Explainable machine learning for preoperative prediction of acute kidney injury after intertrochanteric fracture surgery in older adults.

BACKGROUND

Acute kidney injury (AKI) is an important postoperative complication in older patients undergoing surgery for intertrochanteric femoral fracture. This study aimed to develop and internally evaluate an explainable model for preoperative AKI risk assessment.

METHODS

This retrospective study included 616 patients aged ≥60 years, including 54 (8.8%) who developed postoperative AKI. Candidate predictors were processed and selected using development-cohort data, and seven machine-learning algorithms were compared. Model performance was assessed across 100 repeated stratified 70/30 partitions and by 500-repetition model-specific and complete-pipeline bootstrap analyses. Calibration, decision-curve analysis, and SHAP were used to evaluate and interpret the selected model.

RESULTS

The reported Naive Bayes model comprised baseline serum creatinine, blood urea nitrogen, chloride, uric acid, and chronic kidney disease or renal insufficiency. Across 100 repeated partitions, its mean validation AUC was 0.7231 (SD, 0.0260). The optimism-corrected AUC was 0.8004 for the fixed five-predictor Naive Bayes specification and 0.756 when the complete development pipeline was repeated; the latter analysis yielded a corrected Brier score of 0.084, calibration intercept of 0.02, and calibration slope of 1.70. In the original algorithm-comparison cohort, the observed AUC was 0.795 (95% CI, 0.698-0.892). When the first serum creatinine measurement obtained after admission was used as the alternative reference value, the number of AKI events decreased from 54 to 43 and the validation AUC was 0.706 (95% CI, 0.586-0.826).

CONCLUSION

The model showed moderate internal discrimination across repeated partitions. Model-specific and complete-pipeline bootstrap analyses provided complementary estimates of internal optimism. The model may complement routine preoperative renal assessment while its calibration and transportability are evaluated in additional cohorts.