Discussion

Using multicenter data from four development centers, we developed FLMs to predict postoperative delirium and acute kidney injury in older surgical patients. Our principal finding is that FLMs achieved predictive performance comparable to CLMs for both outcomes in a simulated decentralized training setting without direct pooling of raw patient data. More importantly, the full‐variable models showed substantially better external discrimination than the Age + ASA and preoperative‐only benchmark models, indicating meaningful incremental value from broader perioperative information. Compared with LLMs, FLMs showed a consistent trend toward numerically better or comparable discrimination across centers for both outcomes, with a statistically significant improvement for POD prediction in Center 3 by DeLong testing. Overall, our findings support federated learning as a feasible privacy‐preserving strategy for multicenter perioperative prediction modeling.

POD and AKI significantly affect the recovery and quality of life of older patients. Accurate prediction of these complications may facilitate targeted preoperative intervention and perioperative monitoring, thereby potentially reducing their occurrence. Previous researchers, such as Rössler et al. [32] and our team, have shown that ML models can achieve high accuracy in predicting these risks, offering valuable risk stratification for older patients [33]. In a single‐center retrospective cohort study of 111,888 operations, Xue et al. used five ML algorithms to predict postoperative complications, achieving AUCs of 0.848 for AKI and 0.762 for delirium [34]. Although these studies demonstrated the potential of ML in predicting postoperative complications, they also highlighted several limitations inherent to traditional ML approaches. One major drawback is the need for large centralized datasets, which can be difficult to compile owing to data privacy concerns and regulatory restrictions [35]. Most large datasets come from a single center, leading to potential biases. Moreover, the variability in data collection methods across institutions can result in inconsistent data quality, further complicating model development and reducing model generalizability [36]. These challenges underscore the need for innovative approaches that can leverage the collective power of diverse datasets while maintaining data privacy and security.

FL offers key advantages over centralized methods, including enhanced data privacy by keeping patient data local [37, 38]. It also improves model robustness by using diverse data sources, enhancing generalizability and accuracy [39]. Building on these advantages, FL has been successfully applied in various disease prediction tasks, demonstrating its practical potential in real‐world scenarios. Several studies have demonstrated the application of FL in disease prediction. Danek et al. used multi‐omics data from two datasets to predict the onset of Parkinson's disease. The best FL model achieved an AUC‐PR of 87.6%, with a performance difference of less than 2% compared to the CLM [30]. Sheller et al. applied FL to the BraTS 2018 dataset to build a brain tumor segmentation model. The best FLM achieved a Dice coefficient of 0.852, only 0.01 lower than the CLM (0.862). FL not only protects data privacy but also outperforms other collaborative methods (e.g., IIL and CIIL), demonstrating stability even with imbalanced multi‐institutional data distributions [40].

Additionally, FL enables smaller institutions with limited data to access high‐quality models, addressing data imbalance and scarcity effectively [41]. This practical advantage may be particularly important for smaller medical centers or rare‐case settings, where local data are often insufficient for independent model development [27]. In brain tumor segmentation, FL performed well even with as few as six samples from individual institutions [42]. In our study, where some centers provided only a few hundred samples and others thousands, FL showed competitive performance despite marked sample‐size imbalance, supporting its potential adaptability in multicenter settings. This imbalance is common in real‐world multicenter clinical practice and may make our setting more representative of realistic deployment scenarios than an idealized balanced distribution. Under such non‐IID conditions, sample‐size imbalance and center‐specific data distributions may affect centralized and federated learning through different mechanisms. Centralized learning may be influenced by the majority distribution of the pooled data, whereas federated learning may be affected by heterogeneous local updates and sample‐size‐weighted aggregation. In the present study, FLMs nevertheless achieved performance broadly comparable to CLMs under the observed degree of multicenter heterogeneity.

