Machine-Aided Detection of SARS-CoV-2 from Complete Blood Count

Overview

This study investigated whether routine complete blood count (CBC) tests can be used to detect SARS-CoV-2 infection using machine learning. The motivation was to create a cost-effective, scalable screening tool that could complement or preselect patients for RT-PCR testing.

Data Sources

Two datasets were used: the UCC dataset from the University Clinical Centre in Gdańsk, Poland (22,463 patient records, 2019–2020), and a Zenodo dataset from San Raffaele Hospital, Italy (1,624 patients, 2020–2021, from Cabitza et al.). COVID-19 status was determined by RT-PCR, with ambiguous or borderline cases excluded.

Preprocessing and Features

Only CBC features common to both datasets were selected, and highly correlated pairs (correlation > 0.5) were pruned. Missing continuous values were imputed using k-nearest neighbors, features were standardized, and SMOTE was used to address class imbalance. The final feature set comprised nine CBC parameters — HGB, MCV, MCHC, PLT, WBC, BA, LY, MO, EO — plus age and sex.

Models and Results

Four models were trained and compared: XGBoost, CatBoost, a fully connected ANN, and TabNet. On the balanced UCC dataset, TabNet achieved the highest performance with an F1-score of 87.4% and AUC of 87.1%, substantially outperforming XGBoost (66.3% F1), CatBoost (70.5%), and the ANN (72.0%). Balancing the data notably increased specificity and overall F1 across all models.