Development and validation of an interpretable machine learning model for predicting systemic inflammatory response...

Area of Science:

• Urology
• Nephrology
• Medical Informatics

Background:

• Systemic Inflammatory Response Syndrome (SIRS) is a frequent complication post-percutaneous nephrolithotomy (PCNL).
• Current risk prediction models for SIRS lack generalizability and interpretability.
• Timely intervention requires accurate perioperative risk stratification.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting SIRS after PCNL.
• To identify key perioperative predictors of postoperative SIRS.
• To improve clinical decision-making for infectious complications.

Main Methods:

• Multicenter retrospective cohort study (2,684 patients).
• Feature selection using LASSO regression and Boruta algorithm.
• Development and comparison of seven machine learning models, with AUPRC as the primary metric.
• Validation using internal and external cohorts, with calibration and SHAP analysis for interpretability.

Main Results:

• Six consistent predictors identified: stone size, urine nitrite, urine culture, operative time, residual stone, and neutrophil-to-albumin ratio.
• Random forest model demonstrated balanced performance (AUPRC 0.581-0.641, AUROC 0.873-0.920).
• SHAP analysis confirmed clinically relevant, non-linear predictor associations; an online tool was created.

Conclusions:

• An interpretable ML model using routine perioperative data reliably predicts SIRS post-PCNL.
• This model facilitates early risk stratification and timely clinical decisions.
• Further prospective and multi-regional validation is recommended.