Abstract
The COVID-19 pandemic has resulted in significant mortality and poses a major threat to global public health. Accurate predictions of mortality risk and the need for medical interventions are essential for patient survival and the efficient allocation of limited medical resources. However, machine learning classifiers often struggle with limited data characterized by missing attributes and class imbalances. To address these challenges, we propose HexaGAN, a generative adversarial network with a hint mechanism designed to predict patient survival and interventions such as intubation and supplemental oxygen. Our experiments demonstrate that HexaGAN outperforms existing benchmark techniques, particularly in correctly predicting mortality, where it showed approximately twice the performance of standard methods. This approach offers a promising path toward timely treatment and optimized resource allocation.
Introduction
Predicting critical outcomes for COVID-19 patients, such as the need for mechanical ventilation or the risk of mortality, is vital for hospital resource management. However, Electronic Health Records (EHRs) are frequently incomplete and unbalanced. Specifically, clinical datasets often suffer from three simultaneous "limited data" problems:
- Missing Data: Not all patients receive every diagnostic test, leading to missing attributes.
- Missing Labels: Outcomes for some patients may go unrecorded due to administrative gaps or labeling costs.
- Class Imbalance: The majority of patients survive, making the "deceased" class significantly smaller and harder for standard algorithms to learn.
Traditional approaches typically address these issues sequentially (e.g., imputation followed by oversampling). However, this "cascade" method often fails because it ignores the inherent dependencies between these data flaws. This paper explores a unified framework to address these challenges simultaneously.
Methods and Data
The proposed solution utilizes HexaGAN, a generative framework designed to address missing data, missing labels, and class imbalance within a single architecture. HexaGAN consists of six interacting components, including an encoder (E), multiple generators for imputation ($G_{MI}$) and oversampling ($G_{CG}$), discriminators ($D_{MI}$, $D_{CG}$), and a classifier (C).
A key innovation is the Hint Mechanism. During the data imputation phase, a "hint generator" provides the discriminator with partial information regarding which values were originally missing. This forces the generator to produce more realistic, distribution-faithful data, which in turn improves the final classifier’s performance.
The study utilized the publicly available covid-chestxray-dataset, which includes chest X-ray images and tabular metadata such as age, sex, temperature, and various blood cell counts. The dataset comprised 846 total records, with 468 confirmed COVID-19 cases used as labeled data to predict survival, intubation, and supplemental oxygen requirements.
Results and Discussion
HexaGAN was evaluated against 24 combinations of standard techniques, such as MICE for imputation and SMOTE for oversampling. The results indicated that treating imputation, oversampling, and classification as a single adversarial game is superior to sequential pipelines, as the components "help" each other learn the true underlying distribution of patient data.
Performance Metrics
- Mortality Prediction: HexaGAN achieved a specificity of 0.8071, nearly doubling the performance of the best benchmark combination (0.4141). This jump in specificity is crucial for correctly identifying high-risk patients despite severe class imbalance.
- Intervention Prediction: The model achieved a sensitivity of 1.0000 for intubation, successfully identifying every patient in the test set who required a ventilator. For oxygen requirements, it reached an AUROC of 1.0000 at specific hint rates.
The Role of the Hint Rate
The "hint rate" (the probability with which the hint generator provides information to the discriminator) is a critical hyperparameter. Analysis showed that a hint rate between 0.7 and 0.8 yielded the best balance for both mortality and intervention prediction, as shown in the performance curves below.
Conclusion
HexaGAN demonstrates significant potential in predicting COVID-19 outcomes from messy, real-world clinical data. By integrating data imputation and class balancing into the training process via an adversarial framework, the model achieves superior predictive accuracy for high-stakes medical interventions. While the current study is cross-sectional, future work should extend this framework to longitudinal, time-series data to provide real-time clinical recommendations and validate the model on larger, multi-hospital datasets.
Summary updated