(022) Machine Learning and Deep Learning Techniques for the Survival Prediction of Oropharyngeal Squamous Cell Carcinoma: Ahead of the Pack

Background: Oropharyngeal squamous cell carcinoma (OPSCC) is prevalent globally, particularly among men. Artificial intelligence techniques like deep learning and machine learning can be used for survival prediction with better accuracy than traditional models.

Objectives: This study uses machine learning and deep learning techniques to develop models for identifying and visualizing significant predictors of OPSCC survival rates.

Methods: A retrospective cohort study was conducted at a tertiary care cancer hospital in Assam. A dataset of 1042 OPSCC patients was retrieved from a hospital-based cancer registry in northeast India between 2018 and 2020. The dataset contained 12 predictor variables (independent variables) associated with the survival status (dependent variable) of the patients (alive or dead). Logistic regression, support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP) were the prediction models used to determine the significant predictors of OPSCC survival rate. The key variables were ranked using random forest variable selection techniques. The study used various metrics to assess the classification performance of the models, including sensitivity, specificity, accuracy, precision, F1 score and Matthews’ correlation coefficient. All the methods were performed using the scikit-learn library in Python, designed for machine learning and deep learning tasks.

Results: The MLP model achieved the best performance across all evaluation metrics in this study. The MLP model was observed to have higher accuracy (accuracy=83.25%; area under the curve (AUC)=0.83) and precision (precision=0.83) than RF (accuracy=81.34%; AUC=0.80; precision=0.80), logistic regression (accuracy=77.99%; AUC=0.78; precision=0.75) and the lowest in SVM (accuracy=75.12%; AUC=0.75; precision=0.70). Variables such as age, treatment (surgery followed by radiotherapy), metastasis, T stage, and nodal status can be used to predict survival rates among OPSCC patients.

Conclusions: A deep learning technique incorporating the clinical variable can be an excellent model for predicting survival rates among OPSCC patients. The MLP model was observed to have better predictability than other models; however, clinical validation is warranted to confirm these findings.