| Sumario: | Objective: To investigate the potential of machine learning systems for the clinical classification of palatal salivary gland tumors into five diagnostic categories using only demographic and clinical data. Methods: Four machine learning models—Multilayer Perceptron (MLP), Support Vector Machine (SVM), Random Forest (RF), and XGBoost—were implemented based on demographic and clinical attributes from 100 patients. The methodology involved a four‐step process: Hyperparameter optimization using systematic search for combinations (Grid Search), fivefold cross‐validation comprising training and testing, classifier training, and tests, followed by the acquisition of standardized performance metrics. The input attributes included age, sex, location, and symptoms. Performance was evaluated using accuracy, macro‐average sensitivity, specificity, precision, and F1‐score. Results: According to the mean accuracy values, XGBoost and MLP achieved the highest performance (81%), followed by SVM (80%) and RF (79%). Nevertheless, for all models, both macro‐averaged sensitivity and F1‐score were relatively low, remaining below 75%. Specificity emerged as the most consistent metric, ranging from 85% to 90% across all classifiers. All models reached a perfect score (1.0) in the classification of PA, whereas performance declined for malignant tumors, particularly for the rarer subtypes. Conclusions: Machine learning is a feasible approach for classifying palatal salivary gland tumors, demonstrating high specificity but limited sensitivity, primarily due to the uneven distribution of tumor subclasses, particularly the malignant ones, which are underrepresented owing to their rarity. XGBoost proved to be the most robust model with a low computational cost.
|
|---|