Diagnostic Accuracy of Artificial Intelligence - Based Models for Oral Potentially Malignant Disorders Compared to Histopathological Investigation - A Systematic Review and Meta-analysis
Mrinal V. Shete, Anagha V. Shete · Journal of Head & Neck Physicians and Surgeons · 2025
ABSTRACT Objective: The objective is to evaluate the diagnostic ability of artificial intelligence (AI)-based models for oral potentially malignant disorders compared to histopathological investigation. Methodology: The review was performed in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analysis–Diagnostic Test Accuracy checklist, and the review protocol is registered under PROSPERO (CRD42024592176). Databases were searched from January 2000 to April 2024 to identify the diagnostic potential of AI-based tools and models. True-positive, false-positive, true-negative, false-negative, sensitivity, and specificity values were extracted or calculated if not present for each study. The quality of selected studies was evaluated based on the quality assessment of diagnostic accuracy studies-2 tool. Meta-analysis was performed in Meta-Disc 1.4 software and Review Manager 5.3 using a bivariate model parameter for the sensitivity and specificity, and summary points, summary receiver operating curve, confidence region, and area under curve (AUC) were calculated. Results: Fifteen studies were included for qualitative synthesis and for meta-analysis. Included studies had the presence of low to moderate risk of bias. Sensitivity and specificity were calculated with AUC. Meta-analysis showed a pooled sensitivity and specificity of 0.71 (confidence interval [CI] 0.43–0.91) and 0.31 (CI 0.05–0.73) respectively, with a pooled positive likelihood ratio 0.82 (0.56–1.32) and negative likelihood ratio of 11.14 (0.26–5.03) was observed with diagnostic odd’s ratio of 0.63 (0.08–5.20) and overall accuracy (AUC) being 0.52 suggesting that the overall diagnostic accuracy of AI-based tools being moderate to good in diagnosing the desired condition. Conclusion: AI-based models are a valid tool and overall have good diagnostic potentiality in diagnosing the target condition and can be used as an alternative adjunct to histopathology. AI-based models can be undertaken for early diagnosis and prompt treatment under secondary level of prevention.