Predicting Leukoplakia and Oral Squamous Cell Carcinoma Using Interpretable Machine Learning: A Retrospective Analysis

Abstract

Purpose: The purpose of this study is to assess the effectiveness of the best performing interpretable machine learning models in the diagnoses of leukoplakia and oral squamous cell carcinoma (OSCC). Methods: A total of 237 patient cases were analysed that included information about patient demographics, lesion characteristics, and lifestyle factors, such as age, gender, tobacco use, and lesion size. The dataset was preprocessed and normalised, and then separated into training and testing sets. The following models were tested: K-Nearest Neighbours (KNN), Logistic Regression, Naive Bayes, Support Vector Machine (SVM), and Random Forest. The overall accuracy, Kappa score, class-specific precision, recall, and F1 score were used to assess performance. SHAP (SHapley Additive ExPlanations) was used to interpret the Random Forest model and determine the contribution of each feature to the predictions. Results: The Random Forest model had the best overall accuracy (93%) and Kappa score (0.90). For OSCC, it had a precision of 0.91, a recall of 1.00, and an F1 score of 0.95. The model had a precision of 1.00, recall of 0.78, and F1 score of 0.88 for leukoplakia without dysplasia. The precision for leukoplakia with dysplasia was 0.91, the recall was 1.00, and the F1 score was 0.95. The top three features influencing the prediction of leukoplakia with dysplasia are buccal mucosa localisation, ages greater than 60 years, and larger lesions. For leukoplakia without dysplasia, the key features are gingival localisation, larger lesions, and tongue localisation. In the case of OSCC, gingival localisation, floor-of-mouth localisation, and buccal mucosa localisation are the most influential features. Conclusions: The Random Forest model outperformed the other machine learning models in diagnosing oral cancer and potentially malignant oral lesions with higher accuracy and interpretability. The machine learning models struggled to identify dysplastic changes. Using SHAP improves the understanding of the importance of features, facilitating early diagnosis and possibly reducing mortality rates. The model notably indicated that lesions on the floor of the mouth were highly unlikely to be dysplastic, instead showing one of the highest probabilities for being OSCC.

Keywords

• oral cancer
• white lesion
• Random Forest classifier
• SHAP
• predictive modelling