Artificial Intelligence for the Verification of Nasogastric Tube Position in Chest Radiographs: A Systematic Review

Nasogastric tube (NGT) malposition is a significant patient-safety concern, and chest radiography is commonly used to confirm tube position before the tube is used.This systematic review evaluated the diagnostic performance and clinical usefulness of artificial intelligence (AI) systems in assessing NGT or related enteric feeding-tube position on radiographs.PubMed, Embase via Ovid, IEEE Xplore, and the National Health Service (NHS) Knowledge and Library Hub were searched from inception up to September 1, 2026, with additional backward and forward citation searching.Studies were included if they evaluated AIbased assessment of tube position on radiographs, and risk of bias and applicability were assessed using appropriate tools.A total of 131 records were identified, 32 duplicates were removed, and 99 records were screened.Nineteen reports were sought for retrieval, three could not be retrieved, 16 underwent full-text assessment, and 13 studies were included.Most studies were retrospective, and one was a prospective silent NHS validation.Task-specific AI models generally performed well in internal or development datasets, and some external studies reported area under the receiver operating characteristic curve (AUROC) values above 0.90.However, performance was inconsistent across different hospitals and datasets, and several studies were limited by small numbers of unsafe tubes, class imbalance, selected patient groups, and the reuse of the same public datasets or related models.In the prospective NHS study, sensitivity for unsafe NGT position was 86.5% and negative predictive value (NPV) was 98%, although false-negative cases and AI processing failures still occurred.General-purpose multimodal large language models (LLMs) performed poorly, with sensitivity for abnormal NGT position ranging from 0% to 18.4%.Overall, AI may be useful for triage and clinical decision support, but current evidence does not support autonomous NGT verification.Further prospective, independent, multicentre studies in representative clinical populations are needed.