Given the limited clinical data available for surgically treated infected pancreatic necrosis, researchers sought to develop interpretable, procedure-specific machine learning models to predict 90-day post-necrosectomy mortality and characterize the specific risk patterns linked to each surgical approach.
The retrospective investigation analyzed 143 patients treated at Ruijin Hospital after propensity-score matching on age and sex and the subsequent exclusion of one patient in the minimally invasive surgery group. The study population comprised 96 patients who received open surgical necrosectomy between 2014 and 2018, alongside 47 patients managed with video-assisted retroperitoneal debridement between 2019 and 2024. The primary clinical outcome tracked across both cohorts was all-cause mortality recorded within 90 days following the initial necrosectomy procedure.
To capture procedure-dependent risk variations, investigators built separate CatBoost machine learning models utilizing preoperative clinical measurements, laboratory parameters, and diagnostic imaging variables. The internal model validation protocol integrated repeated feature selection under fixed CatBoost parameter settings together with an additional ten-repetition nested cross-validation analysis, ensuring that both hyperparameter tuning and feature selection remained strictly confined within each outer training set.
Model calibration and conditional bootstrap intervals were subsequently calculated using stored predictions generated from the earlier fixed-parameter analysis.
During the additional nested cross-validation assessment, the open surgical necrosectomy model achieved a median out-of-fold area under the receiver operating characteristic curve of 0.810, displaying a repetition range of 0.773 to 0.833, alongside a median Brier score of 0.149. In parallel, the video-assisted retroperitoneal debridement model yielded a median out-of-fold area under the receiver operating characteristic curve of 0.726, with across-repetition values spanning 0.670 to 0.788, accompanied by a median Brier score of 0.100.
Analysis of feature contributions demonstrated notable differences between the surgical strategies. In the selected open surgical necrosectomy model, platelet count, body mass index, and preoperative procalcitonin levels emerged as the most prominent contributors to mortality prediction. Conversely, in the video-assisted retroperitoneal debridement model, preoperative procalcitonin as well as renal and respiratory status served as the primary predictive determinants.
These dominant risk factors were repeatedly selected across the resampling iterations, demonstrating stable procedure-specific prognostic configurations.
The study authors concluded that interpretable procedure-specific machine learning models successfully identify heterogeneous prognostic patterns following open surgical necrosectomy and video-assisted retroperitoneal debridement. These exploratory findings establish an empirical foundation for developing and prospectively validating personalized surgical management algorithms in patients diagnosed with infected pancreatic necrosis.
A major impetus for developing dedicated machine learning architectures was the recognized scarcity of clinical datasets for surgically managed infected pancreatic necrosis.
Subject of Research: Procedure-specific machine learning mortality prediction in infected pancreatic necrosis
Article Title: Interpretable machine learning for procedure-specific mortality prediction after open surgical necrosectomy or video-assisted retroperitoneal debridement in infected pancreatic necrosis
Article References: Li, C., Zhang, H., Wang, G., Dai, D., Pan, C., Mao, E., Wei, X., Zhao, B., Zhang, H., & Shen, M. (2026). Interpretable machine learning for procedure-specific mortality prediction after open surgical necrosectomy or video-assisted retroperitoneal debridement in infected pancreatic necrosis. Frontiers in Medicine, 13, Article 1932030. https://doi.org/10.3389/fmed.2026.1932030
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1932030
Keywords: infected pancreatic necrosis, open surgical necrosectomy, video-assisted retroperitoneal debridement, machine learning, CatBoost, mortality prediction, prognostic patterns, SHAP analysis, procalcitonin, platelet count, body mass index, renal status, respiratory status, AUROC, Brier score, Ruijin Hospital
News Source: Teresa Odom. (November 20, 2026). Interpretable Machine Learning Models Predict Procedure-Specific Mortality in Infected Pancreatic Necrosis. Scienmag.



