Advanced artificial intelligence vs simpler models for 1-year death prediction among patients receiving hemodialysis

Area of Science:

• Nephrology
• Artificial Intelligence
• Data Science

Background:

• Predicting short-term mortality in hemodialysis patients is crucial for clinical decision-making.
• Evaluating the performance of AI models against traditional methods is essential for optimizing patient care.

Purpose of the Study:

• To assess the data requirements for advanced AI models to surpass simpler methods in predicting mortality for hemodialysis-dependent kidney failure patients.
• To compare the efficacy of logistic regression, boosting, and transformer models across varying dataset sizes and feature complexities.

Main Methods:

• Trained logistic regression, boosting, and transformer models on datasets ranging from 500 to 490,197 patients.
• Utilized feature sets from last-visit data to full patient trajectories, incorporating temporal information.
• Measured model performance using Area Under the Receiver Operating Characteristic Curve (AUC-ROC) and Area Under the Precision-Recall Curve (AUC-PR).

Main Results:

• Temporal features improved performance across all evaluated models.
• On the largest dataset, transformer models (AUC-ROC = 0.8568) and boosting models (AUC-ROC = 0.8598) demonstrated comparable predictive accuracy.
• Transformer models required significantly larger datasets to outperform boosting models.

Conclusions:

• Advanced AI tools, such as transformers, necessitate substantial data volumes to justify their computational expense over simpler models like boosting.
• The utility of complex AI models is constrained in scenarios with limited data availability, even with datasets as large as 500,000 samples.
• Incorporating complex feature sets, particularly those with temporal information, offers benefits regardless of the model's complexity.