Development and Evaluation of a Urinary Na/K Ratio Prediction Model: A Systematic Comparison from Attention-Based...

Area of Science:

• Biomedical Informatics
• Cardiovascular Health
• Machine Learning in Medicine

Background:

• The urinary sodium-to-potassium (Na/K) ratio is a key biomarker for predicting blood pressure and cardiovascular disease risk.
• Accurate estimation of the Na/K ratio is crucial for clinical risk assessment.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for estimating the urinary Na/K ratio.
• To compare the performance of classical ML models against complex deep learning architectures using physiological variables.

Main Methods:

• Utilized physiological data (body weight, blood pressure, pulse rate) from 82 participants.
• Developed an attention-based deep learning model (MIDIP) and compared it with classical ML models (Random Forest, Gradient Boosting, Linear Regression) and ensemble strategies.
• Employed a nested cross-validation framework for rigorous generalization assessment.

Main Results:

• Simple averaging of Random Forest, Gradient Boosting, and Linear Regression (Group A) yielded the best performance (MAE = 1.756, RMSE = 2.349, R² = 0.390).
• Complex models, including an attention-based Transformer, either degraded performance or showed instability due to overfitting in the small sample size.
• Adaptive weighting strategies did not improve prediction accuracy.

Conclusions:

• Model simplicity and equal-weight ensemble methods offer superior generalization for urinary Na/K ratio prediction in small datasets (N=82).
• Overly complex deep learning architectures are prone to overfitting with limited data.
• Emphasizes the critical role of strict validation and controlled model complexity for clinical ML applications.