Deep Generative Learning-Based 1-SVM Detectors for Unsupervised COVID-19 Infection Detection Using Blood Tests

Area of Science:

• Computational biology
• Machine learning in healthcare
• Infectious disease diagnostics

Background:

• Real-time reverse transcription polymerase chain reaction (rRT-PCR) tests are standard for COVID-19 detection but have limitations.
• Limitations include potential false positives/negatives, high cost, need for specialized labs, and lengthy result times.
• A more accessible, rapid, and cost-effective diagnostic method is needed.

Purpose of the Study:

• To introduce flexible, unsupervised data-driven approaches for COVID-19 detection using blood tests.
• To frame COVID-19 detection as an anomaly detection problem solvable with an unsupervised deep hybrid model.
• To amalgamate variational autoencoder (VAE) for feature extraction and one-class support vector machine (1SVM) for detection sensitivity.

Main Methods:

• Developed an unsupervised deep hybrid model combining VAE and 1SVM for anomaly detection in blood samples.
• Imputed missing blood test data using a random forest regressor.
• Evaluated model performance using blood test datasets from Brazil and Italy.

Main Results:

• The proposed VAE-based 1SVM detector demonstrated superior discrimination performance for potential COVID-19 infections.
• Outperformed other models including GANs, DBNs, RBMs, and standalone 1SVM.
• Deep learning-driven 1SVM approaches showed promising detection performance compared to conventional deep learning models.

Conclusions:

• The VAE-based 1SVM offers a promising, data-driven solution for COVID-19 detection via blood tests.
• This approach presents a more accessible and potentially faster alternative to existing diagnostic methods.
• Further validation of deep learning-driven 1SVM for infectious disease diagnostics is warranted.