MLP-CKD: a clinically informed deep learning framework for admission laboratory-based screening and risk...

Area of Science:

• Artificial Intelligence in Medicine
• Nephrology
• Critical Care Medicine

Background:

• Advanced renal dysfunction, particularly uremia-associated, poses significant clinical challenges.
• Early screening and risk stratification are crucial for timely intervention and improved patient outcomes.
• Current methods may not fully leverage the predictive power of routine laboratory data.

Purpose of the Study:

• To develop and validate a deep learning framework (MLP-CKD) for screening and risk stratification of uremia-associated advanced renal dysfunction.
• To assess the model's performance against conventional machine learning and rule-based approaches.
• To identify key drivers of model performance, particularly the role of clinically informed feature grouping.

Main Methods:

• Utilized the MIMIC-IV database, extracting laboratory measurements from adult ICU patients within 24 hours of admission.
• Developed MLP-CKD, a multi-branch multilayer perceptron incorporating clinical domain grouping and missingness indicators.
• Compared MLP-CKD against XGBoost, random forest, and an eGFR rule-based baseline using discrimination, calibration, and decision curve analysis.

Main Results:

• MLP-CKD achieved an AUC of 0.913 and AUPRC of 0.848 on an independent test set, outperforming other models.
• Clinically informed feature grouping was identified as the primary driver of performance, with interaction modules offering incremental benefits.
• The model demonstrated superior discrimination compared to XGBoost, random forest, and the eGFR baseline.

Conclusions:

• MLP-CKD offers a clinically structured framework for screening and risk stratification of uremia-associated advanced renal dysfunction.
• The model's strength lies in its clinically informed feature grouping, providing competitive discrimination.
• MLP-CKD serves as a valuable decision-support tool for clinicians, rather than a standalone diagnostic system.