Improving five-year survival prediction via multitask learning across HPV-related cancers

Oncology is a highly siloed field of research in which sub-disciplinary specialization has limited the amount of information shared between researchers of distinct cancer types. This can be attributed to legitimate differences in the physiology and carcinogenesis of cancers affecting distinct anatomical sites. However, underlying processes that are shared across seemingly disparate cancers probably affect prognosis. The objective of the current study is to investigate whether multitask learning improves 5-year survival cancer patient survival prediction by leveraging information across anatomically distinct HPV related cancers. Data were obtained from the Surveillance, Epidemiology, and End Results (SEER) program database. The study cohort consisted of 29,768 primary cancer cases diagnosed in the United States between 2004 and 2015. Ten different cancer diagnoses were selected, all with a known association with HPV risk. In the analysis, the cancer diagnoses were categorized into three distinct topography groups of varying specificity. The most specific topography grouping consisted of 10 original cancer diagnoses differentiated by the first two digits of the ICD-O-3 topography code. The second topography grouping consisted of cancer diagnoses categorized into six distinct organ groups. Finally, the third topography grouping consisted of just two groups, head-neck cancers and ano-genital cancers. The tasks were to predict 5-year survival for patients within the different topography groups using 14 predictive features which were selected among descriptive variables available in the SEER database. The information from the predictive features was shared between tasks in three different ways, resulting in three distinct predictive models: 1) Information was not shared between patients assigned to different tasks (single task learning); 2) Information was shared between all patients, regardless of task (pooled model); 3) Only relevant information was shared between patients grouped to different tasks (multitask learning). Prediction performance was evaluated with Brier scores. All three models were evaluated against one another on each of the three distinct topography-defined tasks. The results showed that multitask classifiers achieved relative improvement for the majority of the scenarios studied compared to single task learning and pooled baseline methods. In this study, we have demonstrated that sharing information among anatomically distinct cancer types c