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

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Data Fusion Techniques

Background:

• Chronic Kidney Disease (CKD) poses a global health challenge requiring accurate staging for effective management.
• Machine learning (ML) shows potential for CKD prediction, but faces challenges with incomplete or biased datasets.
• Current ML approaches often involve a trade-off between small, complete datasets and larger, imputed (biased) ones.

Purpose of the Study:

• To develop a novel ML framework for improved CKD classification by maximizing sample utility.
• To minimize reliance on data imputation by fusing models trained on different data sources.
• To enhance the generalizability and accuracy of CKD staging models, especially when complete datasets are scarce.

Main Methods:

• Fusion of models trained on distinct data sources (e.g., blood, urine, demographics).
• Integration of intermediate model representations with global representations from a Mixed Model.
• Utilization of cross-attention and self-attention mechanisms for enhanced feature integration.

Main Results:

• Improved accuracy in CKD staging compared to traditional methods.
• Enhanced model generalizability across different data subsets.
• Demonstrated effectiveness in clinical decision support scenarios with limited complete data.

Conclusions:

• The proposed model fusion framework effectively addresses data scarcity and imputation challenges in CKD classification.
• This approach offers a robust solution for clinical decision support, leveraging available partial datasets.
• The method shows significant promise for advancing ML applications in nephrology and personalized medicine.