Automated Detection of COVID-19 Using Machine Learning Analysis of White Blood Cell Flow Cytometry Images

Background: COVID-19 is usually diagnosed by real-time polymerase chain reaction testing of nasopharyngeal swabs, but this approach may be associated with a prolonged time to diagnosis. Machine learning incorporating findings from standard laboratory tests may more rapidly identify individuals infected with SARS-CoV-2, enabling early management and protection for healthcare staff. Methods: A machine learning model was developed to detect patients infected with COVID-19 using only white blood cell flow cytometry without using symptom or clinical history data. The model included images from 106 patients positive for SARS-CoV-2 and 211 controls admitted to a university hospital with respiratory symptoms. A total of 17 texture-feature analysis methods were tested using three classifiers. Finally, the Particle Swarm Optimization algorithm was used to combine the best-performing models. The performance of the final model was assessed using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, precision, and F1-Score. Results: Using the five-fold cross-validation technique, the final model achieved an accuracy of 88.96% for diagnosing patients with a SARS-CoV-2 infection. It also had a sensitivity of 78.30%, a specificity of 94.31%, a precision of 87.83%, an F1-score of 0.83, and an AUROC of 0.86. Conclusions: The final algorithm showed good diagnostic performance even when compared with models that included clinical, epidemiological, and other laboratory or ancillary test data. The use of cytometry images could represent a significant advance in the early diagnosis of COVID-19.