Predicting Overall Survival Using Machine Learning Algorithms in Oral Cavity Squamous Cell Carcinoma.

Area of Science:

• Oncology
• Data Science
• Spatial Epidemiology

Background:

• Machine learning (ML) models are increasingly used for cancer prognosis prediction.
• Existing models often overlook crucial spatial factors within specific geographic regions.
• This study addresses the gap by incorporating Local Government Areas (LGAs) into ML models for oral cancer prognosis.

Purpose of the Study:

• To explore machine learning algorithms for spatially predicting 3- and 5-year oral cancer patient prognosis in Queensland, Australia.
• To provide clinical interpretability of the ML model's predictions.
• To investigate the impact of spatial factors (LGAs) on oral cancer survival prediction.

Main Methods:

• Utilized data from 3,841 oral cancer patients from the Queensland Cancer Registry (QCR).
• Applied the SMOTE-ENN technique for pre-processing unbalanced datasets.
• Trained and evaluated five ML models (logistic regression, random forest, XGBoost, Gaussian Naïve Bayes, Voting Classifier), using age, sex, LGAs, tumour site, and differentiation as predictors.
• Employed SHapley Additive exPlanations (SHAP) for model interpretation.

Main Results:

• The Voting Classifier achieved the highest performance with F1 scores of 0.58 (3-year) and 0.64 (5-year) survival.
• Age was identified as the most significant predictor for both 3- and 5-year prognosis.
• Local Government Areas (LGAs) emerged as a top 3 predictive feature for both survival periods.

Conclusions:

• The Voting Classifier model demonstrated superior performance in predicting oral cancer patient survival in Queensland.
• The SHAP method successfully provided clinical insights into the predictive features influencing the model's outcomes.
• Integrating spatial data (LGAs) enhances the predictive accuracy and interpretability of ML models for cancer prognosis.