Machine Learning Models to Discriminate COVID-19 Severity with Biomarkers Available in Brazilian Public Health

Abstract

Despite advances in vaccination and treatment, the emergence of Long COVID cases has highlighted the continued public health concern posed by the disease. Studies on the early prediction of COVID-19 severity and the identification of associated biomarkers are decisive for preventing Long COVID. The objective is to utilise laboratory test data from patients diagnosed with COVID-19 and apply machine learning techniques to predict disease severity and identify associated biomarkers. From a university hospital in southern Brazil, we processed biochemical and haematological data from patients with COVID-19 (non-severe = non-ICU admission; severe = ICU admission). The data were used to train 15 machine learning algorithms to predict patient prognosis. The Light Gradient Boosting Machine (LGBM) model demonstrated the most effective performance in predicting the prognosis of patients with COVID-19, with accuracy, sensitivity, specificity, and precision values between 80 and 88%. Biomarkers associated with disease severity included Platelets, Creatinine, Erythrocytes, C-reactive protein, Lymphocytes, Albumin, Glucose, Urea, and Sodium. The results of this study demonstrate that machine learning, particularly LGBM, is an effective method for predicting the severity of COVID-19. Identifying specific biomarkers associated with disease severity is crucial for the early intervention and prevention of Long COVID, thereby improving clinical outcomes and patient management. LGBM maintained its performance across different age groups.

1. Introduction

Severe Acute Respiratory Syndrome (SARS) is a contagious illness caused by a coronavirus. COVID-19, which is related to SARS, has the SARS-CoV-2 coronavirus as its agent [1], also described as HCoV-19 or COVID-19 (Coronavirus Disease), according to Thevarajan et al. (2020) [2]. In December 2019, SARS-CoV-2 was detected in the city of Wuhan, China, and in March 2020, the World Health Organization (WHO) recognised the infection with the new coronavirus as a serious public health problem and characterised it as a pandemic [3].

Common symptoms associated with COVID-19 include fever (83–98%), cough (59–82%), shortness of breath (19–55%), and muscle pain (11–44%), with most individuals (>85%) making a good recovery, according to Huang et al. (2020) [4]. Notwithstanding the considerable progress made in the field of vaccination and the development of efficacious treatments, the ongoing threat posed by the novel coronavirus (COVID-19) persists due to the emergence of novel variants and the phenomenon of Long COVID. The World Health Organization (WHO) estimates that approximately 10–20% of individuals infected with SARS-CoV-2 develop prolonged symptoms, which are collectively referred to as Long COVID [5]. The persistence of these symptoms represents an additional challenge for healthcare systems, underscoring the pressing need for effective strategies for disease prevention and management [6]. To increase the complexity of this scenario, there is the possibility that a fully recovered infected individual may be re-infected, suggesting that permanent immune protection may not be available [7].

In the medicine field, a biomarker is a measurable indicator of the current physiological status and indicates the presence and severity of some illness. Laboratory biomarkers provide a dynamic and powerful approach to understand the disease spectrum, with applications in observational and analytical epidemiology, randomised controlled trials, screening, diagnosis and prognosis [8,9].

Laboratory medicine plays an essential role in the early detection, diagnosis, and treatment of many diseases, as the laboratory diagnosis goes beyond the etiological diagnosis and those epidemiological surveillance aspects of in vitro testing [10].

Due to the complexities associated with the evolution of COVID-19, many different biomarkers are used in distinct stages of the pathological process [11]. Moreover, some biomarkers are directly associated with the pathology severity prognosis. The selection of severity indicators for COVID-19 is relevant in the context of optimising the diagnosis and in the election of patients who require priority attention in the Public Health System (SUS)—publicly funded healthcare system in Brazil.

Research carried out in China revealed that patients with a positive diagnosis of severe acute respiratory syndrome (SARS) showed abnormalities in laboratory tests directly related to the progression of the disease. Furthermore, the study concluded that biomarkers can be used to assess the health status of patients infected with SARS [12].

We emphasise that these are routine laboratory biomarkers and not special laboratory tests. The list of biomarkers and their biological system source is presented in

Table 1. Laboratory biomarkers by profile/biological system source.

Profile (Class)	Biomarker
Glycemic	Glucose
	HbA1c
Liver	Alanine transaminase
	Aspartate transaminase
	Albumin
	Total Protein
	Globulin
	Ferritin
	C-reactive protein
Lipid	Total Cholesterol
	HDL-C
	LDL-C
	Triglycerides
Pancreatic	Amylase
	Lipase
Renal	Creatinine
	Urea
	Potassium
	Sodium
Acid-Base Balance	pCO2
	pO2
	sO2
	pH
	HCO3-standard
	HCO3-actual
	BE-CF
	BE-B
	CTCO2
Coagulation	Prothrombin Time
	Prothrombin Time–Relation
	Prothrombin Time—International Normalised Ratio
	Partial Thromboplastin Time
	Partial Thromboplastin Time–Relation
	D-Dimer
Complete Blood Count	Erythrocytes
	Haemoglobin
	Leukocytes
	Mature Neutrophils
	Immature Neutrophils
	Neutrophils
	Basophils
	Eosinophils
	Lymphocytes
	Atypical Lymphocytes
	Monocytes
	Platelets
Inflammatory	Procalcitonin
Cardiac	Troponin I (cTnI)

The table above corresponds to 10 classes of circulating and functional biomarkers, with a total of 48 routine laboratory biomarkers used as final data to analyse the severity of COVID-19 through the application of machine learning.

.

Artificial intelligence (AI) is a branch of computer science that aims to build devices that simulate the human ability to reason, perceive, make decisions, and solve problems. AI is divided into five fields: Natural Language Processing; Automated Reasoning; Machine Learning; Computer Vision; and Robotics [13].

Machine Learning (ML) is a field of AI that consists of algorithms capable of learning from models or examples. Its fundamentals are used in other areas of AI with applications seen in several areas [13], like in the creation of models that allow predictions applied in the diagnosis of diseases and the simulation of prognoses [14,15].

One of these algorithms of ML is the Light Gradient Boosting Machine (LGBM). LGBM is a decision tree-based algorithm that applies a gradient boost. It is designed to run models efficiently by both high-performance computers and small mobile devices, with the following advantages: high speed and efficiency in training; low memory usage; better accuracy; support for parallel and Graphics Processing Unit (GPU) learning; and ability to work with huge amounts of data [16].

