Machine Learning Methods in Classification of Prolonged Radiation Therapy in Oropharyngeal Cancer: National Cancer
Seungjun Ahn1,2, Eun Jeong Oh3, Matthew I Saleem4,5
1Institute for Healthcare Delivery Science, Department of Population Health Science and Policy, Icahn School of Medicine at Mount Sinai, New York City, New York, USA.
Objective:
To investigate the accuracy of machine learning (ML) algorithms in stratifying risk of prolonged radiation treatment duration (RTD), defined as greater than 50 days, for patients with oropharyngeal squamous cell carcinoma (OPSCC).
Study Design:
Retrospective cohort study.
Setting:
National Cancer Database (NCDB).
Methods:
The NCDB was queried between 2004 to 2016 for patients with OPSCC treated with radiation therapy (RT) or chemoradiation as primary treatment. To predict risk of prolonged RTD, 8 different ML algorithms were compared against traditional logistic regression using various performance metrics. Data was split into a distribution of 70% for training and 30% for testing.
Results:
A total of 3152 patients were included (1928 prolonged RT, 1224 not prolonged RT). As a whole, based on performance metrics, random forest (RF) was found to most accurately predict prolonged RTD compared to both other ML methods and traditional logistic regression.
Conclusion:
Our assessment of various ML techniques showed that RF was superior to traditional logistic regression at classifying OPSCC patients at risk of prolonged RTD. Application of such algorithms may have potential to identify high risk patients and enable early interventions to improve survival.
