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

In this study, following specific inclusion and exclusion criteria, a comprehensive retrospective analysis was conducted on 2114 COVID-19 patients. Based on their clinical features, patients were categorized into two groups: the fever group (n = 1057) and the pneumonia group (n = 1057). It was observed that the patients in the pneumonia group were older compared to those in the fever group. This could be attributed to the natural immunosenescence that accompanies aging, leading to a decreased immune response and rendering the immune system less effective in identifying and eliminating the virus, consequently increasing susceptibility to COVID-19. These findings align with the research conducted by Petter Brodin and colleagues, who identified various factors influencing the susceptibility and severity of COVID-19 infection and emphasized the crucial role of age [10]. This observation further validates the outcomes of our study.

In our laboratory analysis of hematological parameters, out of the 33 indicators evaluated, except for Hs-CRP and Bas.%, 31 indicators including WBC, Lym.%, Eos.%, Lym.#, Eos.#, PLT, PCT, Neu.%, Mon.%, Neu.#, Mon.#, HGB, RBC, HCT, MCV, MCH, MCHC, MPV, and PDW showed statistically significant differences between the two groups. Furthermore, these 31 indicators were subjected to collinearity analysis, and those with correlation coefficients exceeding 0.9 were excluded, resulting in the creation of four models (Adaboost, XGBoost, RandomForest, LogisticRegression) based on the remaining indicators. Among these, the XGBoost model demonstrated the best performance in terms of AUC, and it automatically selected the key feature variables, assigning corresponding weights to them.

The XGBoost algorithm has garnered substantial attention from researchers and has been extensively explored for application in predicting various diseases, including, but not limited to, forecasting and classifying heart diseases [11], establishing diagnostic models for breast cancer [12], staging liver cancer [13], optimizing the dosage of immunosuppressive drugs in kidney transplant patients [14], and analyzing survival factors influencing early mortality in colorectal cancer patients [15]. The XGBoost algorithm, a machine learning method based on gradient boosting trees, is renowned for its efficiency, flexibility, and scalability. It leverages abundant data and features to construct intricate nonlinear models, capturing risk factors and potential interactions contributing to the development of diseases. By integrating multifaceted data, including clinical data, radiographic data, and laboratory data, the XGBoost algorithm can establish predictive models that offer physicians accurate judgement and decision support, thereby enhancing therapeutic outcomes and mitigating adverse reactions [16, 17].

Weights are numerical parameters that signify the importance of different features or inputs within a model. They are learned and adjusted through the training data to enhance the model’s ability to fit and predict the data accurately. In the realm of clinical diseases, researching weights in relation to disease relevancy has become a topical issue. Gopi Battineni et al. diagnosed chronic diseases by analyzing weights within machine learning models, demonstrating high utility [18]. Guided by the importance of feature variables, we selected the top ten weights (Age, Mon.%, MPV, Lym.%, SIRI, Eos.#, PLT, HGB, PDW, and Neu.#) and rebuilt the XGBoost model, undergoing training and validation processes. Compared with the model constructed using 31 indicators, the model based on these ten indicators performed superiorly, exhibiting an AUC of 0.977 in the training set and an AUC of 0.912 in the validation subset. This presents high accuracy and stability in predicting COVID-19.

In the predictive model for COVID-19, the weight of the age factor prominently stands out, surpassing other variables in terms of its influence. Numerous studies corroborate that age undeniably serves as one of the crucial factors impacting susceptibility and severity of COVID-19 [19]. As age advances, the risk of infection correspondingly escalates, potentially attributable to the decline in immune function and the presence of other latent health issues. Research conducted by Liu and others compared the age distribution of SARS-CoV-2 infection and mortality rates in China, Italy, and South Korea, revealing a substantial number of fatalities among individuals over the age of 60 in these regions [20]. Hence, in predicting and managing COVID-19, an accurate assessment and substantial consideration of the age factor play a pivotal role in formulating effective prevention and treatment strategies.

Coronavirus Disease 2019 (COVID-19), an acute respiratory infection instigated by the novel coronavirus, implicates aberrations in the hematological system throughout its pathological process. This research unveils that through weight analysis, standard hematological parameters have gained substantial prominence in assessing predictive models for COVID-19. These weights reflect the impact degree of each feature variable on the predictive outcome. For the predictive model of COVID-19, the significance of these standard hematological parameters is indispensable.

The significance of monocytes in the predictive model for COVID-19 is underscored by their weight in the model, highlighting their crucial role in forecasting outcomes. Monocytes play a pivotal role in immune responses, engaging in both antiviral and anti-inflammatory processes. Following SARS-CoV-2 infection, substantial alterations occur in the phenotype and function of monocytes, which tightly correlate with the severity of the patient’s condition. Specifically, using a computational technique referred to as “virus tracking,” Pierre et al. conducted single-cell RNA sequencing on bronchoalveolar lavage fluid samples from patients with severe and mild COVID-19. Their findings revealed a concomitant infection of human metapneumovirus within monocytes of severe patients, particularly pronounced in those monocytes affected by interferon signaling [21]. These insights further substantiate the importance of monocytes in the COVID-19 predictive model, providing crucial clues to comprehend their mechanistic roles and disease progression.

