Pancreatic fluid collections (PFCs) are associated with high healthcare utilization and multiple interventions. However, clinicians still lack reliable baseline risk-stratification tools, leading to reactive decision-making rather than proactive risk-adapted care[1]. We therefore developed a machine-learning algorithm to predict step-up therapy requirements at index EUS and compared its performance with established clinical predictors.

Data on EUS-guided drainage of consecutive PFCs performed between 2016 and 2024 across three endoscopy units at Charité Berlin, irrespective of etiology, were retrospectively collected. The primary outcome was need for step-up therapy, defined as endoscopic necrosectomy, percutaneous drainage, or laparoscopic/open surgery performed after the index procedure to achieve clinical success. An extreme gradient boosting (XGBoost) machine learning algorithm was developed and compared with the established QNI score (Quadrant, Necrosis, Infection) to predict need for step-up therapy[2,3]. The cohort was randomly split into training (70%) and validation (30%) sets. Performance was assessed using 10-fold cross-validation, with discrimination quantified by area under the receiver operating characteristic curve. Of 163 evaluated patient-level features, 20 were selected for the final decision-support model based on Gini coefficients, clinical relevance, and practical considerations to preserve predictive performance while minimizing overfitting.

Among 674 consecutive patients, 486 were included in the final analysis (mean age 56.1±14.3 years; 68.9% male). Overall, 199 patients (40.9%) required step-up therapy, including endoscopic necrosectomy (n=145; 29.8%), percutaneous drainage (n=62; 12.8%), multigate technique (n=22; 4.5%), and laparoscopic or open surgery (n=35; 7.2%). Compared with patients managed without step-up therapy, those requiring escalation had significantly larger PFCs (105±46 vs. 67±32 mm; p<0.001), higher prevalence of >60% necrosis (11.6% vs. 2.4%; p<0.001), elevated CRP levels (median 176 mg/L [IQR 111–216] vs. 100 mg/L [IQR 55–173]; p<0.001), increased infection rates (71.3% vs. 44.3%; p=0.002), and more organ failure (31.2% vs. 8.0%; p<0.001). Patients classified as high-risk by the QNI score (n=125; 25.7%) were more likely to undergo step-up therapy than low-risk patients (65.6% vs. 32.4%; p<0.001). Extent of necrosis was the strongest predictor of step-up therapy (OR 3.9, 95% CI 2.9-5.3; p<0.001). The XGBoost model showed superior discriminative performance compared to the QNI score (AUC 0.86 vs. 0.77; p<0.01) (Table), correctly identifying an additional 12.7% of patients needing step-up therapy while reducing false-positive predictions by 10.8%.

The machine learning algorithm demonstrates superior discriminative ability for predicting step-up therapy compared with the QNI score and enhances risk stratification at the index procedure.