Ping Wang, Xingyu Chen, Tianjiao Hao, Jisong Chen
Early retrospective studies show promising discrimination for AI-based GC prediction; however, evidence remains insufficient for routine clinical decision-making due to methodological heterogeneity, limited external validation, absence of prospective impact studies, and lack of calibration and clinical utility analyses. Prospective multicenter validation and implementation research are needed before clinical adoption.
BACKGROUND: Gangrenous cholecystitis (GC) is a severe pathological subtype of acute cholecystitis, yet its preoperative diagnostic rate remains below 10%. Conventional scoring systems based on logistic regression demonstrate limited predictive performance. Artificial intelligence (AI) approaches may offer improved risk stratification, but the evidence base remains limited.
METHODS: A scoping review was conducted across PubMed, Web of Science, Embase, and CNKI databases for publications from January 2020 to June 2026. The search combined MeSH terms and free-text keywords related to gangrenous cholecystitis and artificial intelligence. After screening 412 records, 18 studies were included: 3 direct GC AI prediction studies, 3 traditional GC scoring systems, and 12 indirect or methodologically related studies. A narrative synthesis was adopted given the substantial heterogeneity in study design.
RESULTS: Three studies directly addressed AI-based GC prediction. Machine learning models using structured clinical data achieved validation AUCs of 0.818-0.944, though these were derived from retrospective, single-center or limited multicenter cohorts. Deep learning models integrating non-contrast and contrast-enhanced CT achieved independent validation AUCs of 0.879 and 0.887 (training-set AUC 0.965). Explainable AI methods identified model-associated predictors, including hypokalemia/hyponatremia, though these require pathophysiological validation. No study reported calibration, decision curve analysis, or prospective clinical impact evaluation.
CONCLUSIONS: Early retrospective studies show promising discrimination for AI-based GC prediction; however, evidence remains insufficient for routine clinical decision-making due to methodological heterogeneity, limited external validation, absence of prospective impact studies, and lack of calibration and clinical utility analyses. Prospective multicenter validation and implementation research are needed before clinical adoption.