QCovSML: A reliable COVID-19 detection system using CBC biomarkers by a stacking machine learning model

Area of Science:

• Medical Diagnostics
• Machine Learning in Healthcare
• Hematology

Background:

• The reverse transcription-polymerase chain reaction (RT-PCR) test is the gold standard for COVID-19 diagnosis but has limitations including long turnaround times, high false-negative rates, and high costs.
• There is a critical need for efficient, accurate, accessible, and widely available diagnostic alternatives to RT-PCR for COVID-19.

Purpose of the Study:

• To develop and validate a novel COVID-19 detection system utilizing complete blood count (CBC) biomarkers.
• To employ a stacking machine learning (SML) model as a rapid and cost-effective diagnostic tool for COVID-19.

Main Methods:

• Utilized seven public datasets, with a primary dataset of 1624 patients (52% COVID-19 positive) from San Raphael Hospital, Italy.
• Identified key biomarkers including white blood cell count, monocyte percentage, lymphocyte percentage, and age using five feature selection techniques.
• Developed and validated a stacking machine learning model and a nomogram-based scoring system (QCovSML).

Main Results:

• The SML model achieved high performance metrics: 91.44% weighted precision, 91.44% sensitivity, 91.44% specificity, 91.45% overall accuracy, and 91.45% F1-score.
• The QCovSML system demonstrated excellent calibration and discrimination, with an Area Under the Curve (AUC) of 0.961 for internal validation and 0.967 for external validation.
• External validation across six datasets from three countries confirmed the model's generalizability and robustness, with average weighted precision of 92.02% and accuracy of 93.34%.

Conclusions:

• The proposed CBC biomarker-based SML model and QCovSML nomogram offer a promising, fast, and cost-effective alternative for COVID-19 diagnosis.
• The SML model outperformed other state-of-the-art machine learning classifiers in COVID-19 prediction.
• The robust performance and generalizability across diverse datasets highlight the potential clinical utility of this approach for widespread COVID-19 screening.