The lower POD incidence in the external validation cohort (5.7% vs. 13.8%), together with the significant differences in baseline and perioperative characteristics between the development and external validation cohorts, suggests the presence of domain shift. These significant between‐dataset differences may partly explain the modest decline in external validation performance, particularly for AKI prediction. This is likely attributable to differences in patient case mix, perioperative management, and institutional practice patterns. In particular, the external validation center is a specialized neuroscience institution, where non‐neurosurgical older patients may represent a more selected population with a different perioperative risk profile from those in the development cohort. Consistent with the lower POD incidence in the external cohort, the negative calibration intercepts indicated systematic overestimation of absolute POD risk, although the calibration patterns remained broadly similar across the centralized and federated models. The relatively small sample size and low POD incidence of the external validation cohort may also have reduced the precision and robustness of the external performance estimates. The limited number of POD events likely increased statistical uncertainty, as reflected by the relatively wide 95% confidence intervals in the external validation results. Therefore, the exact magnitude of model performance in this cohort should be interpreted with appropriate caution. Despite these limitations, the federated models maintained clinically meaningful discrimination in the external cohort, suggesting transportability of their risk‐ranking performance, although local recalibration may be required before clinical implementation.

Post hoc sensitivity and reduced‐variable benchmark analyses further helped clarify the clinical value of the proposed framework. Specifically, the Age + ASA minimal models showed substantially lower discrimination than the full‐variable centralized benchmark models for both POD and AKI, particularly in external validation. The preoperative‐only models retained some predictive information but also performed worse than the full‐variable models. These findings suggest that the observed model performance was not explained by simple baseline clinical characteristics alone and that intraoperative and perioperative variables contributed meaningful incremental predictive value. Accordingly, the practical value of federated learning in this setting lies not merely in marginal gains in AUC, but in its ability to approximate centralized‐model performance without direct data sharing while outperforming simpler baseline models in a multicenter privacy‐constrained environment.

An additional strength of our study is that the SHAP analysis provided insight not only into globally important predictors but also into center‐specific heterogeneity across both POD and AKI models. For POD, age and surgery duration were stable shared predictors across federated and centralized models, whereas for AKI, surgery duration, surgical department, and ASA classification were more prominent. Beyond these common signals, the relative importance of secondary features varied across institutions and generally involved medication‐related variables, physiological or laboratory indicators, and surgery‐related characteristics. From a perioperative perspective, these differences likely reflect variation in patient case mix, operative spectrum, anesthetic and analgesic pathways, and local documentation practices. Although these SHAP patterns should be interpreted as model‐level explanations rather than causal effects, they suggest that the federated models captured both shared risk factors across centers and center‐specific variation under multicenter heterogeneity.

Despite its rigorous design, several limitations remain. First, although we included five centers and applied multiple modeling strategies, model performance remained limited, likely due to suboptimal center‐level performance, distributional differences between training and external validation cohorts, and limited perioperative feature granularity. FL enables privacy‐preserving multicenter collaboration but cannot overcome limitations in data quality or feature completeness. Future studies with larger samples and more detailed perioperative variables may improve performance. Second, the federated learning framework was evaluated as a retrospective simulation rather than a live deployed inter‐hospital network. Centralized feature selection in this simulation does not fully reflect an end‐to‐end federated workflow. In a prospective federated deployment, candidate variables could be harmonized using a prespecified common data dictionary, with locally derived feature statistics or importance scores combined through privacy‐preserving aggregation. The internal validation results should be interpreted as development‐stage evaluations because these datasets informed architecture selection. Therefore, practical implementation issues such as communication latency, system connectivity, asynchronous updating, and site‐level computational variability were not assessed in this study. Third, data heterogeneity could impair the generalizability of federated models, as data in FL frameworks are often non‐IID [43]. Although FedLSD was included as an aggregation strategy designed to address heterogeneity, the imbalanced sample distribution across centers remains a structural limitation but is also common in real‐world clinical practice. This setting realistically reflects current clinical conditions and may have greater practical applicability than an idealized balanced distribution [44]. Future studies should include more balanced multicenter samples to ensure more equitable and robust contributions.

In conclusion, FL holds substantial potential for developing privacy‐preserving prediction models for postoperative complications in older patients. By utilizing decentralized data without direct sharing of raw patient information, it may support collaborative model development across healthcare institutions. This approach may facilitate early risk stratification for complications such as POD and AKI and help guide targeted perioperative monitoring and resource allocation.