According to the British Medical Journal (BMJ) treatment algorithm for COVID-19, the disease severity scale was classified as mild to moderate (non-severe), severe or critical. The BMJ is a global health knowledge provider that offers a variety of services, including the British Medical Journal (BMJ), BMJ Learning and BMJ Best Practice. BMJ Best Practice is an evidence-based clinical decision support tool being applied to diagnosis, prognosis, treatment and prevention using a robust methodology based on evidence and expert opinion [17]. The World Health Organisation (WHO) classification was also consulted, specifically in the Clinical Management of COVID-19: Living guideline, 18 August 2023 [18]. For this study, ICU admission (severe) and non-ICU admission (mild to moderate) were used as criteria for classifying COVID-19 patients by applying Structured Query Language (SQL) commands in the MySQL Database Management System where data obtained from a university hospital was stored.

The infectious process of COVID-19, in severe cases, generates a greatly amplified inflammatory response in the host, known as a ‘cytokine storm’ or ‘inflammatory storm’, associated with severe evolutions of viremia. In addition to the inflammatory response, and in association with it, multiple different biomarker laboratory parameters undergo significant changes in their concentrations/characteristics in the blood. Even in the early days of the pandemic, the literature already described laboratory changes associated with COVID-19, without, however, identifying a biomarker specific to the process [19]. In this study, using computational tools combined with machine learning, we sought to evaluate multiple laboratory biomarkers, with two relevant outcomes for the decision-making of health professionals who are on the front line of patient care. Firstly, it is possible to identify or estimate the severity and, consequently, the prognosis of COVID-19 based on laboratory results.

The aim of this study is to contribute to mitigating the impacts of COVID-19 by offering computational tools that can be used for the early detection of severe cases and the prevention of the development of long COVID, identifying the best machine learning model for the prognosis of the severity of the disease and verifying the accuracy of the model applied to the biomarkers in this study.

2. Materials and Methods

This study was approved by Human Research Ethics Committee under CAAE no. 43948621.7.3001.0096.

2.1. Study Location

The system developed is the result of a research project conducted in the city of Curitiba, the capital of Paraná, which is one of the three states that compose Brazil’s Southern Region. The population of the State of Paraná is approximately 65% Euro-Brazilians, 34% Afro-Brazilians and 1% Orientals and Indigenous people. The estimated population of Curitiba is 1,963,726, with 253 facilities belonging to the Unified Health System and a Human Development Index (HDI) of 0.823 [20].

The Clinics Hospital of the Federal University of Paraná (HC-UFPR), a federal public teaching hospital and a leading institution in medical-hospital care in Curitiba city that aims to serve teaching, research, and extension programmes in the areas of health sciences while providing healthcare to the community. It is the largest public hospital in Paraná State and the third largest university hospital in Brazil. It is a reference in several areas of healthcare, performing all its services free of charge, being fully financed by the Unified Health System. It assists only medium and severe risk cases [21]. HC-UFPR supplied the data analysed.

2.2. Collected Data

Samples of laboratory test results from patients with and without COVID-19 were exported from the database at HC-UFPR. The raw data was obtained in spreadsheets separated by semicolons in Comma Separated Values file format. At the Clinics Hospital of the Federal University of Paraná, data were extracted using SQL (Structured Query Language) commands in IBM’s DB2© Database Management System©, generating five CSV files. The laboratory biomarkers used correspond to the period between March 2020 and September 2022. The collected records were de-identified and lacked a specific timestamp.

The exported database commenced with an estimated total of 2.2 million samples. The samples are biomarkers that were selected based on the criteria of low cost and routine use by Brazil’s Unified Health System (SUS) to treat patients. Following the completion of all processing and organisation, the number of samples reached 35,109, comprising 50 features: 2 anthropometric features (sex and age) and 48 laboratory biomarkers (Table 1).

2.3. Data Processing

The results of laboratory tests extracted from the studied institution’s database were processed with computational tools. The data processing consisted of generating a single CSV file, which was achieved using the Python programming language (version 3.13.5) in the Jupyter Notebook (version 7.3.2) computing environment, with the Pandas (version 2.2.3) and Matplotlib libraries (version 3.10.0) [22].

Initially, five files were received from HC-UFPR in spreadsheet format. The first file contained 1,006,369 samples, the second 149,809, the third 826,829, the fourth 149,704, and the fifth 98,833. Pre-processing was performed by organising and removing redundant and incompatible data. Subsequently, the data from the disparate files exported by the HC-UFPR laboratory system was consolidated. This was achieved through the utilisation of the Python programming language in conjunction with the Jupyter Notebook computing environment, which facilitated the joining of the aforementioned files, thereby generating a single file comprising 2,231,544 samples (Figure 1).

The results of the biomarker samples from this single file were arranged in chronological order, with the date the tests were carried out on the left and the patient’s record from which the laboratory tests were collected on the right. After being arranged in this manner, a total of 142,662 samples were identified (Figure 1).

Subsequently, the classification of COVID-19 positivity was conducted for each sample, with a value of 1 assigned to positive samples and 0 to negative samples. This classification was possible due to the fact that the patients were tested for COVID-19 using different laboratory tests for IgG/IgM detection and diagnostic tests employing RT-PCR. The second classification pertained to the severity of the disease. Positive samples from patients who were admitted to the hospital were classified as 1, indicating severe COVID-19. Samples from non-hospitalised patients were classified as 0, indicating mild to moderate COVID-19. By segregating only positive samples for COVID-19, a file was generated with 35,109 samples, comprising 27,390 severe and 7719 mild to moderate cases (Figure 1).

Following the positioning of the biomarkers, the file exhibited a lack of data. This was because the corresponding laboratory tests (biomarkers) were not conducted on the same date and for the same patient in the same line (sample). To resolve this issue, PyCaret (version 3.3.2) [23], an open-source machine learning library developed with the computer programming language called Python [24], was employed. PyCaret allows the data to be prepared for machine learning model deployment. The missing data were completed using mean calculation (for numeric imputation) as suggested by Pycaret [23]. In scientific health studies, this issue of missing data may arise, and a way to complete missing data is to use the mean of the variable [25].

The PyCaret machine learning library includes functionalities for dealing with data balancing during the training of machine learning models, offering methods for dealing with unbalanced datasets, which is important for avoiding bias in the resulting models [23]. PyCaret performs data balancing by oversampling (increasing the minority class) and undersampling (decreasing the majority class) to balance the distribution of classes in the dataset [23]. PyCaret separates training and test data. Pre-processing takes place within a scikit-learn pipeline. So any transformation, normalisation or imputation is only performed on the training data before being applied to the test or cross-validation data. This ensures that all transformations are applied consistently and without data leakage between training and testing. By default, PyCaret uses a decision threshold of 0.5 for probabilistic classifiers [23].

A source code (https://github.com/COVID-19-Severity, accessed on 18 September 2024) was developed in Python with the use of PyCaret machine learning library to analyse the treated file [26]. Phase 7 of the Python script shows the use of the Fold Generator as StratifiedKFold and, Phase 8 show the function compare_models() [26]. PyCaret automatically performs cross-validation during various stages of the machine learning workflow. When using functions like compare_models(), PyCaret trains and evaluates multiple models using k-fold cross-validation by default (10 folds). Cross-validation in PyCaret is primarily performed on the training set, while a separate hold-out test set is maintained for a final, unbiased evaluation of the selected model. This provides a robust assessment of model performance across different subsets of the training data [23].

