Machine Learning Models Predict Gangrenous Cholecystitis Before Surgery

Because its early symptoms overlap with those of ordinary gallbladder inflammation, surgeons often discover it only during the operation or after pathological examination of the removed organ. A research team at Tongren Hospital, Shanghai Jiao Tong University School of Medicine, has now built a set of interpretable machine learning models that use routine preoperative data to flag these high-risk patients before the scalpel touches the abdomen.

The study, published in BMC Medical Informatics and Decision Making, was a single-center retrospective cohort analysis of 487 patients who underwent cholecystectomy between February 2020 and July 2023. All participants met the Tokyo Guidelines 2018 diagnostic criteria for acute cholecystitis, and the final diagnosis of gangrenous change was confirmed by postoperative pathology. Of the 487 patients, 94, or 19.3 percent, had gangrenous cholecystitis, while the remaining 393, or 80.7 percent, did not. Every laboratory test and imaging study included in the analysis was obtained within 48 hours before surgery, and all operations were completed within 48 hours of admission.

The researchers deliberately excluded patients who had undergone percutaneous transhepatic gallbladder drainage, because no pathological specimen could be obtained from them. This group tends to include patients too unstable for general anesthesia or those who already have gallbladder gangrene or perforation, and the authors acknowledged that excluding them may introduce selection bias. Other exclusion criteria covered gallbladder carcinoma, elective surgery after biliary interventions, severe hepatic dysfunction, pancreatitis, acute cholangitis, and incomplete clinical records.

The preoperative dataset was broad. It included age, sex, maximum body temperature, duration of symptoms, comorbidities and a peritonitis score based on tenderness, rebound tenderness and muscular guarding. Laboratory measurements spanned routine blood tests, coagulation profiles, procalcitonin and biochemical indicators. A radiologist with more than five years of clinical experience, blinded to patient grouping and study protocol, independently measured gallbladder dimensions, wall thickness, pericholecystic effusion and stone incarceration on ultrasound and computed tomography scans. Features missing in more than 20 percent of patients were dropped, remaining gaps were filled by multiple imputation, and continuous variables were standardized by Z-score normalization.

Variable selection proceeded in stages. Independent two-sample t-tests compared preoperative characteristics between the gangrenous and non-gangrenous groups, and univariate logistic regression quantified candidate risk effects. The team then applied Least Absolute Shrinkage and Selection Operator, or LASSO, regression, which uses an L1 penalty to compress correlated variables and retain only those with meaningful independent contributions. Five predictors survived this filter: preoperative C-reactive protein, gallbladder diameter, pericholecystic exudation, serum albumin and basophil percentage. Notably, variables such as prothrombin time-international normalized ratio, glucose, alkaline phosphatase and activated partial thromboplastin time, despite showing group-level differences, received near-zero LASSO coefficients and were set aside as redundant.

The baseline differences were striking. Mean preoperative C-reactive protein was 151.73 milligrams per liter in the gangrenous group versus 57.82 milligrams per liter in the non-gangrenous group. Mean gallbladder diameter was 5.73 centimeters versus 4.57 centimeters, and pericholecystic effusion was far more common among gangrenous cases. Serum albumin was lower in the gangrenous group, 42.58 grams per liter versus 46.54 grams per liter, although this difference did not reach conventional statistical significance in the reported comparison. Basophil percentage was markedly lower in gangrenous patients, a finding the authors themselves treat cautiously, describing it as hypothesis-generating rather than evidence of a causal role, since basophils are not dominant players in acute surgical inflammation.

Nine classification algorithms were then trained on the LASSO-selected features: logistic regression, random forest, XGBoost, LightGBM, support vector machine, artificial neural network, k-nearest neighbors, linear discriminant analysis and a Stacking ensemble. Hyperparameters were tuned by grid search with five-fold cross-validation inside an 80:20 training-test split, and the held-out test set was never used during training or tuning. On that independent test set, XGBoost achieved the highest area under the receiver operating characteristic curve at 0.90, followed by random forest at 0.89 and LightGBM at 0.88. A conventional logistic regression nomogram built on the same five variables reached an AUC of 0.84 with an accuracy of 84.35 percent, a precision of 82.66 percent and an F1-score of 0.82.

Statistical testing reinforced the ranking. Pairwise DeLong tests on the test-set AUC values showed that XGBoost significantly outperformed several classical models, including k-nearest neighbors, support vector machine, artificial neural network and linear discriminant analysis, with absolute Z statistics exceeding 2.0 in most comparisons. The two ensemble leaders differed in secondary metrics: random forest showed slightly greater stability in accuracy and specificity, which the authors suggest suits surgical decision-making where false positives carry costs, while XGBoost offered marginally better recall, at 0.54, and F1-score, at 0.50, favoring early detection scenarios. Decision curve analysis indicated that XGBoost, random forest and Stacking delivered net clinical benefit across a broad range of threshold probabilities, outperforming both treat-all and treat-none strategies, and their calibration curves tracked the 45-degree reference line more closely than those of the shallower learners.

Interpretability came from SHAP, or SHapley Additive exPlanations, values, which assign each feature an additive contribution to every individual prediction. For both XGBoost and random forest, C-reactive protein emerged as the most influential positive predictor, with higher values consistently raising predicted risk, and gallbladder diameter ranked alongside it. The presence of pericholecystic effusion acted as a clear risk-enhancing binary signal. Albumin showed a predominantly negative contribution, consistent with hypoalbuminemia marking poor nutritional reserve, while basophil percentage displayed inverse or sample-dependent effects that the authors attribute to possible nonlinear interactions or dataset-specific patterns rather than established biology.

The authors are explicit about the limits of their work. This was a retrospective, single-center study with a modest number of gangrenous cases, so selection bias and overfitting cannot be fully excluded, and all performance figures come from internal validation at one institution. External validation in multicenter prospective cohorts will be needed before the framework can be considered generalizable. Within those boundaries, the study offers a transparent, non-invasive preoperative risk tool built entirely from variables already collected in routine care, potentially helping surgeons prioritize early intervention and intensified monitoring for patients whose inflamed gallbladders are quietly turning gangrenous.

Subject of Research: Machine learning prediction of gangrenous cholecystitis from preoperative clinical variables

Article Title: Development of a multimodel predictive framework for gangrenous cholecystitis with SHAP-based interpretability

Article References: Niu, Q.-G., Yu, Y., Xu, F., Kong, W.-Q., Li, J.-Y., Tang, D.-F., & Wang, Y. (2026). Development of a multimodel predictive framework for gangrenous cholecystitis with SHAP-based interpretability. BMC Medical Informatics and Decision Making, 26(1), Article 434. https://doi.org/10.1186/s12911-026-03705-9

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03705-9

Keywords: gangrenous cholecystitis, machine learning, XGBoost, random forest, SHAP, LASSO regression, C-reactive protein, preoperative prediction, acute cholecystitis, retrospective cohort study, decision curve analysis, clinical decision support