In work led by Intae Moon, we present SurvLatent ODE, a new method for survival analysis now out as a pre-print with source code. In the clinical setting, we often want to predict outcomes for patients with a complicated history. This can involve modeling irregularly collected patient covariates that change over time (longitudinal data), multiple competing outcomes (competing risks), while allowing for flexible relationships between the features. While some of these challenges have been addressed in prior models, SurvLatent ODE is a method to address these needs within a unified framework by leveraging Ordinary Differential Equations and Recurrent Neural Networks to flexibly model longitudinal data, together with non-parametric estimation of survival functions for multiple outcomes. We apply SurvLatent ODE to multiple datasets including a unique study of cancer associated thrombosis at Dana-Farber (where patient history and competing risks are particularly relevant) and show that it outperforms the established clinical risk scores based on static features, as well as existing predictive algorithms. We are excited about the potential to apply this model to other complex clinical outcomes.

For more details, see the tweet-orial from Intae and the pre-print here:

SurvLatent ODE: A Neural ODE based time-to-event model with competing risks for longitudinal data improves cancer-associated DVT prediction.
Moon I, Groha S, Gusev A. 2022

SurvLatent ODE method