For this study, the standard PyCaret library hyperparameters were used, such as: Imputation = simple, Numerical Imputation = average, Fold Generator = KFold stratified and Number of folds = 10, so the hyperparameters were not adjusted. The machine learning models were trained under homogeneous conditions with the same set of features and balancing techniques. Phase 7 of the Python script developed (https://github.com/COVID-19-Severity/DataSet/blob/main/COVID19_Severity.ipynb.pdf, accessed on 18 September 2024) shows how to carry out reproducibility and fixes the data split (session_id = 8801). The script is in the Supplementary Materials (File S1).

In order to illustrate the entire process in an appropriate manner, the aforementioned steps can be summarised as follows: (1) The results of tests on patients who were either positive or negative for the novel coronavirus were selected from a database called “HOSPITAL”. (2) The second file (dataset), designated “PROGNOSIS”, consisted of positive and negative samples positioned side by side. (3) The third dataset, comprising solely positive samples, was classified by “SEVERITY”. The dataset is in the Supplementary Materials (File S2). (4) In the phase of severity data preprocessing of the dataset, features that were deemed unnecessary were removed, including patient identification, patient birth, sample result date, and sample requesting unit. Following data processing, the final number of features defined in the dataset was 2 anthropometric features and 48 laboratory biomarkers. (5) For the exploratory analysis, all biomarkers were considered and compared graphically using boxplot to visualise the variation in biomarker levels between samples regarding the severity of COVID-19 (severe; non-severe = mild to moderate). (6) In the data division stage, for training and testing the machine learning models, 24,576 (70%) of the training samples and 10,533 (30%) of the testing samples were used. The total set comprised 35,109 samples. (7) Fifteen machine learning models were trained and tested on the dataset to validate the machine learning. (8) The main biomarkers for prognosing the severity of COVID-19 were identified through graphs. (9) Ultimately, following the performance assessment stage utilising metrics such as accuracy, the Light Gradient Boosting Machine (LGBM) was identified as the optimal performance model for the sample design of this study (Figure 1).

The Boosting technique, which is present in LGBM, has been demonstrated to improve classification accuracy. This is achieved by using multiple copies of the classifier to learn about the dataset interactively. To classify a result, a weight is given to each prediction of these classifiers, and then the final prediction is calculated. A similar process occurs during the learning phase. Each training tuple, which is made up of a die and its class, has a weight. Once the initial classifier has completed its learning process, the weights are adjusted to enhance focus on the tuples that have been misclassified by that classifier. Consequently, the subsequent classifier has a greater probability of achieving accurate classification than the initial classifier, and so on. The final classifier combines the votes of each classifier individually, where the weights of each classifier are a function of its accuracy. Such a classifier employs the LGBM algorithm and performs the learning and classification steps on features extracted from the given dataset [27].

For the task of classifying the severity of COVID-19 based on biomarkers, the Light Gradient Boosting Machine (LightGBM) model, an optimised implementation of the gradient boosting algorithm based on decision trees, was identified as ideal for this study. The characteristics of the LightGBM are its efficiency in modelling tabular data, its ability to capture complex non-linear relationships between variables and its robustness to missing values, dispensing with artificial imputations [28].

Another determining factor observed in this study for its choice was its ability to deal with unbalanced data, a common feature in clinical studies, by applying adjustable weights to classes and optimising specific metrics. In addition, the model incorporates structural regularisation techniques, such as restricting the depth of trees and controlling the minimum number of samples per leaf, helping to reduce the risk of overfitting.

3. Results

The anthropometric features and laboratory measurements of the present study’s population are presented in the Supplementary Materials. The Sex feature is represented by the values 1, 2, which correspond to female and male, respectively. The median Age feature for females in both groups (mild to moderate COVID: 52; severe COVID: 55) was lower than for males (mild to moderate COVID: 56; severe COVID: 57). The median Erythrocyte count was found to be lower for individuals with severe COVID (3.42 × 10⁶/µL for females and 3.57 × 10⁶/µL for males) than for those with mild to moderate COVID (4.20 × 10⁶/µL for females and 4.25 × 10⁶/µL for males). Further laboratory measurements can be observed in the Supplementary Materials (File S3), where descriptive statistics and box plot graphics are presented.

To assess the performance of the machine learning models used in this study, various statistical metrics widely accepted in the literature were employed. Accuracy was used as an initial metric, representing the proportion of correct classifications in relation to the total number of predictions made; however, it can be limited in unbalanced datasets. In order to assess the discriminative capacity of the models, the Area Under the ROC Curve (AUC-ROC) was also used, which measures the classifier’s ability to distinguish between classes, with values closer to 1 indicating better performance. To complement the analysis, precision was calculated, which indicates the proportion of correct positive predictions, and recall (or sensitivity), which measures the proportion of correctly identified positive instances. Both metrics are essential for evaluating performance in tasks with differentiated costs between false positives and false negatives. In order to balance precision and recall, the F1 score was used, defined as the harmonic mean between these two metrics, which is particularly useful in situations where there is an imbalance between classes. In addition, Cohen’s Kappa index was calculated, which assesses the degree of agreement between the model’s predictions and the actual labels, correcting for chance. Finally, the Matthews Correlation Coefficient (MCC) was adopted, which is considered to be a robust and balanced measure, even in scenarios with unequal distribution between classes, offering a general assessment of the quality of the classification [29].

Table 2. Metrics employed in the construction of machine learning models.

