Deep learning-based prediction of acute pancreatitis severity from abdominal CT with multicenter external validation.

Abstract

Background: Acute pancreatitis (AP) is a common gastrointestinal disease with a rising global incidence. While most cases are mild, severe AP (SAP) carries high mortality. Early and accurate severity prediction facilitates management optimization. However, existing clinical severity prediction models, such as Bedside Index of Severity in Acute Pancreatitis (BISAP) and Modified CT Severity Index (mCTSI), have modest accuracy and often rely on data unavailable at admission.

Purpose: This study proposes a deep learning (DL) model to predict AP severity using abdominal contrast-enhanced CT scans acquired within 24 hours of admission.

Materials and methods: We collected 10 130 studies from 8335 patients across a multi-site U.S. health system and 3488 studies from public datasets. The DL model was trained in 2 stages: (1) self-supervised pretraining on 11 896 unlabeled studies and (2) fine-tuning on 550 labeled studies. Performance was evaluated against mCTSI on a hold-out internal test set (n = 100 patients) and against both mCTSI and BISAP on an external Hungarian AP registry (n = 518 patients).

Results: On the internal test set, the model achieved areas under receiver operating curve (AUROCs) of 0.888 (95% CI: 0.800-0.960, sensitivity 0.50, specificity 0.92) for SAP and 0.888 (95% CI: 0.819-0.946, sensitivity 0.75, specificity 0.85) for mild AP (MAP), outperforming mCTSI (P = .002). External validation showed AUROCs of 0.887 (95% CI: 0.825-0.941, sensitivity 0.73, specificity 0.88) for SAP and 0.858 (95% CI: 0.826-0.888, sensitivity 0.38, specificity 0.98) for MAP, surpassing mCTSI (P = .024) and BISAP (P = .002). In retrospective triage analysis, the model correctly identified 50%-73% of patients who progressed to SAP and 40%-73% of those with MAP.

Conclusion: The proposed DL model achieved performance comparable to or better than established prognostic tools and maintained robust external performance. These findings suggest that AI-assisted CT analysis may support early, automated risk stratification of AP.