Diagnostic Efficacy of Artificial Intelligence Models for Predicting Endodontic Outcome – A Systematic Review and Meta-Analysis

This systematic review was conducted to evaluate the diagnostic ability of artificial intelligence (AI) models for predicting an endodontic radiographically inferred condition. Review was performed in accordance to PRISMA-DTA checklist and registered under PROSPERO (CRD42025631782). Databases were searched from January 2000 to December 2024 for studies comparing the diagnostic ability of AI models compared to dental specialists. Risk of bias (ROB) assessment was done through QUADAS (Quality assessment of diagnostic accuracy studies)-2 tool and meta-analysis was performed in Meta-Disc 1.4 software and Review Manager 5.3 for pooled sensitivity, specificity, and summary receiver operating characteristics (SROCs). Five studies were included for analysis. Included studies revealed the presence of moderate to low ROB. Various AI models analysed and evaluated as an index test were artificial neural network, convolutional neural network, direct learning, and direct learning network. Meta-analysis revealed a pooled sensitivity of 0.83 (95% confidence interval (CI) 0.31-1.00) and a pooled specificity of 0.33 (95% CI 0.03-0.81); the summary receiver operating characteristics (SROC) through area under curve (AUC) was 0.54. The included AI models were trained and evaluated on radiographic data only; therefore, findings reflect diagnostic accuracy of image-based AI in detecting radiographic signs associated with endodontic disease rather than comprehensive clinical prognoses. While AI demonstrated moderate sensitivity for identifying these endodontic conditions, low specificity indicates a high false-positive rate when used as a standalone radiograph-based tool. These models may serve as adjunctive screening aids but require prospective validation that integrates clinical and treatment variables before they can be used to predict longitudinal treatment outcomes.