Machine learning based prediction of recurrence in oral tongue cancer: a systematic review with quantitative synthesis.

PURPOSE: Oral tongue squamous cell carcinoma (OTSCC) is characterized by aggressive local invasion and a high risk of cervical nodal metastasis and mortality. Earlier detection of recurrent OTSCC is associated with improved survival. This systematic review with quantitative synthesis aimed to evaluate the performance of machine learning (ML) models in predicting recurrence in OTSCC. METHODS: This review was conducted in accordance with the PRISMA 2020 guidelines. Studies published between August 2019 and July 2025 that used ML models to predict recurrence in OTSCC were included. Five eligible retrospective cohort and registry-based studies using various ML models were identified. Extracted performance metrics included the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. RESULTS: ML-based models showed generally favorable predictive performance for OTSCC recurrence. Reported AUCs ranged from 0.69 to 0.97 in exploratory quantitative synthesis. Model accuracies ranged from 68% to 95%, sensitivities from 47% to 94%, and specificities from 79% to 96%. Multimodal models that integrated whole-slide imaging (WSI) features with clinicopathological variables, particularly depth of invasion, achieved the highest reported performance across the included studies. CONCLUSION: ML models show promising predictive performance for recurrence in OTSCC, particularly in early-stage disease. Multimodal models that combine WSI with clinicopathological data appear to improve predictive accuracy. Although ML models are not a substitute for clinical expertise, they are rapidly evolving into valuable adjunctive tools for recurrence risk stratification. Prospective validation and external testing across diverse populations are warranted before routine clinical implementation.