Risk assessment of oral leukoplakia by DNA content enhanced by machine learning models

Abstract

The difficulty in risk assessing oral leukoplakia (OL) for malignant transformation has recently been reduced as novel strategies based on DNA content measurements have been providing consistent and reproducible predictive values, with a major advantage of using archived material. Regrettably, most such approaches are based on costly equipment and human resources that are not widely available, especially in populations with limited access to dedicated technology. The aim of this study was to investigate DNA content as a predictive marker of malignant transformation adapting novel image-based cytometry advances to a conventional flow cytometry context. Nuclei isolation was performed enzymatically on thick sections from paraffin embedded tissue from 97 cases, 18 that progressed to oral carcinoma and 79 that did not. Flow cytometry was used to establish DNA content based on propidium iodide fluorescent labeling of nuclear suspensions. Multiple logistic regression was used to establish DNA content thresholds for DNA content parameters (G1, S-phase, G2, 4cER) to facilitate risk classification criteria. The predictive values of each marker were calculated from Kaplan-Meier and the Log-rank tests (p <0.05). Random Forest (RF) models were developed to enhance predictive values. The most useful DNA content parameters were 4cER fraction and G1/G2 ratio. A 3-step algorithm using these two parameters to classify lesions into low risk and high risk resulted in 52% positive predictive value (PPV) and 93% negative predictive value (NPV). The final RF model increased PPV to 96% with high sensitivity and specificity. DNA content by flow cytometry is a very accessible method, reaching a PPV of 96% in high-risk lesions and 93% NPV in low-risk lesions, with high sensitivity/specificity.