Artificial Intelligence in the Diagnosis of Esophageal Diseases: A Systematic Review With Meta-Analysis

Abstract

Background: For the evaluation of upper gastrointestinal disorders, artificial intelligence (AI) tools have been applied in various settings, including during endoscopy, and patient interviews (questionnaires), with encouraging results. We conducted a systematic review with meta-analysis to evaluate the performance of AI in the diagnosis of both malignant and benign esophageal diseases. Methods: PubMed, MEDLINE, EMBASE, EMBASE Classic, and the Cochrane Library (from inception to 30th March 2021) were searched to identify observational studies that reported the performance of AI in the diagnosis of malignant or benign esophageal diseases. Data were extracted from published study reports, and pooled accuracy, sensitivity, specificity, and 95% confidence intervals were calculated. Findings: The literature search resulted in 2568 studies. All studies were screened, and 67 full articles were assessed. Forty-two studies reported the performance of AI in the diagnosis of various esophageal diseases and were included in the qualitative synthesis. Nineteen studies reported extractable data and were included in the meta-analysis. For the diagnosis of Barrett’s neoplasia, AI (with white light) accuracy, sensitivity, and specificity were 89.6, 89.5, and 90.0%, respectively, and 91.0, 88.6, and 92.8%, respectively, using AI with narrow band imaging (NBI). For the diagnosis of esophageal squamous cell carcinoma, AI systems achieved accuracy, sensitivity, and specificity of 92.7, 97.6, and 87.8% with WLE, and 93.8, 95.2, and 95.1% with NBI, respectively. In the detection of abnormal interpapillary capillary loops, AI showed up to 97.8% accuracy, 98.8% sensitivity, and 96.7% specificity. For the diagnosis of GERD exclusively based on symptoms, AI showed accuracy, sensitivity, and specificity of up to 99.4, 99.2, and 99.7%, respectively. Overall, in the diagnosis of included benign and malignant esophageal diseases, AI achieved accuracy, sensitivity, and specificity of 92.9%,94.2%, and 92.5%, respectively. Interpretation: Our results demonstrated high performance scores of AI systems in the clinical and endoscopic diagnosis of esophageal diseases. The support of AI tools might virtually overcome the variability in the diagnostic performance due to different levels of expertise among clinicians, ultimately improving the standard of care. Funding Information: None. Declaration of Interests: Pierfrancesco Visaggi: none. Brigida Barberio: none. Cesare Hassan: reports personal fees from Medtronic, Fujifilm, Olympus, outside the submitted work. Prateek Sharma: Consultant Medtronic, Olympus, Boston Scientific, Fujifilm and Lumendi; Grant Support: Ironwood, Erbe, Docbot, Cosmo pharmaceuticals and CDx labs, outside the submitted work. Edoardo Savarino: has received lecture or consultancy fees from Abbvie, Alfasigma, Amgen, Aurora Pharma, Bristol-Myers Squibb, EG Stada Group, Fresenius Kabi, Grifols, Janssen, Johnson&Johnson, Innovamedica, Malesci, Medtronic, Merck & Co, Reckitt Benckiser, Sandoz, Shire, SILA, Sofar, Takeda, Unifarco, outside the submitted work. Nicola de Bortoli: has received lecture or consultancy fees from Malesci and Reckitt Benckiser, outside the submitted work.