Andre Goncalves | 18-ERD-017
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. In this project, we developed machine-learning-based, multitask learning models to improve survival predictions for cancer patients by leveraging information that spans multiple datasets regarding anatomically distinct human papillomavirus (HPV)-related cancers. Ten different cancer diagnoses were selected, all with a known association with HPV risk. The tasks were to predict five-year survival for patients within the different topography groups. Our results showed that multitask classifiers achieved relative improvement for the majority of the scenarios we studied, particularly for rare cancer sites, as compared to single-task learning models, where no information is shared across datasets, and pooled baseline methods, where all information is pooled into a single dataset.
The methods developed in this research can more accurately predict cancer patient outcomes, which supports Lawrence Livermore National Laboratory's core competencies in bioscience and bioengineering, and high-performance computing, simulation, and data science, including the ability to use multitask learning to improve predictive capabilities.
Publications, Presentations, and Patents
Goncalves, A., et al. 2018a. "Multitask Learning and its Applications to Alzheimer's Disease Progression and Cancer Survival Prediction." Lawrence Livermore National Laboratory CASIS Workshop, Livermore, CA, LLNL-PRES-75158
——— 2018b. "Deep Entity Embeddings for Cancer Survival Prediction." Lawrence Livermore National Laboratory Data Science Institute Workshop, Livermore, CA. LLNL-POST-75259
——— 2019. "Bayesian Multitask Learning Regression for Heterogeneous Patient Cohorts." Journal of Biomedical Informatics X 4:100059. LLNL-JRNL-767154
Ladd, A., et al. 2020. "Recommender Systems for Screening Cancer Treatment." Lawrence Livermore National Laboratory Research Slam, Livermore, CA. LLNL-PRES-812906