Background: Limited data availability impedes the advancements of artificial intelligence (AI) applications in surgery to date. Federated learning (FL), a novel privacy-focused machine learning technique, introduces a decentralized framework to facilitate multicenter modeling. This study constructs a FL network, including differential privacy (DP), to predict 90-day mortality following colorectal surgery. Methods: Patients undergoing elective colorectal surgery across three tertiary centers (C1–C3) between January 2015 and December 2021 were retrospectively enrolled. Neural networks (NN) for mortality prediction were trained and validated for all three centers individually and after data aggregation using centralized and distributed FL analysis. Local and central DP was then applied as additional data protection framework components. Areas under the receiver operating characteristic (AUROC) and precision-recallcurves (AUPRC) including 95%-confidence intervals were calculated. Results: A total of 2959 patients (mean [SD] age: 56.9 [16.8] years; n = 1677 [56.7%] male) were enrolled. The 90-day mortality rate was 3.1% (n = 92). Local NNs achieved AUROCs of 0.80 ([0.74–0.87]; C1), 0.81 ([0.75–0.87]; C2), and 0.84 ([0.75–0.92]; C3) and corresponding AUPRCs reaching 0.26 [0.11–0.41], 0.31 [0.21–0.41], and 0.21 [0.11–0.32], respectively. An aggregated centralized NN (NN _CZ ) achieved an AUROC of 0.81 [0.76–0.85] and AUPRC of 0.29 [0.22–0.36]. Distributed FL (NN _FL ) was comparable (AUROC: 0.78 [0.72–0.84], P = 0.67; AUPRC: 0.26 [0.17–0.35], P = 0.44) to the centralized model. Central DP reduced the performance of FL-based prediction by 5% in AUROC (0.74 [0.64–0.84]) and 35% in AUPRC (0.17 [0.10–0.23]). Local DP almost diminished the performance (AUROC: 0.52 [0.49–0.55]; AUPRC: 0.05 [0.03–0.06]). Feature importance analysis revealed age, blood status and the Charlson Comorbidity Index as highest weighted features for both NN _CZ and NN _FL . Conclusion: Federated learning demonstrates similar performance to centralized machine learning in preoperative mortality prediction, providing an encouraging framework to accelerate the development of future surgical AI applications. Improving data privacy through DP is associated with compromises in model performance.

Background: Machine Learning (ML) is increasingly being adopted in biomedical research, however, its potential for outcome prediction in visceral surgery remains uncertain. This study compares the potential of ML methods for preoperative 90-day mortality (90DM) prediction of an aggregated multi-organ approach to conventional scoring systems and individual organ models. Methods: This retrospective cohort study enrolled patients undergoing major elective visceral surgery between 2014 and 2022 across two tertiary centers. Multiple ML models for preoperative 90DM prediction were trained, externally validated and benchmarked against the American Society of Anesthesiologists (ASA) score and revised Charlson Comorbidity Index (rCCI). Areas under the receiver operating characteristic (AUROC) and precision recall curves (AUPRC) including standard deviations were calculated. Additionally, individual models for esophageal, gastric, intestinal, liver, and pancreatic surgery were developed and compared to an aggregated approach. Results: 7711 cases encompassing 78 features were included. Overall 90DM was 4% (n = 309). An XBoost classifier demonstrated the best performance and high robustness following external validation (AUROC: 0.86 [0.01]; AUPRC: 0.2 [0.04]). All models outperformed the ASA score (AUROC: 0.72; AUPRC: 0.08) and rCCI (AUROC: 0.81; AUPRC: 0.11). rCCI, patient age and C-reactive protein emerged as most decisive model weights. Models for gastric (AUROC: 0.88 [0.13]; AUPRC: 0.24 [0.26]) and intestinal surgery (AUROC: 0.87 [0.05]; AUPRC: 0.17 [0.09]) revealed the highest organ-specific performances, while pancreatic surgery yielded the lowest results (AUROC: 0.66 [0.08]; AUPRC: 0.22 [0.12]). A combined multi-organ approach (AUROC: 0.84 [0.04]; AUPRC: 0.21 [0.06]) demonstrated superiority over the weighted average across all organ-specific models (AUROC: 0.82 [0.07]; AUPRC: 0.2 [0.13]). Conclusion: ML offers robust preoperative risk stratification for 90DM in elective visceral surgery. Leveraging training across multi-organ cohorts may improve accuracy and robustness compared to organ-specific models. Prospective studies are needed to confirm the potential of ML in surgical outcome prediction.

