Prognosticating Colorectal Cancer Recurrence using Machine Learning Techniques

Danliang Ho, Dawn Qingqing Chong, Brenda Tay, Iain Bee Huat Tan, Mehul Motani

Abstract: Colorectal cancer is among the top three most commonly occurring cancers worldwide, and around 30-40% of patients treated by curative intent surgery will experience cancer recurrence. Proactive prognostication would enable clinicians to better plan treatment modality and intensity, and follow-up frequency to reduce recurrence. Here, we study the application of machine learning models to predict cancer recurrence in a cohort of 904 post-resection colorectal cancer patients. We employ heterogeneous structured and temporal clinical features including demographic and diagnostic information, tumour stage and location details, biochemistry and molecular typing results, as well as surgical details and treatment parameters. We characterize the performance of multiple machine learning classifiers including logistic regression, support vector machine, gradient boosting and multi-layer perceptron on structured data. Our best model achieved a sensitivity of 80.7% and a specificity of 88.2%. This is comparable to and even exceeding the performance of carcinoembryonic antigen (CEA), a tumour marker commonly used in the clinic for colorectal cancer monitoring. We also demonstrate feasibility for accurate forecasting of recurrence up to 4 months in advance, as well as the possibility of predicting recurrence as early as 6 months post-surgery. Our results have positive implications for better management of colorectal cancer patients in the post-resection setting.