CLFC: Contrastive Learning with Feature Concatenation Framework for Chronic Kidney Disease

Abstract

Patients with thyroid dysfunction have a high likelihood of underdiagnosis of Chronic Kidney Disease (CKD) and Acute Kidney Injury (AKI) because the relationship between endocrine, renal, and metabolic biomarkers are complex and nonlinear. Conventional risk predictive models are unable to model these heterogeneous relationships, which leads to delayed nephroprotective interventions. To overcome this hurdle, this paper suggests the Contrastive Learning with Feature Concatenation (CLFC) model on the stratification of CKD/AKI disease at early and precise stages during the management of thyroid disease patients. The model uses modality-specific encoders, which learn latent representations using thyroid, renal, and metabolic data, but a self-supervised contrastive learning module uses the NT-Xent loss to ensure consistency of the representations. The trained multimodal embeddings are then combined with late-stage concatenation of features and trained with the help of supervised classification. Empirical analysis of a multimodal clinical dataset proves that the presented solution performs much better than the traditional machine learning and deep learning baselines, with higher accuracy, F1-score, and AUC. The importance of contrastive learning, and renal biomarkers in heightening the risk discrimination is further supported by the ablation studies. The suggested framework provides a clinically interpretable scalable robust solution to the assessment of CKD/AKI risk that can be used to make early nephroprotective decisions in thyroid-impacted populations.