# Predicting Malignant Transformation in Oral Leukoplakia: A Multilayer Perceptron Approach Incorporating Clinicopathological Features and DNA Content

## TL;DR

This study uses machine learning to predict the risk of oral leukoplakia turning into cancer by combining clinical and DNA data.

## Contribution

A multilayer perceptron model integrating DNA content and clinicopathological features for predicting malignant transformation in oral leukoplakia.

## Key findings

- 4cER, G2 phase, dysplasia grading, and inflammatory infiltrate were significant predictors of malignant transformation.
- The model achieved 72% sensitivity, 96% specificity, and an AUC of 85.4% in predicting malignant risk.
- High-risk cases predicted by the model showed significantly poorer outcomes in survival analysis.

## Abstract

Oral leukoplakia (OL) is a potentially malignant disorder of the oral mucosa. Accurate prediction of malignant transformation (MT) remains a clinical challenge. This study aimed to develop and evaluate a machine learning model that integrates histopathological, demographic, and DNA content features to predict MT risk in OL.

We conducted a retrospective cohort study of 97 OL cases—18 with confirmed MT and 79 non‐transformed controls—selected from a larger series. Each case included clinicopathological features, and DNA content data obtained by flow cytometry for cell cycle phases (G1, S‐phase, G2 and excess DNA beyond the tetraploid region [4cER]). All cases had a minimum 5‐year follow‐up or histologically confirmed transformation. A multilayer perceptron (MLP) model was trained on 27 features. Stratified five‐fold cross‐validation and minority class oversampling (positive filling) were used to improve learning and mitigate data imbalance. Performance was evaluated using accuracy, sensitivity, specificity, F1‐score, AUC, and Kaplan–Meier survival analysis.

Significant predictors of MT included 4cER (p = 0.005), G2 phase (p = 0.04), dysplasia grading (p = 0.003), and inflammatory infiltrate (p = 0.01). The optimized model yielded 72% sensitivity, 96% specificity, and an AUC of 85.4%. Survival analysis showed significantly poorer outcomes in the high‐risk cases predicted by the model (p < 0.0001).

Integrating DNA content analysis with machine learning provides an objective and clinically useful model to stratify malignant risk in OL, complementing conventional histopathology and supporting personalized patient management.