Federated learning for early severity prediction in acute pancreatitis: a multi-center study.

backgroundAcute pancreatitis (AP) is a common gastrointestinal disease prone to severe systemic complications, presenting with abdominal pain, nausea, and vomiting. Diagnosis depends on serum markers and imaging. Categorized by the Atlanta classification, about 20% of cases progress to more severe forms. Due to its rapid progression and high misdiagnosis rate, early and accurate assessment is crucial. In this study, we aim to construct a deep neural network to provide a scientific basis for the clinical management of AP.

methodsA total of 1,884 patients diagnosed with AP from four hospitals, namely the Fourth Affiliated Hospital of Zhejiang University School of Medicine(ZJU-H4), the First Affiliated Hospital of Ningbo University(NBU-H1), the Second Affiliated Hospital of Zhejiang Chinese Medical University(ZCMU-H2), and Lishui People’s Hospital(LISHUI-H), were retrospectively included from 2012 to 2024. The examination indicators of the patients within 48 h and 72 h after admission were collected. Clinical models were developed using Arya, a novel privacy computing platform by Healink to predict the severity of AP in patients. Two learning methods—logistic regression (LR) and deep neural networks (DNN)—were applied within the federated learning framework. Moreover, to address the challenge of fair contribution evaluation in federated learning, a novel assessment method integrating data quality and model performance impact was proposed.

resultsIn the 48-hour scenario, the DNN model showed higher prediction performance than the logistic regression model in terms of sensitivity (78.20% vs. 78.70%), specificity (78.40% vs. 72.90%) and accuracy (78.30% vs. 75.90%); the same trend was observed in the 72-hour scenario: sensitivity (78.10% vs. 83.30%), specificity (80.90% vs. 70.60%) and accuracy (79.40% vs. 77.10%). Moreover, the model based on data within 72 h was more accurate than that within 48 h. The optimal model among the four scenarios was the DNN - based model using indicators within 72 h after admission.

conclusionThis study successfully integrated collective knowledge and data resources from various medical institutions while ensuring patient privacy and data security through the adoption of federated learning. It demonstrated a significant leap in the prediction of AP severity using machine learning. Furthermore, by introducing a novel contribution evaluation mechanism, it addressed the limitations of traditional data-sharing methods, enhancing fairness and incentivizing multi-center participation.