Development of a penalized clinical risk score for oral squamous cell carcinoma

Abstract Introduction Prompt diagnosis of Oral Squamous Cell Carcinoma (OSCC) is key to favorable outcomes, but in low- and middle-income settings, constrained access to timely histopathology and imaging often leads to delayed detection. The NDB UFES dataset provides demographic, behavioral, and clinical lesion attributes that can be leveraged to generate interpretable risk models. Materials and methods The study utilized the NDB UFES dataset, comprising annotated images of oral lesions with corresponding demographic, behavioral, and lesion specific data. Cases with confirmed histopathological diagnoses were included and categorized as OSCC or non OSCC. Predictor variables included lesion size, localization, tobacco and alcohol use, sun exposure, gender, skin tone, and age group. Data were standardized and binned into clinically meaningful categories. A penalized logistic regression model with five-fold cross-validation was developed to estimate OSCC risk from the selected clinical predictors. Model coefficients were bootstrapped to estimate standard errors and 95% confidence intervals. The penalized coefficients were then transformed into a point-based risk score. Results The developed model demonstrated strong discrimination, with an AUC of 0.89 (95% CI 0.80–0.97), and a Brier score of 0.13. The simplified point-based score showed similar discrimination, with an AUC of 0.92. Larger lesion size in the 2–4 cm category, floor-of-mouth and tongue localization, older age, and male sex were among the model features associated with higher predicted risk. Former/unknown tobacco use received a positive coefficient relative to current use, whereas current tobacco use received a negative point assignment in the fitted model; these counterintuitive directions are interpreted cautiously because of the limited sample size and category construction. Conclusion An interpretable point-based risk score was developed from routinely available clinical features and internally evaluated using the NDB UFES dataset. The model demonstrated promising discrimination and calibration within this dataset. External and prospective validation are required before the score can be evaluated for use in clinical triage, referral, or other management decisions.