Model	Complete Name	ACC	AUC	Recall	Prec.	F1	Kappa	MCC	TT (s)
lightgbm	Light Gradient Boosting Machine	0.8808	0.9177	0.8808	0.8757	0.8757	0.6250	0.6314	6.9560
xgboost	Extreme Gradient Boosting	0.8796	0.9153	0.8796	0.8747	0.8751	0.6247	0.6297	0.7510
rf	Random Forest Classifier	0.8786	0.9060	0.8786	0.8732	0.8725	0.6136	0.6220	5.2900
et	Extra Trees Classifier	0.8779	0.9047	0.8779	0.8726	0.8714	0.6098	0.6191	4.7040
gbc	Gradient Boosting Classifier	0.8636	0.8939	0.8636	0.8572	0.8525	0.5460	0.5645	5.7600
ada	Ada Boost Classifier	0.8481	0.8744	0.8481	0.8383	0.8375	0.5024	0.5153	1.3780
dt	Decision Tree Classifier	0.8144	0.7358	0.8144	0.8168	0.8155	0.4656	0.4658	0.5530
lda	Linear Discriminant Analysis	0.8087	0.7924	0.8087	0.7868	0.7794	0.3057	0.3401	0.2300
lr	Logistic Regression	0.8031	0.7783	0.8031	0.7785	0.7668	0.2619	0.3045	4.4330
ridge	Ridge Classifier	0.8029	0.7927	0.8029	0.7818	0.7587	0.2326	0.2909	0.1790
knn	K Neighbors Classifier	0.7882	0.6947	0.7882	0.7582	0.7616	0.2545	0.2740	1.0320
dummy	Dummy Classifier	0.7802	0.5000	0.7802	0.6086	0.6838	0.0000	0.0000	0.1610
svm	SVM—Linear Kernel	0.5547	0.5315	0.5547	0.7456	0.4531	0.0113	0.0414	1.0880
qda	Quadratic Discriminant Analysis	0.4920	0.7272	0.4920	0.7780	0.4170	0.0281	0.0930	0.3640
nb	Naive Bayes	0.2816	0.7064	0.2816	0.7608	0.2069	0.0263	0.0890	0.1780

The performance metrics for the optimal machine learning algorithms for sample studies in this work were obtained through the script developed in the Python programming language, Phase 8: Build and compare models, which can be found at https://github.com/COVID-19-Severity, accessed on 18 September 2024 [26]. The performance metrics, expressed as decimals, of the machine learning algorithms are presented in the following order: Accuracy (ACC), Area Under the receiver operating characteristic Curve (AUC), proportion of true positives, Precision, F1 score (the harmonic mean of the Precision and Recall), Kappa coefficient (a statistical method for evaluating the level of agreement or reproducibility between two sets of data), scalar metric for binary classification: the Matthews Correlation Coefficient (MCC), and the average training time and testing time, in seconds, of machine learning algorithms (Time Take—TT) are also considered.

presents the percentages for the Accuracy, AUC (Area Under the Curve) of the ROC, Recall, Precision, F1 score, Kappa, and Matthews Correlation Coefficient (MCC) for 15 machine learning models and the metric Time Taken (TT) in seconds (s). Table 2 highlights in yellow the best performances for the metrics used for the machine learning models.

From the metrics presented in the table above, converting decimal values to percentages, it can be observed that the first five machine learning models exhibit a high degree of similarity in terms of Accuracy (between 86% and 88%), Area Under the Curve (between 89% and 92%), Recall (between 86% and 88%), Precision (between 86% and 88%), F1 score (between 85% and 88%), Kappa (between 55% and 63%), and Matthews Correlation Coefficient (MCC) (between 56% and 63%), as well as in terms of Test Time (TT) (between 0.2 and 6.9 s).

The ROC curve is employed to assess the accuracy of a test, with a value of 1.0 on the Y axis (True Positive Rate) signifying 100% Specificity and a value of 1.0 on the X axis (False Positive Rate) indicating 100% Sensitivity (Figure 2). A value below 0.5 indicates that the test lacks diagnostic capacity. The larger the Area Under the Curve, the more accurate the test, the greater its diagnostic capacity, and the more sensitive and specific it is. Values exceeding 0.9 are considered to be excellent, while values exceeding 0.8 are regarded as good. Values exceeding 0.7 are considered to be adequate [29].

The diagonal line representing random guessing is also shown as an example in Figure 2. The closer the ROC curve of a model is to the diagonal line, the less accurate it is. If the model is good, it is more likely to encounter true positives as it moves down the ranked list. Thus, the curve moves steeply up from zero. As the number of true positives decreases and the number of false positives increases, the curve flattens and becomes more horizontal. One method for assessing the accuracy of a model is to measure the area under the curve. Several software packages can perform this calculation. The closer the area is to 0.5, the less accurate the corresponding model is. A model with perfect accuracy has an area of 1.0 [29].

Figure 2 illustrates the Receiver Operating Characteristic (ROC) curves of five classification machine learning models that demonstrated superior performance for the samples under investigation.

The study revealed that the Area Under the ROC Curve between the five most effective machine learning models (LGBM, XGBoost, Random Forest, Extra Trees, Gradient Boosting) was remarkably consistent, with values ranging from 90% to 92% for classes 0 (Mild to Moderate COVID) and 1 (Severe COVID).

The values for the sensitivity calculations and other indicators were obtained through the use of a confusion matrix, as exemplified in Appendix A, in

Table A1. The structure of the Confusion Matrix employed is as follows.

Results	Confusion Matrix
	Positive	Negative
Positive	a True Positive	b False Positive
Negative	c False Negative	d True Negative

a = True Positive, results considered as “Severe COVID-19”. b = False Positive, results considered as “Severe COVID-19”. c = False Negative, results considered as “Mild to Moderate COVID-19”. d = True Negative, results considered as “Mild to Moderate COVID-19”. Sensitivity = a/(a + c). Specificity = d/(b + d). False Positive or Positive Predictive Value = a/(a + b). False Negative or Negative Predictive Value = d/(c + d). Accuracy = a + d/(a + b + c + d).

.

Table 3. Confusion Matrix presents the five most effective machine learning models for the samples under study.

Machine Learning Model	Severity Classification Based of the Samples
LGBM (n = 10,533)	(0) Mild to Moderate (20%)	(1) Severe (80%)
(0) Mild to Moderate	1462 (14%)	854 (5%)
(1) Severe	362 (6%)	7855 (75%)
XGBoost (n = 10,533)	(0) Mild to Moderate (18%)	(1) Severe (82%)
(0) Mild to Moderate	1475 (14%)	841 (8%)
(1) Severe	425 (4%)	7792 (74%)
Random Forest (n = 10,533)	(0) Mild to Moderate (17%)	(1) Severe (83%)
(0) Mild to Moderate	1422 (13%)	894 (9%)
(1) Severe	387 (4%)	7830 (74%)
Extra Trees (n = 10,533)	(0) Mild to Moderate (16%)	(1) Severe (84%)
(0) Mild to Moderate	1396 (13%)	920 (9%)
(1) Severe	364 (3%)	7853 (75%)
Gradient Boosting (n = 10,533)	(0) Mild to Moderate (15%)	(1) Severe (85%)
(0) Mild to Moderate	1213 (12%)	1103 (10%)
(1) Severe	326 (3%)	7891 (75%)

The number n = 10,533 samples assigned to each machine learning model was defined by the script, based on the 30% of samples used for testing. Phase 10 involved extracting the metrics results from the five top models, which can be found at https://github.com/COVID-19-Severity, accessed on 18 September 2024 [26].

presents the percentages of the confusion matrix for the machine learning models. The values 1 and 0 correspond to the severity levels of “Severe” and “Mild to Moderate” for COVID-19, respectively.

The parameters identified for the studied samples are presented in

Table 4. Comparison of the parameters in machine learning models.

