Prediction of SARS-CoV-2-positivity from million-scale complete blood counts using machine learning

Area of Science:

• Hematology
• Infectious Diseases
• Computational Biology

Background:

• Complete Blood Count (CBC) is a cost-effective blood test measuring key cell components.
• CBC variations offer insights into potential diseases, aiding medical decision-making.
• Machine learning (ML) can unlock predictive power from complex analyte relationships.

Purpose of the Study:

• Develop ML models for COVID-19 diagnosis using CBC data.
• Leverage non-linear relationships between blood analytes for enhanced prediction.
• Assess model performance across different COVID-19 waves and in the presence of other respiratory infections.

Main Methods:

• Compiled a large dataset of CBCs and RT-PCR tests for SARS-CoV-2 and other respiratory viruses.
• Developed an ensemble ML procedure combining models for various respiratory infections.
• Analyzed model performance during the first and second waves of COVID-19 in Brazil.

Main Results:

• Achieved high Area Under the Receiver Operating Characteristic Curve (AUROC) exceeding 90% in validations.
• Identified bias in models trained solely on SARS-CoV-2 data, leading to poor performance with other viruses.
• Demonstrated that incorporating data on other respiratory diseases is crucial for model robustness.

Conclusions:

• A novel ML approach using CBC shows potential for COVID-19 diagnosis.
• Aggregating data on other respiratory diseases is essential for robust diagnostic models.
• The low-cost, rapid CBC-based tool offers versatility for pandemic and post-pandemic scenarios.