Uiwon Hwang , Euideuk Hwang , Minsoo Kang , Sungroh Yoon

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

Seoul National University

Kyungpook National University

1

2

1

1

1

2

Coronavirus disease 2019

​

​

​

​

​

​

​

​

​

​

​

• Numerous countries have faced shortage of medical resources, including beds and medical staff

🡪 COVID-19 pandemic poses an unprecedented threat to health and public health around the world

COVID-19

COVID-19

Since

December 2019

# of confirmed cases�> 500 million

Death toll

> 6 million

(up to May 16, 2022)

​

​

40%

40%

Silent

20%

​

​

10%

Clinical manifestation

Pneumonia

Hypoxemic pneumonia

: requires oxygen treatment

40%

Benign upper respiratory ds.

Wynants, Laure, et al. "Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal." bmj 369 (2020).

Zhang, Qian, et al. "Human genetic and immunological determinants of critical COVID-19 pneumonia." Nature 603.7902 (2022): 587-598.

Coronavirus disease 2019

​

​

• Mortality prediction in COVID-19 patients is necessary
• To efficiently allocate limited medical resources and provide the best treatment to patients

​

• Timely clinical interventions (e.g. intubation, supplemental oxygen) is important
• To decrease the death rate of COVID-19 patients

​

🡪 ML models that predict the mortality and intervention can be of great help to � the public health crises and patients’ lives

​

Dataset

• Limited data problems in electronic health records (EHRs)
• Missing data problem: some attributes can be missing
• Class imbalance problem: outcomes can be imbalanced
• Missing label problem: outcomes can be missing

​

Attributes in EHR

Class

# of samples

4

9

Dataset

• Covid-chestxray-dataset
• Chest X-ray images + Metadata
• 846 records: 468 records obtained from COVID-19 patient + 378 records obtained from other pneumonias

+

Cohen, Joseph Paul, Paul Morrison, and Lan Dao. "COVID-19 image data collection." arXiv preprint arXiv:2003.11597 (2020).

Method

Real?�Fake?

real data distribution

generated data distribution

loss metrics

Arjovsky, Martin, Soumith Chintala, and Léon Bottou. "Wasserstein generative adversarial networks." International conference on machine learning. PMLR, 2017.

Method

• HexaGAN

​

​

​

​

​

​

​

• Hint mechanism

7/11

1

1

0.5

0.5

Hint generator�(𝑤.𝑝. hint rate)

Hwang, Uiwon, et al. "Hexagan: Generative adversarial nets for real world classification." International Conference on Machine Learning. PMLR, 2019.

Yoon, Jinsung, et al. "Gain: Missing data imputation using generative adversarial nets." International conference on machine learning. PMLR, 2018.

Overview

​

​

​

​

​

​

​

​

​

​

​

Chest X-ray

Metadata

(w/ missing data)

1

1

1

0

Missingness�vector

1

1

0.5

0.5

HexaGAN

survival

intubation

oxygen

Prediction

Proper�treatment

Hint mechanism

Hint generator�(𝑤.𝑝. hint rate)

Dataset

Results

• Mortality prediction (survival)

x2

Results

• Intervention prediction
• Intubation prediction

​

​

​

​

​

​

• Supplemental O2 prediction

Discussion

• ML perspective
• Cascade combination of existing ML methods
• Requires building a pre-processing pipeline that depends on problems present in the training data
• Ignores the connections between the problems

​

​

​

​

• HexaGAN
• No need to build a specific preprocessing pipeline
• Deals with limited data problems simultaneously through interaction between its components

​

​

​

​

​

​

Real world data

HexaGAN

Discussion

​

​

• Medical perspective
• The model presented in this study can predict survival, the use of supplemental oxygen, and intubation which are closely related to the severity of COVID-19
• The early use of medicines according to the prediction of the model could be further investigated

​

• GAN methods trained on datasets obtained from longitudinal studies can predict whether an intervention should be recommended at a particular moment
• This will be an important avenue for future work to provide more timely treatment to COVID-19 patients in the presence of the limited data problems

​

Thank you!

• TL;DR
• We construct a deep generative model to provide accurate prediction of mortality and interventions for COVID-19 patients from the dataset with limited data problems

​

• Summary
• Use HexaGAN which addresses limited data problems simultaneously and apply a hint mechanism to enhance the prediction performance
• Significantly outperform combinations of existing methods

​

• Anticipate that our approach could help
• Provide appropriate treatments on time
• Allocate limited medical resources efficiently
• Reduce the mortality rate of COVID-19 patients