Implementing GAN-based method for real-valued medical time series data generation

Student

Anushree Choudhary

Advisor

Jaideep Srivastava

Abstract

It is difficult to do analysis or prediction using medical data because it is highly sparse and imbalanced in nature which hampers the detection of unlikely events due to the introduced bias. A generative model is a framework which tackles both the issues by generating synthetic data which is almost like the real data. In this project, I try to implement an oversampling technique using the framework of Generative Adversarial Networks (GANs). Using GANs, we overcome the problem of imbalanced class in the data and in turn it helps in building a highly accurate predictive model for the sparse data. The data is in time-series format and hence, we need to take care of the temporal dynamics of the data. The model architecture used for this project includes generating realistic time-series data after combining the flexibility of the unsupervised paradigm and exercising some control over the network dynamics by introducing supervised training. This framework is then used to predict the in-hospital mortality risk for ICU patients using the original data and the synthesized data.

Video

Implementing GAN-based method for real-valued medical time series data generation