Artificial Intelligence for Preoperative Prediction of Lymph Node Metastasis and Depth of Invasion in Oral Tongue Squamous Cell Carcinoma: A Systematic Review and Meta-Analysis.

Background

Occult lymph node metastasis (OLNM) and depth of invasion (DOI) are key determinants of elective neck dissection in clinically node-negative oral tongue squamous cell carcinoma (OTSCC), yet accurate preoperative risk stratification remains challenging. This study evaluated the diagnostic performance of artificial intelligence (AI)-based predictive models for OLNM and DOI in OTSCC.

Methods

A systematic review and meta-analysis were conducted in accordance with PRISMA 2020 guidelines. A structured search of PubMed identified twelve eligible studies, nine of which provided extractable 2 × 2 contingency data for inclusion in the primary bivariate meta-analysis. One additional study modeling DOI-derived pT stage was synthesized narratively. Pooled sensitivity and specificity were estimated using a bivariate random-effects model. Heterogeneity, threshold effects, and publication bias (Deeks' test) were assessed. Methodological quality was evaluated using QUADAS-2 supplemented by an AI-specific methodological appraisal.

Results

Across nine studies included in the primary meta-analysis, pooled sensitivity was 0.679 (95% CI: 0.604-0.745) and pooled specificity was 0.762 (95% CI: 0.705-0.811), with a summary AUC of 0.786. Heterogeneity was moderate for sensitivity (I2 = 41.8%) and low for specificity (I2 = 23.4%), with no significant threshold effect (ρ = -0.117, p = 0.776). No significant publication bias was detected (p = 0.596). Subgroup analyses showed comparable performance between OLNM-specific and general LNM models, whereas deep learning or hybrid approaches demonstrated higher accuracy than traditional machine learning methods. Notably, only one out of nine primary studies incorporated true external validation.

Conclusions

AI-based models demonstrate moderate discriminative performance for predicting LNM and DOI in OTSCC and may serve as adjunctive tools in preoperative risk stratification rather than standalone decision-makers. However, the near absence of external validation, limited calibration reporting, and lack of clinician-comparator analyses substantially constrain current clinical translation. Future research should prioritize multi-center prospective validation, systematic calibration and decision-curve analyses, and adherence to TRIPOD-AI and CLAIM reporting standards.