Construction of Prediction Model of Early Glottic Cancer Based on machine learning

Abstract

Abstract Objective To explore the clinical applicability of the diagnosis of early glottic cancer based on machine learning (ML) combined with narrow-band imaging (NBI). Methods Chi-square test and multivariate logistic regression analysis were used to explore clinical and laryngoscopic features that could potentially predict early glottic cancer. Afterward, three classical ML methods, namely random forest (RF), support vector machine（SVM), and decision tree (DT), were combined with NBI endoscopic images to identify risk factors related to glottic cancer and to construct and compare the predictive models. The patients were randomly divided into a training group and a test group. In the training set, RF, DT, and SVM were used to construct a predictive model to distinguish between benign and malignant laryngeal lesions, and the test set was used to evaluate the predictive value of the model. Results According to the comparative modelling results, the RF‑based model was found to predict more accurately than other methods and have a significant predominance over others.The accuracy, precision, recall, F1 index, and AUC value of the RF model were 0.96, 0.0.90, 1.00, 0.95. The ROC curve analysis results (overall accuracy) indicate the RF model (AUC = 0.97) had the highest score among others. Conclusion RF, DT, and SVM were used to construct a predictive model for early glottic cancer, and the RF model outperformed the other models.