Prediction of loss of heterozygosity in oral cavity dysplasia through vascular pattern

Background and purpose: Loss of Heterozygosity (LOH) is a key genetic alteration associated with progression of oral cavity dysplasia to oral squamous cell carcinoma, yet non-invasive methods to predict LOH status are lacking. The aim of this study is to investigate whether vascular pattern abnormalities detected through Narrow Band Imaging (NBI) can predict LOH status in oral dysplasia using machine learning-based quantitative analysis. Methods: This was a retrospective analysis of prospectively collected data from the IMPEDE multicenter clinical trial (NCT04504552). The study sample included 31 patients with oral potentially malignant disorders (13 LOH- positive, 18 LOH-negative) and 125 region-of-interest (ROI) images. Predictor variables were quantitative vascular morphology features: vessel length, number of intersections, tortuosity, branching angles, fractal dimensions, which were extracted from NBI images using the Jerman Vesselness Filter and pvbm library. Main outcome variable was LOH status (positive versus negative) by genetic testing per EPOC trial criteria. Patient demographics, smoking status, alcohol consumption were the covariates. A Support Vector Machine (SVM) classifier with class balancing and probability calibration was trained using Leave-One-Group-Out Cross-Validation (LOGO-CV) at the patient level. Results: Of 31 patients (median age 65 years; 19 males, 12 females), 13 (42%) were LOH-positive. Significant differences in vascular morphology were observed between LOH-positive and LOH-negative samples. The SVM classifier achieved a patient-level accuracy of 77.4% (24/31 correctly classified), with an area under the ROC curve (AUC) of 0.75. Conclusions:Non-invasive quantitative vascular pattern analysis from NBI images demonstrates good discriminative ability for predicting LOH status in oral dysplasia. This approach has potential as an adjunct diagnostic tool for early cancer risk stratification, potentially reducing the need for invasive biopsies.