ABSTRACT Background This study aimed to develop an explainable machine-learning framework based on three predefined analytical endpoints: treatment allocation, complication risk, and TG18 severity discrimination. Methods One hundred twenty-one patients diagnosed using Tokyo Guidelines 2018 criteria underwent medical treatment, laparoscopic cholecystectomy, conversion from laparoscopy to open surgery, and percutaneous transhepatic cholecystostomy. The analyses included ANOVA, Random Forest classification, logistic regression, and ROC–AUC evaluation. Relationships between the model and clinical outcomes were examined through complications and hospitalization durations. Results Most patients were GRADE I (43.0%)/GRADE II (53.7%), only 3.3% being GRADE III. Neutrophil (F = 14.82, p < 0.001), CRP (F = 12.91, p < 0.001), monocyte (F = 10.15, p < 0.001), and albumin (F = 3.47, p = 0.0184) levels, wall thickness at computed tomography (F = 9.86, p < 0.001), and age (F = 5.01, p = 0.0027) exhibited significant correlations with treatment modalities. The most effective predictors of treatment decisions were neutrophil count (29.4%), CRP (24.3%), and monocyte (9.5%). Medical treatment and conversion to open surgery were predicted with high accuracy. The complication rate was 77.8% in the conversion to open surgery group, while no complications were observed in the medical or drainage groups. Hospitalization was significantly longer in patients developing complications (11.8 vs. 6.5 days, p < 0.001). Conclusions This model converting multidimensional clinical data into algorithmic form may provide meaningful predictive insight for individualized treatment strategies and early identification of complication risk. These findings suggest that the framework may have exploratory decision-support value; however, prospective validation and assessment of clinical impact are required before its implementation in clinical practice.
Dadük et al. (Mon,) studied this question.