SurvivalNet

The SurvivalNet project is creating next-generation survival modeling methods using contemporary machine learning techniques to predict disease progression, metastasis, and survival. Data produced in clinical labs or research are increasingly high-dimensional, and are outgrowing the capacity of conventional survival modeling approaches. Advanced molecular assays can generate thousands of features describing the transcriptional, genetic, and epigenetic state of a patient's disease, and new modeling capabilities are needed to realize the potential of these data for precision treatment. In addition to providing accurate predictions, these methods be transparent, and should convey their confidence to clinicians.

The methods we have developed combine conventional statistical approaches with neural networks and modern optimization techniques, creating models that can learn highly-predictive representations from censored data. We have demonstrated these models using genomic and proteomic data in breast cancer and brain tumors. These methods can also utilize convolutional networks to learn from clinical outcomes and imaging data like histology.