Predicting Hospitalization in Emergency Department Patients with Acute Abdominal Pain Using AI Analysis of Supine Abdominal Radiographs

This study aimed to develop and validate a deep learning model capable of predicting hospitalization in emergency department (ED) patients presenting with acute, non-traumatic abdominal pain using only supine abdominal radiographs (AXRs). In this retrospective diagnostic-prediction study conducted at a tertiary ED, 1,011 patients with a triage diagnosis of unspecified abdominal pain and a corresponding supine AXR were included. After standardized preprocessing—including histogram normalization, CLAHE enhancement, and resizing to 224×224 pixels—the dataset was split into training (708), validation (101), and independent test (202) subsets using patient-level stratified sampling. A transfer-learning approach with an EfficientNet-B0 backbone and five-fold cross-validation was employed, and final predictions were generated using a probability-averaged ensemble. Model performance on the independent test set was assessed using standard discrimination metrics, predictive values, and calibration analysis. Hospi-talization prevalence in the test set was approximately 74%. The model achieved an accuracy of 64.8%, sensitivity 72.3%, specificity 49.1%, precision (PPV) 82.1%, and negative predictive value 34.7%. Discrimination was moderate, with an AUROC of 0.74 and AUPRC of 0.79. Calibration analysis showed slight overestimation of hospitalization risk at intermediate probability levels (0.6–0.8), with a Brier score of 0.21. Performance was stable across cross-validation folds. Despite the diagnostic limitations of supine AXRs, the deep learning model extracted meaningful prognostic signals associated with ED hospitalization. The system functioned primarily as a high-precision “rule-in” adjunct, reliably identifying patients likely to require admission, although specificity and negative predictive value remained modest. These findings suggest that AXRs contain underutilized prognostic information, supporting future development of multimodal triage tools integrating imaging with structured clinical data.