provides a way to incorporate disease-related information into neuroimaging-based model development.

The model achieved consistent performance in the discovery and two external validation cohorts, suggesting that multimodal brain MRI combined with deep learning may offer a promising approach to identifying cognitive impairment in CKD.

Nevertheless, several limitations should be considered. The cross-sectional design cannot capture individual longitudinal trajectories or establish causal relationships between kidney dysfunction, brain alterations, and cognitive impairment. Further prospective studies involving larger and more diverse populations are needed to validate the findings and evaluate the model's potential value in clinical practice.

Overall, this research provides new insights into brain alterations associated with CKD and offers a potential technical framework for the identification of cognitive impairment in this patient population.

Authors of the paper include Wenbo Yang, Lijun Song, Hang Lei, Shiqing Wang, Xintian Ren, Yanfang Tan, Boyan Xu, Zhenghan Yang, Min Li, Weikang Guo, Anfu Zhou, Zhenchang Wang, and Hao Wang.

This work was supported by the National Natural Science Foundation of China (82572190), the National Key Research and Development Program of China (2025YFF0515404), Young Elite Scientists Sponsorship Program by CAST (No. YESS20240057), Beijing Natural Science Foundation (JR26027), Beijing Physician Scientist Training Project (BJPSTP-2025-46), Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China (JYB2025XDXM607), and Beijing Friendship Hospital, Capital Medical University (yyqcjh2023-7).

The paper, "Deep Learning to Identify Cognitive Impairment Across the Full Spectrum of Chronic Kidney Disease: Stage- and Cognition-Informed Feature Selection with a Dual-Branch Feature-Gated Network", was published in the journal Cyborg and Bionic Systems on Nov 6, 2026. The article is available at https://doi.org/10.34133/cbsystems.0733 .

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.