Parameters (%)	LGBM	XGBoost	Random Forest	Extra Trees	Gradient Boosting
Sensitivity	80%	78%	79%	79%	79%
Specificity	90%	90%	90%	89%	88%
Positive Predictive Value	63%	64%	61%	60%	52%
Negative Predictive Value	95%	95%	95%	96%	96%
Accuracy	88%	88%	88%	88%	86%
Matthews Correlation Coefficient	64%	63%	62%	62%	57%

The five machine learning models yielded comparable parameter values. To illustrate, the optimal machine learning model (LGBM) exhibited 80% Sensitivity, 90% Specificity, 88% Accuracy, 63% Positive Predictive Value, 95% Negative Predictive Value, and 64% Matthews Correlation Coefficient. The Matthews Correlation Coefficient is a quality measure of two binary classifications that can be used even if the classes have quite different sizes. It returns a value between (−1) and (+1), where a coefficient of (+1) represents a perfect prediction, (0) represents an average random prediction, and (−1) represents an inverse prediction. This statistic is equivalent to the phi coefficient and attempts, like efficiency, to summarise the quality of the contingency table in a single comparable numerical value: phi = (VP × VN – FP × FN)/sqrt((VP + FP) × (VP + FN) × (VN + FP) × (VN + FN)). It should be noted that if any of the sums in the denominator is equal to zero, the denominator can be considered to be one, resulting in a phi equal to zero, which would be the correct limit for this situation.

, along with the comparative performance parameter results between the top five machine learning models.

Feature importance refers to techniques that calculate a score for all input features of a given model. The scores represent the relative importance of each feature. A higher score indicates that the specific feature will have a greater effect on the model being used to predict a given variable [30].

After running our Python script, the five most important biomarkers for the prognosis of COVID-19 severity, as determined by the samples under study, were: Platelets, Creatinine, Erythrocytes, C-reactive Protein and, Lymphocytes (Figure 3).

The Age feature is essential when determining the prognosis of a patient infected by COVID-19. The selection of the dataset for COVID-19 was guided by clinical relevance and the availability of epidemiological records, concentrating on crucial attributes such as age and sex. These factors are known to influence the severity and mortality of COVID-19 [31].

The Age feature is a well-established predictor of severity in COVID-19 and, according to the graph, has greater importance than biomarkers (Figure 3).

Age is a proven predictor of the severity of COVID-19 [31]. In order to explore the effect of age on the LGBM model that was selected in our study, the model’s performance was evaluated on a sample that had been stratified by age groups. The treatment of the selected age groups was consistent with that applied to the total sample. The results of the study are presented in

Table 5. LGBM model performance across age groups to classify COVID-19 severity.

Age Groups	n	AUC (%)	Sensibility (%)	Specificity (%)	Accuracy (%)	PPV (%)	NPV (%)
0–17 y	832	91	82	87	85	72	92
18–29 y	852	91	77	90	87	69	93
30–39 y	1025	91	75	90	87	63	94
40–49 y	1295	93	83	91	90	70	96
50–64 y	3326	91	76	91	89	62	95
65–74 y	1975	90	75	90	87	62	94
75–84 y	1028	88	71	90	88	47	96
>85 y	197	87	76	89	87	63	94

n, sample size (10,533), represents 30% of the test sample. Age in years (y). AUC, area under de curve. PPV, positive predictive value. NPV, negative predictive value.

.

The LGMB model demonstrated a high degree of accuracy for the unstratified sample, with an accuracy of 88% (see Table 4). This level of accuracy is comparable to that observed in the age-stratified sample (see Table 5), where a minimum of 85% accuracy was achieved across all age groups. Additionally, the negative predictive value (NPV) exhibited values of ≥92% across all age groups.

In our study using machine learning models with biomarkers, we found that the presence of platelets, creatinine, erythrocytes, C-reactive protein, lymphocytes, albumin, glucose, urea, and sodium can be associated with the diagnosis of COVID-19 and the severity of the disease (Figure 3). These biomarkers represent potential therapeutic targets that should be investigated more effectively in future clinical trials.

4. Discussion

Since the beginning of the coronavirus pandemic, it has been noted that older people are more susceptible to severe forms of COVID-19. In this study, the Age feature occupies the first place of influence in the predictive model, supporting its direct relationship with the probability of hospitalisation due to physiological ageing and the accumulation of comorbidities. Our study confirms the importance of the Age feature and is in line with previous studies that associate this condition with a higher risk of hospitalisation [31].

4.1. Biomarkers

The assessment of laboratory biomarkers has proved to be a fundamental tool in supporting clinical decision-making in COVID-19 patients, both for risk stratification and for the early detection of clinical deterioration. Among the main biomarkers used are those related to the inflammatory response, renal function, hydroelectrolyte balance and energy metabolism.

The severe respiratory syndrome of the COVID-19, caused by the novel virus SARS-CoV-2, has been shown to be associated with an exaggerated immune response, which has been described as a “cytokine storm” [32]. The intense inflammatory process has been demonstrated to affect multiple organs, consequently altering various laboratory biomarkers.

The present study captured records of multiple laboratory parameters that are utilised in a hospital environment (see Table 1). These parameters are frequently employed in patients considered to be at high risk.

The laboratory parameters employed in the context of the pandemic vary according to the technological and financial resources available at the institutions responsible for patient treatment. Another factor that influences the selection of laboratory biomarkers is the evolving knowledge of the pathophysiology of the disease.

The laboratory parameters selected in this study are similar to those described for other referral centres treating patients with the disease [33,34].

As illustrated in Figure 3, the variables that were most relevant as predictors of severity in relation to the novel strain of the virus were identified in the present study. These biomarkers are associated with different aspects of the response to pathophysiological changes resulting from the action of SARS-CoV-2. The following discussion and presentation will therefore be in decreasing order of importance in the models studied.

As demonstrated in a plethora of studies, age has been identified as a pivotal factor associated with the severity and mortality of cases of Coronavirus Disease 2019 (henceforth referred to as “COVID-19”). As demonstrated in numerous studies, age has been identified as a pivotal factor in the severity and mortality outcomes associated with COVID-19 [35].

For instance, in a publication by the Centers for Disease Control and Prevention (CDC), age is identified as the strongest factor associated with severe outcomes from the disease [36].

In this study, the risk of contracting, being hospitalised for, and succumbing to, the effects of the novel coronavirus (SARS-CoV-2) increases exponentially with age. To elaborate, the probability of infection, hospitalisation, and death is 25-fold higher for the 18–29 age group than for those aged 50–64, and 60-fold higher for those aged 65–74 [36].

Platelets play an essential role in coagulation and the inflammatory response. Patients with COVID-19 often have thrombocytopenia, which has been associated with greater disease severity and an increased risk of thromboembolic events [37]. The reduction in platelets may reflect exacerbated consumption due to systemic inflammation and coagulopathy.

