An improved domain-adversarial network for predicting hemodialysis adequacy

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Nephrology

Background:

• Hemodialysis (HD) is crucial for end-stage renal disease (ESRD) patients.
• Current real-time monitoring of dialysis adequacy (Kt/V) is often expensive and complex.
• There is a need for accurate, low-cost, and versatile monitoring solutions.

Purpose of the Study:

• To develop a novel, accurate, and low-cost Kt/V prediction method for hemodialysis.
• To enhance the intelligence and efficiency of clinical hemodialysis treatment.
• To create a method compatible with diverse dialysis devices.

Main Methods:

• An improved Domain Adversarial Neural Network (DANN) framework was proposed.
• Feature extraction combined Long Short-Term Memory (LSTM) with Kolmogorov-Arnold Network (KAN) using B-spline activations.
• A multi-head attention mechanism was integrated for temporal dependency capture.

Main Results:

• The proposed model demonstrated high prediction accuracy for Kt/V.
• It significantly outperformed conventional deep learning models (RNN, LSTM, DANN).
• The model effectively mitigated data distribution shifts due to patient and device variability.

Conclusions:

• The novel DANN-based model enables precise, non-invasive estimation of dialysis adequacy (Kt/V).
• It shows excellent generalization and potential for clinical application.
• This offers a practical solution for real-time adequacy assessment, especially for lower-end dialysis devices.