An Interpretable Machine Learning Survival Model for Predicting Long-term Kidney Outcomes in IgA Nephropathy

Area of Science:

• Nephrology
• Data Science
• Biostatistics

Background:

• IgA nephropathy (IgAN) is a prevalent kidney disease with varied clinical presentations.
• Accurate prediction of long-term patient outcomes is crucial for effective clinical management.
• Standard survival analysis methods can be biased by right-censored data common in long-term follow-up studies.

Purpose of the Study:

• To develop and validate an accurate survival model for predicting IgA nephropathy patient prognosis.
• To address the challenge of biased risk estimation caused by right-censored data in IgAN cohorts.
• To provide an interpretable model for understanding patient-specific risk factors and their impact on renal progression.

Main Methods:

• Construction of a survival model using EXtreme Gradient Boosting for survival (XSBoost-Surv).
• Incorporation of time-to-event data into the modeling process to handle censored observations.
• Application of Shapley Additive exPlanations (SHAP) for model interpretability and analysis of predictor-outcome relationships.

Main Results:

• The developed XSBoost-Surv model demonstrated superior discrimination performance compared to conventional survival methods.
• The model accurately predicted long-term outcomes for IgA nephropathy patients.
• SHAP analysis provided insights into individual predictions and non-linear predictor effects.

Conclusions:

• XSBoost-Surv offers an accurate and explainable approach to predicting IgAN patient prognosis.
• The model's ability to handle censored data and provide interpretable results enhances clinical understanding of renal progression.
• This tool can potentially benefit therapeutic strategies and clinical decision-making for IgA nephropathy.