Creatinine and urea are classic markers of kidney function. Elevation of these biomarkers in COVID-19 patients indicates possible acute kidney injury, which is a common finding in severe cases, resulting from both direct effects of SARS-CoV-2 on the kidneys and systemic hypoperfusion [38]. Monitoring these parameters is essential for therapeutic adjustments, including the management of fluids and nephrotoxic drugs.

Erythrocytes (red blood cells that transport haemoglobin, an oxygen carrier) may be reduced in some patients, reflecting chronic inflammatory processes or haemolysis. Anaemia, when present, can compromise tissue oxygenation and is an additional risk factor for respiratory complications [39].

C-reactive protein (CRP) is a sensitive and nonspecific marker of systemic inflammation. Its elevated levels have been consistently associated with the severity of COVID-19, reflecting a cytokine storm and correlating with adverse outcomes such as the need for mechanical ventilation and mortality [40].

Lymphocytes tend to be reduced (lymphopenia) in critically ill patients, possibly due to immune exhaustion and direct cytotoxic effects of the virus. Lymphopenia has prognostic value and can be used as a screening criterion for hospitalisation or more intensive monitoring [41].

Albumin, synthesised by the liver, with several relevant physiological functions (control of osmotic pressure, a carrier for various molecules in the blood, among others) is also plasma protein with anti-inflammatory and antioxidant functions. Its reduction (hypoalbuminaemia) is a marker of poor clinical evolution and may reflect systemic inflammation or liver dysfunction [42]. Low albumin levels are associated with higher hospital mortality.

Elevated glucose (hyperglycaemia), even in non-diabetic patients, has been observed in association with worse outcomes in COVID-19, possibly as a result of metabolic stress and exacerbated inflammatory response [43]. Strict glycaemic control has been recommended to reduce complications.

Sodium in serum is an indicator of hydroelectrolyte balance, and disorders such as hyponatraemia have been reported in patients with severe COVID-19, possibly due to inappropriate secretion of antidiuretic hormone or gastrointestinal fluid loss [44]. Changes in sodium levels can worsen the clinical state, requiring careful management.

Thus, combined monitoring of these biomarkers can help with risk stratification, prioritisation of resources (such as ICU admission), and the implementation of early therapeutic interventions. The integrated use of these laboratory parameters reinforces the personalised clinical approach to COVID-19, allowing for timely interventions that can have a positive impact on clinical outcomes.

When evaluated in context, the machine learning models demonstrated the capacity to identify routine laboratory biomarkers that are relevant for discriminating severe cases of COVID, a finding that is consistent with the findings of other studies and with the established pathophysiological changes associated with viremia.

4.2. Comparison with Previous Studies

The utilisation of Haematochemical values derived from routine blood tests may represent an alternative approach to the early detection and cost-effective diagnosis of COVID-19. This can be achieved through the development, evaluation and validation of machine learning models based on routine blood tests. In the present study, the values were derived from routine laboratory tests (n = 35,109), as previously outlined in Table 1. In order to assess the severity of the novel coronavirus, the Light Gradient Boosting Machine (LGBM) model was utilised, with the most optimal performance being demonstrated (80% sensitivity, 90% specificity and 88% accuracy) (see Table 2 and Table 3). The machine learning models XGBoost, Random Forest, Extra Trees and Gradient Boosting also demonstrated satisfactory performance (see Table 4). The LGBM model enabled the classification of the severity of patients admitted to hospitals with confirmed cases of the disease, with a high negative predictive value of 95% (see Table 4).

Cabitza et al. [45] utilised three distinct training datasets comprising Haematochemical values from 1624 patients (52% of whom were positive for COVID-19) admitted to San Raphael Hospital (OSR) between February and May 2020 to develop machine learning models. The full OSR dataset comprised 72 features. The dataset comprised Complete Blood Count (CBC), Biochemistry, Coagulation, Blood Gas Analysis and CO-Oximetry values, age, sex and specific symptoms at triage. Two sub-datasets were also created: a COVID-specific sub-dataset (32 features) and a CBC sub-dataset (21 features). A total of 58 cases (50% of which were positive for COVID-19) were collected from another hospital and 54 negative patients were included in the study in 2018 at OSR. These cases were used for internal-external and external validation. Five machine learning (ML) models were developed, namely Random Forest (RF), Naive Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). The Area Under Curve (AUC) of the Receiver Operating Characteristic (ROC) of the algorithms for the full OSR dataset ranged from 83% to 90%. For the COVID-specific dataset, the AUC of the ROC ranged from 83% to 87%. Finally, for the CBC dataset, the AUC of the ROC ranged from 74% to 86%. The validations also yielded satisfactory outcomes, with the respective AUC of ROC reaching 75% to 78% and specificity reaching 92% to 96% [45]. In a manner analogous to the present study, machine learning (ML) models exhibited capacity for discrimination when employed with disparate laboratory data pertaining to the diagnosis of COVID-19.

In our study, we utilised values derived from routine laboratory tests, as previously outlined in Table 1. The Light Gradient Boosting Machine (LGBM) model demonstrated the most optimal performance, with 80% sensitivity and 88% accuracy, closely followed by other models with nearly identical results: eXtreme Gradient Boosting (XGBoost), Random Forest Classifier, Extra Trees Classifier and Gradient Boosting Classifier.

Also using ML techniques, Albuquerque Paiva de Oliveira et al. [14], namely Random Forest (RF), Multi-Layer Perceptron (MLP) and Support Vector Machines (SVM) Regression, to construct disease prediction models utilising clinical parameters. The researchers verified the existing correlations between the laboratory parameters and the results of the COVID-19 test and developed two classification models. The first classified the test results for patients suspected of having COVID-19, while the second classified the hospitalisation units of patients with COVID-19, both according to the laboratory parameters. The models achieved an accuracy above 96%, indicating that they are promising for the classification of COVID-19 tests and the screening of patients by inpatient unit [14]. The authors of the study in question have reinforced the ability of machine learning (ML) techniques to diagnose cases of the novel coronavirus, severe acute respiratory syndrome (SARS-CoV-2), with laboratory data, a finding similar to that of our study.

