An interpretable machine learning model to predict postoperative acute kidney injury following pancreatic surgery

Technical Summary: An Interpretable Machine Learning Model to Predict Postoperative Acute Kidney Injury Following Pancreatic Surgery

Problem Statement
Post-pancreatectomy acute kidney injury (AKI) is a severe and prevalent complication associated with increased mortality, prolonged ICU stays, and elevated healthcare costs. Current clinical diagnosis relies on serum creatinine changes, which are often delayed and insensitive due to perioperative interventions like fluid resuscitation and diuretic use. Existing prediction models face significant limitations: they frequently rely on traditional logistic regression with limited ability to capture complex non-linear relationships, utilize small or broadly defined cohorts, and often function as "black boxes" lacking interpretability, which hinders clinical trust and adoption. Furthermore, many models underutilize dynamic intraoperative data.

Methodology
This study employed a retrospective cohort design involving 2,493 adult patients who underwent elective pancreatectomy at Shanghai Changhai Hospital between January 2023 and August 2025.

• Data Processing: Patients were excluded if they had preoperative renal replacement therapy dependence, severe preoperative renal impairment (eGFR < 15), emergency surgery, or missing data required for AKI diagnosis. AKI was defined using KDIGO criteria (serum creatinine increase ≥1.5x baseline within 7 days or ≥0.3 mg/dL within 48 hours).
• Feature Selection: From a comprehensive set of baseline, intraoperative, and laboratory variables, LASSO regression with 10-fold cross-validation was used to select the most predictive features, addressing multicollinearity and reducing dimensionality.
• Model Construction: Seven machine learning algorithms were developed and compared: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), XGBoost, LightGBM, and Artificial Neural Network (ANN). The dataset was split into training (70%) and testing (30%) sets. SMOTE oversampling was applied to the training set to address class imbalance. Hyperparameters were tuned using five-fold cross-validation and grid search.
• Evaluation: Models were assessed using discrimination (AUC, ROC), calibration (Brier score, calibration curves), and clinical utility (Decision Curve Analysis).
• Interpretability: The best-performing model was interpreted using SHAP (SHapley Additive exPlanations) to quantify feature importance, identify global and local impacts, and visualize non-linear relationships and threshold effects.

Key Results

• Incidence: Postoperative AKI occurred in 291 patients (11.67%).
• Predictors: LASSO regression identified five key predictors: age, hemoglobin, albumin, cystatin C (CysC), and intraoperative blood loss.
• Model Performance: The Random Forest (RF) model demonstrated the most balanced and clinically relevant performance. It achieved a test AUC of 0.780 (95% CI: 0.729–0.826), with a specificity of 0.849 and sensitivity of 0.506. While Logistic Regression showed a slightly higher AUC (0.793), its extremely low sensitivity (0.138) rendered it less clinically useful for identifying at-risk patients.
• Interpretability (SHAP):
  • Intraoperative blood loss was identified as the strongest risk driver.
  • Cystatin C and intraoperative blood loss exhibited clear positive associations with AKI risk, with threshold effects observed at CysC > 1.0 mg/L and blood loss > 500–600 mL.
  • Albumin and hemoglobin acted as protective factors, with risk increasing when albumin < 35 g/L and hemoglobin < 100–110 g/L.
  • Age showed a positive correlation with risk, particularly increasing after age 60.

Significance and Claims
The authors claim this study presents the first machine learning prediction model specifically targeting post-pancreatectomy AKI that combines high predictive accuracy with clinical interpretability.

• Clinical Utility: By moving beyond "black box" models, the study provides transparent explanations of risk factors via SHAP, enabling clinicians to understand the decision-making basis. This facilitates early risk identification and informed, personalized clinical decision-making.
• Actionable Insights: The identification of specific thresholds (e.g., blood loss > 600 mL, CysC > 1.0 mg/L) offers actionable criteria for early risk stratification. This allows for targeted interventions, such as goal-directed fluid therapy and avoidance of nephrotoxic drugs, for high-risk patients.
• Methodological Contribution: The study demonstrates that ensemble algorithms like Random Forest, when paired with explainable AI techniques, outperform traditional scoring systems in capturing complex perioperative interactions.

Limitations
The authors acknowledge the study's limitations, including its single-center retrospective design, which may introduce selection bias and requires multi-center external validation for generalizability. Additionally, while core variables were included, unmeasured confounders (e.g., detailed fluid management specifics) may exist, and the model's sensitivity (50.6%) indicates room for improvement. Future work is needed to validate the tool in prospective settings and assess its real-world impact on clinical outcomes.