Online COVID-19 diagnosis prediction using complete blood count: an innovative tool for public health

Abstract

Background: COVID-19, caused by SARS-CoV-2, presents distinct diagnostic challenges due to its wide range of clinical manifestations and the overlapping symptoms with other common respiratory diseases. This study focuses on addressing these difficulties by employing machine learning (ML) methodologies, particularly the XGBoost algorithm, to utilize Complete Blood Count (CBC) parameters for predictive analysis.

Methods: We performed a retrospective study involving 2114 COVID-19 patients treated between December 2022 and January 2023 at our healthcare facility. These patients were classified into fever (1057 patients) and pneumonia groups (1057 patients), based on their clinical symptoms. The CBC data were utilized to create predictive models, with model performance evaluated through metrics like Area Under the Receiver Operating Characteristics Curve (AUC), accuracy, sensitivity, specificity, and precision. We selected the top 10 predictive variables based on their significance in disease prediction. The data were then split into a training set (70% of patients) and a validation set (30% of patients) for model validation.

Results: We identified 31 indicators with significant disparities. The XGBoost model outperformed others, with an AUC of 0.920 and high precision, sensitivity, specificity, and accuracy. The top 10 features (Age, Monocyte%, Mean Platelet Volume, Lymphocyte%, SIRI, Eosinophil count, Platelet count, Hemoglobin, Platelet Distribution Width, and Neutrophil count.) were crucial in constructing a more precise predictive model. The model demonstrated strong performance on both training (AUC = 0.977) and validation (AUC = 0.912) datasets, validated by decision curve analysis and calibration curve.

Conclusion: ML models that incorporate CBC parameters offer an innovative and effective tool for data analysis in COVID-19. They potentially enhance diagnostic accuracy and the efficacy of therapeutic interventions, ultimately contributing to a reduction in the mortality rate of this infectious disease.

Keywords: COVID-19; Complete blood Count (CBC); Machine learning; XGBoost.

Fig. 1

Study flow showing patients excluded from the study and the final cohort included in the study

Fig. 2

Study flow showing patients excluded from the study and the final cohort included in the study

Fig. 3

Receiver operating characteristic curves (ROC) showing the predictions of the four models: XGBoost, Random Forest, Logistic Regression and the AdaBoost

Fig. 4

The top ten feature importance weights

Fig. 5

a AUC for the Training Sets. b AUC for the Validation Sets

Fig. 6

Decision Curve Analysis (DCA)

Fig. 7

a Calibration Curveor the Training Sets. b Calibration Curveor the Training Sets

Fig. 8

The visualization of the prediction model through Deepwise and Beckman Coulter DxAI platform. The Supplementary table 1 and figure 1 are located in file 1