Temporal dynamics and predictive modeling of oral epithelial dysplasia features during carcinogenesis.

Abstract

OBJECTIVE: To investigate the temporal evolution and predictive value of individual histopathological features of oral epithelial dysplasia (OED) during carcinogenesis, and to evaluate the diagnostic utility of optical fluorescence imaging in a murine model. DESIGN: A 4-nitroquinoline 1-oxide (4NQO) mouse model was used to induce oral squamous cell carcinoma (OSCC). Histological evaluation of 20 architectural and cytological OED features was performed at seven time points over 24 weeks. Data were analyzed using descriptive statistics, Spearman's correlation, logistic regression, and principal component analysis (PCA). A random forest model was developed to classify malignant transformation and evaluated using F1-score, cross-validation, ROC, and precision-recall curves. Optical fluorescence imaging was assessed for early lesion detection. RESULTS: Cumulative feature burden, particularly involving loss of stratification, reverse polarity, nuclear atypia, and mitotic activity, was more predictive of transformation than any single feature. PCA revealed two major axes-structural disorganization and proliferative instability. The random forest model achieved high predictive performance (F1 = 0.88, ROC-AUC = 0.97, PR-AUC = 0.98). Autofluorescence failed to improve early detection. CONCLUSION: Feature accumulation is a robust predictor of OSCC risk in dysplastic lesions. Histological quantification combined with machine learning offers potential for improved prognostic modeling in oral precancer.