Prediction of Mortality and Intervention in COVID-19 Patients Using Generative Adversarial Networks

The COVID-19 pandemic hits worldwide with a significant number of deaths and poses a major threat to public health. Accurate predictions of the risk of death and medical interventions are crucial for the survival of infected patients and the distribution of limited medical resources. Although machine learning classifiers can be used to predict mortality and medical interventions, it is problematic to employ the methods because training data are limited whose attributes may be missing and classes may be imbalanced. To effectively cope with these problems, we construct HexaGAN with a hint mechanism to predict the survival of the patients and medical interventions such as intubation and supplemental oxygen. In experiments, our method outperforms combinations of existing techniques for limited data problems. Notably, our method showed about twice higher performance than benchmarks in predicting deceased patients correctly. We anticipate that our approach could help provide appropriate treatments on time, allocate limited medical resources efficiently, and ultimately reduce the mortality rate of COVID-19 patients.