| Description |
For decades, researchers have explored the e ects of clinical and biomolecular factors on disease outcomes and have identi ed several candidate prognostic markers. Now, thanks to technological advances, researchers have at their disposal unprecedented quantities of biomolecular data that may add to existing knowledge about prognosis. However, commensurate challenges accompany these advances. For example, sophisticated informatics techniques are necessary to store, retrieve, and analyze large data sets. Additionally, advanced algorithms may be necessary to account for the joint e ects of tens, hundreds, or thousands of variables. Moreover, it is essential that analyses evaluating such algorithms be conducted in a systematic and consistent way to ensure validity, repeatability, and comparability across studies. For this study, a novel informatics framework was developed to address these needs. Within this framework, the user can apply existing, general-purpose algorithms that are designed to make multivariate predictions for large, hetergeneous data sets. The framework also contains logic for aggregating evidence across multiple algorithms and data categories via ensemble-learning approaches. In this study, this informatics framework was applied to developing multivariate prognisis models for human glioblastoma multiforme, a highly aggressive form of brain cancer that results in a median survival of only 12-15 months. Data for this study came from The Cancer Genome Atlas, a publicly available repository containing clinical, treatment, histological, and biomolecular variables for hundreds of patients. A variety of variable-selection approaches and multivariate algorithms were applied in a cross-validated design, and the quality of the resulting models was measured using the error rate, area under the receiver operating characteristic curve, and log-rank statistic. Although performance of the algorithms varied substantially across the data categories, some models performed well for all three metrics|particularly models based on age, treatments, and DNA methylation. Also encouragingly, the performance of ensemble-learning methods often approximated the best individual results. As multimodal data sets become more prevalent, analytic approaches that account for multiple data categories and algorithms will be increasingly relevant. This study suggests that such approaches hold promise to guide researchers and clinicians in their quest to improve outcomes for devastating diseases like GBM. |
