Maximizing Sample Utilization in CKD Classification: Fusion and Alignment of Locally Trained Models with a Global Model.

Chronic Kidney Disease (CKD) is a significant global health issue that requires accurate classification for effective management. While machine learning techniques have shown promise in predicting CKD stages, obtaining comprehensive datasets with all relevant clinical features remains a major challenge. Existing works often face a trade-off between using small, complete datasets or larger, biased ones due to imputation. Additionally, data is often derived from various sources-such as blood and urine samples, along with demographic information-each requiring individual models that may vary in sample size. In this study, we explore this challenge by maximizing sample usage while minimizing reliance on imputation through the fusion of models trained on different data sources. Our approach integrates intermediate representations from these models with global representations from a Mixed Model, trained on a smaller but complete dataset that incorporates attributes from all sources. By leveraging cross-attention and self-attention mechanisms, our method improves staging accuracy and enhances model generalizability. This framework is particularly beneficial for clinical decision support when complete datasets are scarce, yet partial datasets are available, such as with CKD data from the UK SAIL databank.