Detecting Oral Cancer Using Tabular Deep Learning

Area of Science:

• Biomedicine
• Artificial Intelligence
• Machine Learning

Background:

• Oral cancer has a low survival rate, necessitating early detection for effective treatment.
• Current diagnostic methods primarily rely on image analysis, leaving potential in tabular clinical data unexplored.
• Deep learning applications for oral cancer prediction using structured text data are understudied.

Purpose of the Study:

• To investigate the efficacy of deep learning models in predicting cancerous versus precancerous oral lesions using tabular clinical data.
• To compare the performance of two deep learning methods against a conventional algorithm for oral cancer risk assessment.
• To analyze the interpretability of the models and identify key predictive features.

Main Methods:

• Utilized a subset of 1791 patients from an ongoing oral cancer study, focusing on structured text (tabular) clinical data.
• Implemented and compared two distinct deep learning architectures designed for tabular data.
• Benchmarked deep learning models against a traditional machine learning algorithm for classification accuracy.

Main Results:

• All evaluated models demonstrated strong predictive performance on a hold-out test set, achieving a Youden index greater than 0.6 and an Area Under the Curve (AUC) greater than 0.9.
• Model interpretability analysis revealed that lesion characteristics are critical factors in predicting oral cancer risk.
• The study confirms the potential of AI/ML in analyzing clinical tabular data for oral cancer detection.

Conclusions:

• Deep learning approaches applied to tabular clinical data show significant promise for the early detection of oral cancer.
• Lesion characteristics are identified as key indicators for differentiating between cancerous and precancerous oral lesions.
• This research contributes valuable insights for applying AI/ML in biomedical applications, particularly in oral oncology.