OA10.6. When to Accept the Transfer: A Clinical Prediction Model for Esophageal Perforation in Pneumomediastinum

Abstract

Benign Disease: Iatrogenic Esophageal Disease, Perforation and Postsurgical Complications Pneumomediastinum is a common indication for inter-hospital transfer driven by concern for esophageal perforation, yet most cases are benign and self-limited. Patients often undergo extensive workups that prove unnecessary. No validated model exists to risk-stratify these patients. We aimed to develop a model identifying perforation among patients transferred with pneumomediastinum. A retrospective review was performed of 281 adult patients transferred to a quaternary referral center with radiographically-confirmed pneumomediastinum from 1/2019-11/2025. The primary outcome was esophageal perforation, defined as perforation identified on endoscopy or managed with intervention (i.e., endoscopic stenting, clip/suture repair, primary repair, esophagectomy). Random forest variable importance analysis was performed on all collected clinical, laboratory, and radiographic variables to identify candidate predictors (Figure). These were then entered into a stepwise logistic regression (removal threshold P≥.20). The final model uses all retained variables to generate an individualized predicted probability of perforation. Notably, variables traditionally associated with esophageal perforation, such as recent esophageal instrumentation and heart rate, demonstrated low variable importance and were not retained in the final model. Model discrimination was assessed by area under the ROC curve (AUC) with bootstrap validation (1,000 resamples). Threshold analysis defined a three-tier risk stratification scheme prioritizing 100% sensitivity in perforation detection. Of 281 patients, 90 (32.0%) had esophageal perforation. CT findings most predictive of perforation were mediastinal fluid collection (PPV 87%, specificity 96%), esophageal defect/irregularity (PPV 79%, sensitivity 57%), and pleural effusion (PPV 74%, sensitivity 70%), confirmed by random forest analysis as top-ranking predictors (Figure). Stepwise logistic regression retained 9 variables: mediastinal fluid collection (aOR 14.3), pleural effusion on CT (aOR 9.7), emesis/retching (aOR 8.4), esophageal defect/irregularity (aOR 5.6), dysphagia (aOR 3.3), pleural effusion on chest radiograph (aOR 2.7), chest pain (aOR 2.7), respiratory rate (aOR 1.10/unit), and age (aOR 1.03/year). The model achieved an AUC of 0.96 (95% CI: 0.94–0.98). Patients lacking all three CT findings had an NPV of 94.9%; however, the model’s low-risk classification achieved an NPV of 100%, capturing remaining perforations through clinical variables. Model-derived thresholds stratified patients into low, intermediate, and high-risk groups with perforation in 0%, 44.2%, and 92.1%, respectively (Table). A 9-variable prediction model incorporating imaging findings, presenting symptoms, and vital signs demonstrated excellent discrimination for esophageal perforation among patients with pneumomediastinum, classifying the majority as either low or high-risk with an NPV of 100% and PPV of 92.1%, respectively. Few patients fell into the intermediate category. This tool could guide clinical decision-making by identifying patients who may safely forego workup, require further evaluation, or need expedited transfer. Video Description