Mix Contrast for COVID-19 Mild-to-Critical Prediction

In a few patients with mild COVID-19, there is a possibility of the infection becoming severe or critical in the future. This work aims to identify high-risk patients who have a high probability of changing from mild to critical COVID-19 (only account for 5% of cases).

Using traditional convolutional neural networks for classification may not be suitable to identify this 5% of high risk patients from an entire dataset due to the highly imbalanced label distribution. To address this problem, we propose a Mix Contrast model, which matches original features with mixed features for contrastive learning. Three modules are proposed for training the model: 1) a cumulative learning strategy for synthesizing the mixed feature; 2) a commutative feature combination module for learning the commutative law of feature concatenation; 3) a united pairwise loss assigning adaptive weights for sample pairs with different class anchors based on their current optimization status.

We collect a multi-center computed tomography dataset including 918 confirmed COVID-19 patients from four hospitals and evaluate the proposed method on both the COVID-19 mild-to-critical prediction and COVID-19 diagnosis tasks. For mild-to-critical prediction, the experimental results show a recall of 0.80 and a specificity of 0.815. For diagnosis, the model shows comparable results with deep neural networks using a large dataset. Our method demonstrates improvements when the amount of training data is small or imbalanced.

Identifying mild-to-critical COVID-19 patients is important for early prevention and personalized treatment planning.