The contribution of machine learning to predicting cancer outcome

Artificial intelligence methods may aid physicians to predict long-term outcome of individualized treatments of cancer. Hitherto, in the clinical literature on outcome prediction, traditional statistical methods prevail. This paper addresses the contribution of machine learning as compared to traditional statistical methods in the prediction of the long-term outcome of cancer treatment. Using a dataset of 1552 patients with clinical and pathological features a model was induced using a traditional statistical method (logistic regression) and a state-of-the-art machine learning method (proximal support vector machine). The models were trained to predict three outcome after five years: (1) local recurrence of the cancer, (2) metastases, and (3) overall survival of the patient. The performances of the models were evaluated using the Area-Under-the-Curve (AUC) of the Receiver Operating Characteristic (ROC) curve in combination with 10-fold cross-validation. The results reveal that both models perform on a par with mean AUCs between 0.72 and 0.78. No significant difference in performance could be established between the two methods. We conclude that proximal support vector machines do not improve the long-term cancer outcome prediction as compared to logistic regression. Further research is needed to establish if our result generalizes to other state-of-the-art methods in machine learning.