Diagnostic Performance of Artificial Intelligence-Based Models for the Detection of Early Esophageal Cancers in Barret’s Esophagus: A Meta-Analysis of Patient-Based Studies

Introduction Barret’s esophagus (BE) is a precursor of adenocarcinoma of the esophagus. The detection of high-grade dysplasia and adenocarcinoma at an early stage can improve survival but is very challenging. Artificial intelligence (AI)-based models have been claimed to improve diagnostic accuracy. The aim of the current study was to carry out a meta-analysis of papers reporting the results of artificial intelligence-based models used in real-time white light endoscopy of patients with BE to detect early esophageal adenocarcinoma (EEAC). Methods This meta-analysis was registered with the International Prospective Register of Systematic Reviews (PROSPERO; Reg No. CRD42021246148) and its conduction and reporting followed the Preferred Reporting Items for Systematic Review and Meta-Analysis of Diagnostic Test Accuracy (PRISMA-DTA) statement guidelines. All peer-reviewed and preprint original articles that reported the sensitivity and specificity of AI-based models on white light endoscopic imaging as an index test against the standard criterion of histologically proven early oesophageal cancer on the background of Barret's esophagus reported as per-patient analysis were considered for inclusion. There was no restriction on type and year of publication, however, articles published in the English language were searched. The search engines used included Medline, PubMed, EMBASE, EMCARE, AMED, BNI, and HMIC. The search strategy included the following keywords for all search engines: ("Esophageal Cancer" OR "Esophageal Neoplasms" OR " Oesophageal Cancer" OR "Oesophageal Neoplasms” OR "Barrett's Esophagus" OR "Barrett's Oesophagus") And ("Artificial Intelligence" OR "Deep Learning" OR "Machine Learning" OR "Convolutional Network"). This search was conducted on November 30, 2020. Duplicate studies were excluded. Studies that reported more than one dataset per patient for the diagnostic accuracy of the AI-based model were included twice. Quantitative and qualitative data, including first author, year of publication, true positives (TP), false negatives (FN), false positives (FP), true negatives (TN), the threshold of the index test, and country where the study was conducted, were extracted using a data extraction sheet. The Quality Appraisal for Diverse Studies 2 (QUADS-2) tool was used to assess the quality of each study. Data were analyzed using MetaDTA, interactive online software for meta-analysis of diagnostic studies. The diagnostic performance of the meta-analysis was assessed by a summary receiver operating characteristics (sROC) plot. A meta-analysis tree was constructed using MetaDTA software to determine the effect of cumulative sensitivity and specificity on surveillance of patients with BE in terms of miss rate and overdiagnosis. Results The literature search revealed 171 relevant records. After removing duplicates, 117 records were screened. Full-text articles of 28 studies were assessed for eligibility. Only three studies reporting four datasets met the inclusion criteria. The summary sensitivity and specificity of AI-based models were 0.90 (95% CI, 0.83- 0.944) and 0.86 (95% CI, 0.781-0.91), respectively. The area under the curve for all the available evidence was 0.88. Conclusion Collective evidence for the routine usage of AI-based models in the detection of EEAC is encouraging but is limited by the low number of studies. Further prospective studies reporting the patient-based diagnostic accuracy of such models are required.