Supporting information

Figure S1: ROC curves of POD local learning models developed using different machine learning methods across four centers. (A) Chinese PLA General Hospital; (B) The Second Xiangya Hospital of Central South University; (C) Union Hospital, Tongji Medical College, Huazhong University of Science and Technology; (D) The Affiliated Hospital of Guizhou Medical University.

Figure S2: Decision curve analysis for all POD prediction models. (A) Internal Validation; (B) External Validation.

Figure S3: Calibration curve for all prediction models of POD. (A) Internal Validation; (B) External Validation.

Figure S4: Feature‐specific SHAP values of federated and centralized prediction models for POD across four development centers. (A‐C) Chinese PLA General Hospital; (D‐F) The Second Xiangya Hospital of Central South University; (G‐I) Union Hospital, Tongji Medical College, Huazhong University of Science and Technology; and (J‐L) The Affiliated Hospital of Guizhou Medical University. Within each center, the three panels correspond to the federated aggregation strategies FedAvg, FedLSD, and FedProx, respectively. (M) SHAP summary plot of the centralized learning model (CLM) trained on the pooled data from all four development centers. The x‐axis shows SHAP values, which represent the contribution of each feature to the model output, and the color scale indicates feature values from low to high. All medication‐related variables shown in the figure refer to intraoperative medications.

Figure S5: ROC curves of AKI local learning models developed using different machine learning methods across four centers. (A) Chinese PLA General Hospital; (B) The Second Xiangya Hospital of Central South University; (C) Union Hospital, Tongji Medical College, Huazhong University of Science and Technology; (D) The Affiliated Hospital of Guizhou Medical University.

Figure S6: Decision curve analysis for all AKI prediction models. (A) Internal Validation; (B) External Validation.

Figure S7: Calibration curve for all prediction models of AKI. (A) Internal Validation; (B) External Validation.

Figure S8: Feature‐specific SHAP values of federated and centralized prediction models for AKI across four development centers. (A‐C) Chinese PLA General Hospital; (D‐F) The Second Xiangya Hospital of Central South University; (G‐I) Union Hospital, Tongji Medical College, Huazhong University of Science and Technology; and (J‐L) The Affiliated Hospital of Guizhou Medical University. Within each center, the three panels correspond to the federated aggregation strategies FedAvg, FedLSD, and FedProx, respectively. (M) SHAP summary plot of the centralized learning model (CLM) trained on the pooled data from all four development centers. The x‐axis shows SHAP values, which represent the contribution of each feature to the model output, and the color scale indicates feature values from low to high. All medication‐related variables shown in the figure refer to intraoperative medications.

Table S1: Distribution of variables across the five participating centers.

Table S2: Participant characteristics when building POD model.

Table S3: Participant characteristics when building AKI model.

Table S4: Performance of POD Federated Learning Models in 10‐fold Cross‐validation.

Table S5: Calibration performance of centralized and federated learning models for POD.

Table S6: Validation performance of sensitivity analysis models for POD prediction.

Table S7: Performance of AKI Federated Learning Models in 10‐fold Cross‐validation.

Table S8: Calibration performance of centralized and federated learning models for AKI.

Table S9: Validation performance of sensitivity analysis models for AKI prediction.