Like Cobre et al. [46], our study approached the objectives of the work in a similar manner. The researchers applied machine learning (ML) models to predict the severity of the novel coronavirus disease (COVID-19) as prognostic indicators of the pathological process. The study employed four machine learning-based models: Artificial Neural Networks (ANN), Decision Trees (DT), Partial Least Squares Discriminant Analysis (PLS-DA), and K-Nearest Neighbours. These models were trained on more than 5000 RT-PCR samples and data on biochemical parameters, haematological and urinary tests of patients. The models were able to effectively predict the diagnosis of COVID-19 and its severity in Brazil. The study found that ML-based models (ANN, DT, PLS-DA, and KNN) were able to effectively predict the diagnosis and disease severity of COVID-19 with an accuracy of above 84%. This is comparable to the results obtained by RT-PCR and the minimum threshold recommended for diagnostic tests. ANN was the model with the best performance, achieving more than 94% accuracy, and could therefore be used as a decision support tool for health professionals in practice [46]. The study by Cobre et al. [46] differs from the present study in that it utilised solely RT-PCR results, yielding approximately 550 positive results, constituting 10% of the total sample. The present study employed a range of diagnostic tests, including immunological and RT-PCR methods, in addition to a sample size approximately seven times larger than previous investigations. It is important to note that the machine learning (ML) models employed were able to classify criteria associated with the novel coronavirus (SARS-CoV-2) across laboratory records and various diagnostic tests, which provides consistency to these models.

Kukar et al. [47] constructed a machine learning model for the diagnosis of COVID-19, which was validated using routine blood tests on 5333 patients with various bacterial and viral infections, as well as 160 patients with confirmed COVID-19. The operational ROC point was selected with a sensitivity of 82% and specificity of 98%. The cross-validated AUC was 97%. The five most useful routine blood parameters for the diagnosis of COVID-19, according to the XGBoost algorithm feature importance score, were: MCHC, Eosinophil count, Albumin, International Normalized Ratio (INR, blood test to measure blood clotting time) and percentage of Prothrombin activity. As a result, patients presenting with fever, cough, myalgia, and other symptoms can now have routine blood tests evaluated by the diagnostic tool created by the authors [47]. The study reports that the ML model, XGBoost, which was also evaluated in the present study, is important in the differential classification of infectious processes and of cases of the disease known as “COVID-19”, even in instances where only a small number of cases of “COVID-19” are present in the sample (n = 160). The XGBoost model in our study demonstrated a comparable performance to the LGBM model that was selected for use. The proposal of evaluating multiple ML models in a concurrent manner is a rational approach, given that these models have the capacity to discern subtle variations in the groups to be classified through diverse methodologies.

Huyut, Velichko, and Belyaev conducted a study with the objective of identifying routine blood value predictors of mortality from COVID-19 and determining the lethal risk levels of these predictors during the disease process. The study dataset comprises 38 routine blood values from 2597 patients who died (n = 233) and those who recovered (n = 2364) from COVID-19 between August and December 2021. In this study, the model of Histogram-based Gradient Boosting (HGB) was the most successful machine learning classifier in detecting living and deceased patients with COVID-19 (with squared F1 metrics F1² = 1). The most efficient binary combinations with Procalcitonin were obtained with D-Dimer, erythrocyte sedimentation rate (ESR), direct bilirubin (D-Bil) and Ferritin. The XGBoost method was the most successful model in diagnosing the disease, with an 85% sensitivity and 90% accuracy [48]. In this study, machine learning (ML) models exhibited a capacity for prediction in relation to the outcomes of patients infected with the novel coronavirus (SARS-CoV-2), even in instances where the case group (mortality) comprised a limited sample size. The HGB and XGBoost models, the latter of which was utilised in the present study, are not identical, but they do share numerous concepts. It can be inferred that machine learning (ML) models with the HGB and XGBoost frameworks have the potential to discriminate outcomes based on laboratory data, even in the presence of smaller sample sizes.

A further study, involving patients over the age of 18 with non-COVID-19 and COVID-19 pneumonia who required hospitalisation in a tertiary hospital, developed a machine learning model to classify these two diseases with similar clinical characteristics. The performance of the models based on laboratory tests was excellent. Several potentially important laboratory tests for differential diagnosis were identified, including: The study identified several laboratory tests that demonstrated potential for differentiating the two disease groups, including D-Dimer, Eosinophil, Glucose, Aspartate Aminotransferase and Basophil. In terms of performance, the LGBM model demonstrated the highest Area Under the ROC Curve (99%), followed by XGBoost (98%) and Random Forest (97%). In terms of Accuracy, XGBoost and LGBM achieved a score of 95%, followed by Random Forest with a score of 94%. In terms of the F1 score, which has been identified as an important performance index in studies with data imbalance (non-COVID-19: 769; COVID-19: 296), LGBM achieved the highest score of 94%, followed by XGBoost (93%) and Random Forest (93%) in that order [49]. Like Baik, Hong, Park [49], our study corroborates the hypothesis that the machine learning (ML) models utilised in this study also demonstrated excellent performance in the classification of other outcomes associated with the ongoing global pandemic of the novel severe acute respiratory syndrome (SARS-CoV-2) virus. The findings of our study are in accordance with the conclusions drawn in the aforementioned study, which also confirmed that the LGBM and XGBoost models are the most effective performers.

Finally, the data from patients diagnosed with COVID-19 were collected by Çubukçu et al. [50] between April 2020 and November 2020. This dataset consisted of results from concomitant Clinical Chemistry (CC) and Complete Blood Count (CBC) tests, which were recorded within a 48 h window of the RT-PCR test. The parameters of the CC test included Total Bilirubin, Alanine Aminotransferase (ALT), Aspartate Aminotransferase (AST), Lactate DeHydrogenase (LDH), Creatinine, and C-Reactive Protein (CRP). The Complete Blood Count (CBC) parameters included Eosinophils, Monocytes, Lymphocytes, Red Blood Cell distribution width, Platelets, Mean Corpuscular Haemoglobin, Leukocytes, Mean Corpuscular Volume, Mean Corpuscular Haemoglobin concentration, Neutrophils, Haemoglobin, Haematocrit, Basophils, and Red Blood Cells. The objective of the study was to develop a clinical decision support tool to assist in the diagnosis of COVID-19 disease using Machine Learning (ML) models, based on results from routine laboratory tests. ML models were developed using laboratory data (n = 1391), comprising six Clinical Chemistry (CC) results, 14 CBC parameter results, and results related to the presence of COVID-19. Four machine learning (ML) algorithms, including Random Forest (RF), gradient boosting (XGBoost), Support Vector Machine (SVM), and Logistic Regression, were employed to construct eight ML models utilising Complete Blood Count (CBC) and a combination of CBC and Clinical Chemistry (CC) parameters. Performance evaluation was conducted on the test dataset and the Brazilian validation dataset. The accuracy values of all models ranged from 74% to 91% [50]. In this study, the maximum positive predictive values (PPV) were observed in the machine learning (ML) models for predicting the novel corona virus (SARS-CoV-2) infection, namely 90.4% and 91.8%, as obtained by the random forest and XGBoost algorithms, respectively. The present study is pertinent in the context of ours, given that it employs analogous machine learning (ML) models with a database of the same ethnicity, i.e., Brazilians. The aforementioned machine learning (ML) models demonstrated a high degree of efficacy in predicting the course of the disease using laboratory data.

