Artificial intelligence for predicting post-excision recurrence and malignant progression in oral potentially malignant disorders: a retrospective cohort study.

BACKGROUND

Oral cancer may develop from precursor lesions and conditions termed oral potentially malignant disorders (OPMD), although not all patients progress to cancer in their lifetime. Managing patients with OPMD is challenging due to lesion recurrence and uncertain malignant progression risk following surgical excision. This study developed a multitask AI-based model to predict the risk of treatment failure, malignant progression, and recurrence among patients with OPMD treated by surgery.

METHODS

This study utilized multidimensional data from 366 retrospective patients with OPMD treated in 2 tertiary centers to construct an AI model to predict 3 treatment outcomes among patients with OPMD. Multifaceted prognostic variables were collected for the cohort and used to train four AI supervised learning models, followed by optimal model selection. Area under the receiver operating characteristic curve (AUC) and Brier scores (BSs) were used to assess model performance. External testing of the model was also performed, and metrics were compared to the WHO and binary dysplasia grading systems (current standards). We further assessed the net benefit and explainability of the final multitask model.

RESULTS

The outperforming model (TabPFN) had good AUC values of 0.829 (0.729-0.929), 0.912 (0.836-0.988), and 0.791 (0.683-0.899) for predicting treatment failure, malignant progression, and lesion recurrence at external testing. The BSs of the model for all three treatment outcome predictive tasks were also optimal (0.085-0.147). Furthermore, the AI model had a superior net benefit than the WHO and binary epithelial dysplasia grading systems in assessing the need for close monitoring among patients with OPMD treated by surgery. The explainability of the model was also successfully implemented.

CONCLUSIONS

The multitask AI-based model developed with multidimensional data has good discriminatory performance, calibration, and net benefit, showing potential for comprehensive risk assessment and clinical decision support in the surgical management of patients with OPMD to promote early detection of oral cancer.