Sparse Medical Time Series Modeling with Neural Differential Equations

Pedro Miguel Gonçalves Lopes

Key information

Authors:

Pedro Miguel Gonçalves Lopes (Pedro Miguel Gonçalves Lopes)

Supervisors:

Paulo Alexandre Carreira Mateus (Paulo Alexandre Carreira Mateus); Alexandra Sofia Martins de Carvalho (Alexandra Sofia Martins de Carvalho)

Published in

12/20/2021

Abstract

In the field of time series modeling, the problem of irregularly sampled data proves hard to overcome. Neural ordinary differential equation models (neural ODEs) are a family of models that have proved to be a successful approach to learning sparse and irregularly measured time series, having obtained several state of the art results on data interpolation. Their ability to model data continuously in time give them the ability to make accurate prediction in any time point of the time series. In this thesis, we review the concepts necessary to understand neural ODEs, ranging from numerical ODE solvers to various neural network architectures. We construct neural ODE models on top of multiple widely used neural network architectures, and test their performance on multiple tasks on sparse medical data. This study aims at finding in what conditions does the increased performance of upgraded classical neural networks with neural ODEs out weights the additional computational cost associated. The main contributions of this work are, the further testing of this family of models in order to further study its strengths and weaknesses, on multiple tasks and the testing of a new way to extrapolate data using recurrent neural networks with a neural ODE, without re-feeding predictions and relying solely on the dynamics of the neural ODE.