In Anticancer research
BACKGROUND/AIM : Machine learning (ML) models are often modelled to predict cancer prognosis but rarely consider spatial factors in a region. Hence this study explored machine learning algorithms utilising Local Government Areas (LGAs) in Queensland, Australia to spatially predict 3- and 5-year prognosis of oral cancer patients and provide clinical interpretability of the predicted outcome made by the ML model.
PATIENTS AND METHODS : Data from a total of 3,841 oral cancer patients were retrieved from the Queensland Cancer Registry (QCR). Synthesizing minority oversampling technique together with edited nearest neighbours (SMOTE-ENN) was used to pre-process unbalanced datasets. Five ML models: logistic regression, random forest classifier, XGBoost, Gaussian Naïve Bayes and Voting Classifier were trained. Predictive features were age, sex, LGAs, tumour site and differentiation. Outcomes were 3- and 5-year overall survival of patients. Model performances on test set were evaluated using area under the curve and F1 scores. SHapley Additive exPlanations (SHAP) method was applied to the best performing model for model interpretation of the predicted outcome.
RESULTS : The Voting Classifier was the best performing model with F1 score of 0.58 and 0.64 for 3- and 5-year overall survival, respectively. Age was the most important feature in the Voting Classifier in 3- and 5-year prognosis prediction. LGAs at diagnosis was the top 3 predictive feature for both 3- and 5-year models.
CONCLUSION : The Voting Classifier demonstrated the best overall performance in classifying both 3- and 5-year overall survival of oral cancer patients in Queensland. SHAP method provided clinical understanding of the predictive features of the Voting Classifier.
Tan Jia Yan, Adeoye John, Thomson Peter, Sharma Dileep, Ramamurthy Poornima, Choi Siu-Wai
2022-Dec
Oral cavity cancer, SHapley values, interpretability, machine learning, prognosis
