Single and multiple time-point prediction models in kidney transplant outcomes

Area of Science:

• Nephrology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Kidney transplantation is a vital treatment for end-stage renal disease.
• Accurate prediction of graft and recipient survival is essential for clinical decision-making and patient management.
• Existing prediction models vary in their approach to time-dependent factors and survival analysis.

Purpose of the Study:

• To compare the predictive performance of regression models and artificial neural networks (ANNs) for kidney transplant outcomes.
• To evaluate single time-point versus multiple time-point modeling strategies.
• To identify key factors influencing model accuracy and calibration.

Main Methods:

• Utilized the USRDS dataset for a large-scale retrospective analysis.
• Employed logistic regression and single-output ANNs for single time-point predictions.
• Applied Cox models and multiple-output ANNs for multiple time-point predictions.
• Assessed model performance using discrimination (AUC) and calibration metrics.

Main Results:

• Both regression models and ANNs achieved good prediction discrimination (AUC up to 0.82) and calibration.
• Single and multiple time-point models showed comparable AUC, with exceptions for multiple-output ANNs with high censoring.
• Logistic regression performed similarly to ANNs without significant non-linear relationships or interactions.
• Time-varying effects and appropriate baseline survivor functions are critical for Cox model accuracy.

Conclusions:

• Regression models and ANNs offer viable approaches for predicting kidney transplant survival.
• Model choice depends on data characteristics, including censoring and predictor relationships.
• Accurate modeling of time-varying effects and baseline survival is paramount for reliable clinical decision support.