The RF model trained from CC and CBC analytes demonstrated the most optimal performance on the dataset of the present study, with an Accuracy of 85%, a Sensitivity of 80%, and a Specificity of 91%. The RF model trained only from CBC parameters was able to detect COVID-19 cases with an accuracy of 83%. The SVM model trained with CC and CBC parameters demonstrated the most optimal performance on the external validation dataset, with an accuracy of 91%, a sensitivity of 100%, and a specificity of 84%. The ML models presented in this study can be employed as clinical decision support tools to inform doctors’ clinical judgments regarding the diagnosis of COVID-19 [50].

Similarly, the ML models identified (LGBM, XGBoost, RF, Extra Trees, Gradient Boosting), trained and tested in our study, can also be utilised as clinical decision support tools to assess the severity of COVID-19.

A further salient point pertains to the observation that a significant proportion of studies under scrutiny present considerably diminished sample sizes when juxtaposed with the present study. This observation serves to reinforce the validity of the conclusions drawn. It is important to note that in numerous studies, including our own, specificity and negative predictive value (NPV) exhibited the highest values, typically exceeding 90%. The ML models evaluated, therefore, more efficiently capture those who do not present the proposed outcome. In the context of the management of patients infected with SARS-CoV-2, the utilisation of a tool that can predict those at lower risk, or with a high negative predictive value (NPV), holds considerable potential in circumstances where intensive care unit (ICU) beds and associated resources are in short supply. The our study addresses precisely this point, considering the severity of the ongoing global pandemic of the novel coronavirus (SARS-CoV-2), a factor in patient classification and directly related to the management of intensive care unit (ICU) bed distribution, adoption of more aggressive therapeutic interventions, and, in general, maximisation in the management of high-risk patients for the health service.

The effects of the novel virus known as severe acute respiratory syndrome (SARS-CoV-2) are not limited to any specific age group. The infection fatality rate demonstrates a marked increase in mortality with advancing age [35].

Models that aspire to encapsulate the impact of the severity of the novel Coronavirus (SARS-CoV-2) pandemic, as evidenced in this study, are required to demonstrate responsiveness across all age demographics. The LGBM model that demonstrated efficacy in our study was capable of capturing the severity of the novel strain of the virus across a broad spectrum of age demographics (see Table 5). The model exhibited comparable specificity and sensitivity rates to those observed in the unstratified sample (see Table 4). It is important to note that the negative predictive value (NPV) exceeds 90% across all age groups. This characteristic serves to reinforce the robustness of the ML model under scrutiny, thereby expanding its potential for discriminatory applicability.

Similarly, numerous studies have indicated a correlation between serum Glucose and Eosinophil levels and the severity of COVID-19 [51,52,53,54]. However, the existing literature lacks highly specialised and directed studies employing machine learning models for the purpose of diagnosing the severity of COVID-19. This work aims to address this gap in the literature.

4.3. Limitations

This study has some limitations. All data are from a single hospital in southern Brazil. This limits the generalisability of the results to other populations.

The sample for the present study was obtained from a single university hospital in southern Brazil, and the institution does not require patients to pay any treatment costs (item 2.1 study location). The subjects of the study are characterised by a low income and a high prevalence of comorbidities, including nutritional problems. It is important to note that comparisons with studies conducted in other hospitals, where patient profiles may differ, may introduce biases or yield different results.

The demographic composition of Brazil is characterised by a trihybrid admixed population, encompassing individuals of Caucasian, Black, and Indigenous descent. Comparisons with studies including more homogeneous populations in high-income countries may reveal discrepancies, either due to differences in the standard of care with more resources, or even intrinsically due to ethnic differences associated with viral response.

Patients were categorised into two groups (mild to moderate = non-ICU admission and severe = ICU admission). Clinical or administrative criteria were not used to establish this classification, which represents a methodological limitation. The categorisation was based on COVID-positive patients and ICU admission or not. The study did not select data on the need for supplementary oxygen, specific physiological parameters, institutional protocols or official criteria from the Brazilian Ministry of Health.

The criterion applied to classify severity in this study was simplified (binary), based only on hospitalisation (yes/no), and did not fully represent the clinical complexity of the disease. This simplification can introduce biases if the decision to hospitalise is influenced by non-clinical factors, such as system saturation or administrative decisions.

Among the limitations of the study, we highlight the use of data imputation and automated balancing via PyCaret. Imputation can introduce biases by altering original distributions and reducing variability, especially when assumptions about the absence mechanism are not met. Although automated balancing reduces the impact of class imbalance, it can generate artificial instances with low representativeness or eliminate relevant information, affecting the generalisation of the model. Additionally, the automation of pre-processing steps may limit methodological flexibility and obscure critical decisions, requiring caution in the interpretation of results.

Furthermore, this study did not use explainable visualisation techniques such as SHAP and LIME to analyse the interpretability of the LGBM model to provide a deeper understanding of decision-making mechanisms.

Moreover, the absence of temporal indicators in the retrieval of the records rendered it unfeasible to investigate the behaviour of the machine learning models tested with the anticipated changes in viral behaviour and its treatment during the viremia process over time.

5. Conclusions

The five machine learning models exhibited comparable performance. The Light Gradient Boosting Machine (LGBM) demonstrated the most favourable outcomes, with an accuracy of 88%, a sensitivity of 80%, F1 score of 88%, and MCC of 63%. This model may be suitable for the predictive diagnosis of COVID-19 severity. The laboratory biomarkers incorporated into the LGBM learning model aided in the classification of COVID-19 severity for the samples under study.

The identified machine learning models (LGBM, XGBoost, Random Forest, Extra Trees, and Gradient Boosting) demonstrated satisfactory performance in the differential diagnosis of mild to moderate and severe COVID-19, with an accuracy between 86 and 88%, an AUC between 89 and 92%, an F1 score between 86 and 88%, and recall between 86 and 88%, as well as in identifying crucial characteristics through the feature importance methods. This study also provides evidence for the effectiveness and future potential of artificial intelligence research based on laboratory testing results.

Analysing the importance of the features revealed that age was the most significant predictor for this study. Identifying these predictors can help with clinical applications, including anticipating hospital demand by diagnosing patients who need priority attention and managing the allocation of resources in Brazil’s public health system (SUS).

Future research should focus on carrying out a more granular classification (mild, moderate, severe, critical), integrating additional clinical variables such as ICU admission, oxygen saturation or assisted ventilation to enrich the model and improve its predictive capacity.

6. Patents

The script created in Python called “COVID-19 Severity” was registered with the National Institute of Industrial Property (INPI) with the Software Registration Certificate under number BR512024002371-2, associated with the Federal University of Paraná [55].