Platelets, a type of blood cell, primarily contribute to blood coagulation and hemostasis, also demonstrating significant importance within the COVID-19 predictive model. Existing research indicates that SARS-CoV-2 may interact with platelets, thereby inciting an increase in platelet activation and aggregation [22]. This activation of platelets may correlate with the incidence of inflammatory responses and thrombus formation. Concurrently, anomalies in platelet counts may emerge in COVID-19 patients, with some presenting thrombocytopenia [23]. As crucial actors in immune response and inflammation, platelets can influence the body’s antiviral and anti-inflammatory responses. Hui Liu and colleagues have constructed a risk scoring model based on routine blood examination parameters, named the PAWNN score. Incorporating platelets and related data, this score accurately predicts the mortality risk of hospitalized COVID-19 patients and allows for dynamic monitoring throughout the hospital stay. Therefore, the inclusion of platelets in COVID-19 predictive models to assess their contribution and importance in prediction outcomes bears significant implications [24].

The role of lymphocytes in COVID-19 predictive models is significant. As a crucial category of immune cells, lymphocytes play an essential role in combating pathogen infection and maintaining immune homeostasis. In the process of SARS-CoV-2 infection, research has demonstrated that a reduction in lymphocyte counts correlates closely with disease severity, potentially associated with vascular homeostasis imbalance and immune cell dysfunction triggered by a cytokine storm [25].

The SARS-CoV-2 virus primarily gains entry into host cells and instigates infection by binding to the ACE2 receptor on host cells via its spike protein. Although initially thought to predominantly affect the respiratory system, increasing research suggests that it not only impacts the lungs but also potentially has ramifications on the heart and hematologic system. The virus achieves this by disrupting the binding of the spike protein with the heme in hemoglobin, thereby depriving it of its iron atoms, rendering it incapable of carrying oxygen, and damaging the hemoglobin [26]. Consequently, a state of hypoxia ensues within the human body, leading to symptoms such as dyspnea, cyanosis, and organ damage [27]. We discerned a notable weightage of neutrophils in the COVID-19 prediction model; these cells are critical elements of innate immunity and participate in the process of combating pathogen infections. Their presence is related to inflammation, cytokine storms, and the prognosis of critically ill patients [28, 29]. By conducting weight analysis, we find these features in routine blood tests to be of significant relevance in the COVID-19 prediction model. This discovery provides useful references for our in-depth research and further clinical applications.

However, XGBoost has limitations that need to be considered. Firstly, it performs better with larger training samples and abundant computational resources, but COVID-19 research often involves incomplete, imbalanced, and inconsistent data. Secondly, as COVID-19 is a rapidly evolving disease, relying solely on existing data may not capture real-time changes. XGBoost’s reliance on stable data distributions may affect its accuracy and stability in dynamic situations like COVID-19. Additionally, the dynamic nature of COVID-19 data limits XGBoost’s applicability for prediction. Therefore, combining other methods and techniques with XGBoost is essential to enhance its specificity and sensitivity in analyzing and predicting COVID-19-related issues.

In summary, our investigation reveals that the fusion of complete blood count (CBC) parameters with advanced machine learning techniques offers a powerful approach for the prediction of COVID-19. Our model, which prioritizes CBC as the cornerstone feature, stands out with an AUC of 0.920, underscoring a substantial leap in sensitivity and specificity for the detection of COVID-19 compared to existing models. For instance, our findings suggest that our model outperforms an integrated model using variance analysis coupled with LASSO and Boruta feature selection methods, which reported an AUC of 0.910 [30].

Furthermore, our results advocate for the incorporation of the XGBoost algorithm within the tapestry of current clinical workflows. Utilizing the model’s output as a supplementary aid in diagnostic protocols has the potential to streamline the prioritization process for PCR testing, fostering a more efficient clinical decision-making process. Such a model can serve as a cogent decision-support tool, interfacing seamlessly with hospital information systems to provide timely and accurate assessments for COVID-19 diagnosis and prognosis. To further improve the model’s real-world applicability and more portable for clinical use, the online application of the model utilized CBC biomarkers and could greatly improve the efficiency and coverage of COVID-19 diagnosis.

Nevertheless, while CBC biomarkers serve as potent predictors within our model, we emphasize the necessity of their contextual interpretation in conjunction with other clinical and imaging data. This holistic approach enriches the model’s precision and reliability. The implementation of such a comprehensive predictive model holds the promise of bolstering clinicians’ capabilities in navigating the pandemic’s challenges, optimizing screening processes, and tailoring patient-specific therapeutic strategies, ultimately aspiring to refine the overall management of COVID-19.