New world vs old world : refining oral cancer screening, diagnosis, and prediction

Abstract

Oral cancer is the most common head and neck malignancy worldwide with a 50% five-year survival rate. Early detection of tumors remains the most important factor to improve disease prognosis. Oral cancer is amenable to prevention and screening with a natural history that may involve oral potentially malignant disorders (OPMDs). However, the current practice of oral cancer screening and diagnosis needs to be refined to ensure better survival. Moreover, while the management of oral cancer has improved with the advent of multidisciplinary and multimodal strategies for aggressive tumors, the timely selection of patients for such therapies still poses a challenge. Hence, in this project, a series of investigations were conducted to refine the present workflow and decision-making process in oral cancer screening and management using contemporary innovative approaches. Seven studies were performed following a comprehensive overview and four systematically conducted reviews and meta-analyses. The first study investigated the use of hierarchical Bayesian spatial modeling in identifying the population at risk of oral cancer. Further, two studies assessed and compared the use of artificial intelligence (AI) and statistical risk-scoring techniques in the identification of candidates for oral cancer screening. Then, the use of AI techniques for predicting malignant transformation as discrete and dynamic outcomes in three common OPMDs (oral leukoplakia, oral lichen planus, and oral lichenoid lesions) was investigated in two studies. The sixth study investigated the use of AI techniques for methylome biomarker discovery in salivary DNA following reduced representation bisulfite sequencing (RRBS). Finally, the seventh study compared time-to-event AI methods in predicting disease-specific survival and overall survival of patients with oral cancer. The first study proved that hierarchical Bayesian spatial methods can be used successfully to identify at-risk populations for oral cancer screening which can be visualized easily as choropleth maps. Results from the second and third studies found that an explainable ensemble AI-based model improved the identification of candidates for oral cancer screening and outperformed statistical risk-scoring or crude methods of oral cancer risk stratification following multiple evaluations. Additionally, the fourth and fifth studies successfully developed scalable AI-based models for malignant transformation prediction in oral leukoplakia and oral lichenoid mucositis that outperformed the dysplasia grading system and is applicable in settings irrespective of differences in data standards. The sixth study suggested that RRBS analysis of salivary DNA can yield potential biomarkers for oral cancer and that AI techniques can potentially operationalize salivary DNA methylation biomarker discovery while the seventh study showed that continuous-time AI-based models can satisfactorily forecast disease-specific and overall survival of oral cancer patients and may be useful in selecting candidates for advanced treatment. In conclusion, this study series has shown that innovative methods such as spatiotemporal mapping and AI outperform could be successfully applied in oral oncology and outperform the current methods of decision-support in oral cancer management when available. Overall, these innovative methods could potentially refine the screening, diagnosis, and prediction of oral cancer.