Risk prediction prior to oncologic esophagectomy is crucial for assisting surgeons and patients in their joint informed decision making. Recently, a new risk prediction model for 90-day mortality after esophagectomy using the International Esodata Study Group (IESG) database was proposed, allowing for the preoperative assignment of patients into different risk categories. However, given the non-linear dependencies between patient- and tumor-related risk factors contributing to cumulative surgical risk, machine learning (ML) may evolve as a novel and more integrated approach for mortality prediction. We evaluated the IESG risk model and compared its performance to ML models. Multiple classifiers were trained and validated on 552 patients from two independent centers undergoing oncologic esophagectomies. The discrimination performance of each model was assessed utilizing the area under the receiver operating characteristics curve (AUROC), the area under the precision–recall curve (AUPRC), and the Matthews correlation coefficient (MCC). The 90-day mortality rate was 5.8%. We found that IESG categorization allowed for adequate group-based risk prediction. However, ML models provided better discrimination performance, reaching superior AUROCs (0.64 [0.63–0.65] vs. 0.44 [0.32–0.56]), AUPRCs (0.25 [0.24–0.27] vs. 0.11 [0.05–0.21]), and MCCs (0.27 ([0.25–0.28] vs. 0.15 [0.03–0.27]). Conclusively, ML shows promising potential to identify patients at risk prior to surgery, surpassing conventional statistics. Still, larger datasets are needed to achieve higher discrimination performances for large-scale clinical implementation in the future.

Federated learning (FL) has emerged as a promising solution to deal with the privacy concerns which often limit access to data for training machine learning (ML) models. It encompasses the exchange of model parameters between data owners and a central model aggregation instance without the need to share sensitive data directly. Although FL does not require sharing sensitive information, it may not guarantee complete privacy, as model parameters can potentially reveal sensitive details about the underlying training data. To enhance privacy, FL approaches can additionally integrate differential privacy (DP), which entails the careful addition of noise to the model. In this paper, we investigate the relationship between FL, DP, and highly non-independent and identically distributed (non-IID) data. We present a specific real-life healthcare application of predicting patient mortality and revision surgery after visceral operations. Our contributions encompass assessing the performance of FL with and without DP in the presence of highly non-IID data, analysing the fairness for individual training participants and investigating the efficacy of personalisation to alleviate the privacy-utility trade-off introduced by DP. Our findings indicate that the addition of DP into FL given non-IID datasets disparately favours participants with larger datasets, while the model has limited utility for those with less data. The addition of a subsequent model fine-tuning step can remedy this disparity and enables all participants to reach a model performance comparable to the centralised training scenario.

Background: Comprehensive preoperative risk stratification is essential for improving perioperative outcomes and guiding informed decisions in general surgery (GS). However, data scarcity remains a key challenge to developing robust, high-dimensional artificial intelligence (AI) models. To address this data barrier in surgical AI, transfer learning (TL) enables neural networks (NNs) to transfer and adapt knowledge from pretrained source models to new domains with critically limited data availability. Methods: This multicenter study included patients undergoing advanced GS at three tertiary centers between 2015 and 2023. Multiple large-scale source models for 90-day mortality prediction were trained on 85 preoperative parameters. Subsequently, organ-specific fine-tuning was performed for esophageal, liver, pancreatic, and colorectal surgery individually. TL models were benchmarked against standard ML models and conventional risk scores using the area under the receiver-operating characteristic curve (AUROC), precision–recall curve (AUPRC), and F1-score, including 95% confidence intervals. Feature analyses were performed for each NN to investigate and compare model interpretability. Results: 14 922 patients (mean [SD] age: 58.5 [16.1] years) were included. Conventional ML achieved AUROCs of 0.75 (0.72–0.79; esophageal surgery), 0.80 (0.79–0.82; liver surgery), 0.73 (0.71–0.76; pancreatic surgery), and 0.92 (0.92–0.92; colorectal surgery) with corresponding AUPRCs reaching 0.37 (0.33–0.43), 0.30 (0.29–0.31), 0.29 (0.24–0.34), and 0.57 (0.56–0.58), respectively. TL significantly improved AUPRCs by 38% in esophageal (0.54 [0.51–0.58], P < 0.001), 14% in liver (0.34 [0.32–0.36], P < 0.001), and 8% in pancreatic surgery (0.31 [0.28–0.37], P < 0.001). Patient age and the Charlson Comorbidity Index (CCI) consistently emerged as the highest-weight features across all TL models. All NNs outperformed the American Society of Anesthesiologists Physical Status and CCI as conventional risk scores in predicting mortality. Conclusions: Machine learning outperforms conventional risk modeling in preoperative mortality prediction. TL can significantly enhance model performance in surgical domains with limited data availability, offering a promising approach to overcome persisting data constraints for AI in surgery.