Introduction

Most human coronavirus infections are mild. However, several betacoronaviruses can cause serious diseases or even death1,2. The mortality rates of severe acute respiratory syndrome coronavirus (SARS-CoV) and Middle East respiratory syndrome coronavirus (MERS-CoV) were 10% and 37% respectively. SARS-CoV-2 is the pathogen for 2019 novel coronavirus disease (COVID-19)3,4, which has resulted in thousands of deaths in the world since the beginning of 2020.

The diagnosis of SARS-CoV-2 infection must be confirmed by the real-time reverse transcriptase polymerase chain-reaction (RT-PCR) or gene sequencing of specimens of patients5,6. Chest radiograph and laboratory findings are both important for accessing the severity of the disease7,8,9. Critical patients should be admitted to Intensive Care Unit (ICU) of infectious disease hospital, while mild patients could be kept and treated at isolation. It is very important to effectively prioritize resources for patients with the highest risk because of the large number of infected people10.

ICU patients and non-ICU patients differed significantly in some blood parameters, including: leukocytes, neutrophils, prothrombin time, D-dimer, total bilirubin (TB), lactate dehydrogenase, high sensitivity cardiac troponin I and procalcitonin5,7,11. Ruan et al.12 retrospectively analyzed laboratory findings of 68 nonsurvivors and 82 discharged patients, and found significant differences in lymphocytes, platelets, albumin, TB, urea nitrogen, creatinine, myoglobin, C-reactive protein and interleukin-6 between the two groups. These laboratory findings seemed useful in predicting outcome of SARS-CoV-2 infection. However, an advanced prediction model involving multiple laboratory parameters is urgently required to be applied in a clinical-decision support system to improve the predictive and prognostic accuracy.

As a branch of artificial intelligence, machine learning (ML) helps establish accurate prediction model13,14,15. However, there are few publications reporting prediction of the outcome of SARS-CoV-2 pneumonia using ML methods based on laboratory findings. Thus we retrospectively collected laboratory findings of discharged patients and non-survivors. These data were dealt with a ML method similar to radiomics16,17. We aim to establish a prediction model of outcome of SARS-CoV-2 pneumonia based on laboratory data.

Methods

All methods were carried out in accordance with relevant guidelines and regulations.

Study design and participants

This study was approved by the Ethics Commission of Hospital (TJ-2020-075). Written informed consent was waived by the Ethics Commission of hospital.

The author’s center was the designated hospital for severe and critical SARS-CoV-2 pneumonia. Patients underwent repeated RT-PCR tests to confirm SARS-CoV-2. Laboratory tests for SARS-CoV-2 pneumonia included: blood routine test, serum biochemical (including glucose, renal and liver function, creatine kinase, lactate dehydrogenase, and electrolytes), coagulation profile, cytokine test, markers of myocardial injury, infection-related makers, and other enzymes. Repeated tests were done every 3–6 days for monitoring the patient’s condition.

Oxygen support (from nasal cannula to invasive mechanical ventilation) was administered to patients according to the severity of hypoxaemia. All patients were administered with empirical antibiotic treatment, and received antiviral therapy. Most of patients improved after treatment. However, a few critical patients continued to deteriorate and eventually died.

Data collection

58 fatal cases of SARS-CoV-2 pneumonia (39 male, median age 66 years) were collected by the electronic medical record system. 68 discharged patients with SARS-CoV-2 pneumonia whose age and gender matched the non-survivors were selected (46 male, median age 66 years). The admission date of these patients was from Feb 16, 2020 to Mar 20, 2020. We reviewed all laboratory findings for each patient. Results of repeated tests were carefully compared to find the greatest deviation from normal value. In general, the greatest number in series of values was recorded. However, for platelets, red blood cell, lymphocytes, hemoglobin, calcium, total protein, albumin, estimated glomerular filtration rate (eGFR), and prothrombin activity (PTA), the minimum was recorded. Laboratory findings at the day of mortality were not used. These recorded laboratory findings were considered as lab features of a patient. A initial data set of 126 patients (non-survivor 58, discharge 68) was thus built.

There were 16 patients who did not have the entire group of laboratory features, thus their data were deleted from the dataset. The remaining data of 110 patients (51 non-survivor, 59 discharge) were analyzed by machine learning.

Statistical analysis and modeling

First, all the variables were compared between non-survivors and discharged patients using the Mann–Whitney U test for non-normally distributed features or the independent t test for normally distributed features16,17. Features with P < 0.05 were considered significant variables and selected16,17. Second, Spearman’s correlation coefficient was used to compute the relevance and redundancy of the features16,17. Third, we applied the maximum relevance minimum redundancy (mRMR) algorithm to assess the relevance and redundancy of the features16,17. The features were ranked according to their mRMR scores16,17. Fourth, the top 15 features with high-relevance and low-redundancy were selected for least absolute shrinkage and selection operator (LASSO) logistic regression model. The LASSO logistic regression model was adopted for further features selection16,17. Some candidate features coefficients were shrunk to zero and the remaining variables with non-zero coefficients were finally selected16,17. The model was used for calculating signature for each patient. Mann–Whitney U test was used for comparing signature between two groups16,17. Receiver operator characteristic (ROC), precision recall curve (PRC) analysis and Hosmer–Lemeshow test were used for further evaluation of model.

The statistical analyses were performed using R software (version 3.3.4; https://www.r-project.org)16,17. The following R packages were used: the “corrplot” package was used to calculate Spearman’s correlation coefficient; the “mRMRe” package was used to implement the mRMR algorithm; the “glmnet” was used to perform the LASSO logistic regression model, and the “pROC” package was used to construct the ROC curve16,17.

Results

Nine laboratory features were eliminated in the first step of feature selection because of non-significance. The remaining thirty-eight lab features were significantly different between two groups (P < 0.05), and then mRMR scores were obtained for them. There were seven features having non-zero coefficients after LASSO algorithm, and were selected for the model. Table 1 shows the fifteen features with the highest mRMR scores. Figure 1 shows the correlation matrix heatmap of the thirty-eight significant features. Figure 2 shows the feature selection process with LASSO algorithm. Figure 3 shows the contribution of the seven features to the model. Figure 4 shows the signatures of all patients, as well as ROC. Figure 5 shows the PRC for the model.

Table 1 The fifteen features with higher mRMR scores were selected for the step of LASSO logistic regression.
Full size table
Figure 1
figure 1

Correlation matrix heatmap of 38 significant features. Spearman’s correlation coefficient was used to compute the relevance and redundancy of the features.

Full size image
Figure 2
figure 2

The fivefold cross-validation (A) of the least absolute shrinkage and selection operator algorithm for feature selection process. A vertical line was drawn at the optimal value. Some candidate features coefficients were shrunk to zero (B) and the remaining seven variables with non-zero coefficients were finally selected.

Full size image
Figure 3
figure 3

Contribution of the features to the model. The histogram shows the contribution of the seven features with non-zero coefficients. The features are plotted on the y-axis, and their coefficients are plotted on the x-axis.

Full size image
Figure 4
figure 4

Bar charts of the signature for patients. The red bars indicate the signatures of discharged patients, while the light green bars indicate the signatures of non-survivors. The AUC was 0.997 for the signature.

Full size image
Figure 5
figure 5

The precision recall curve for the model. The area under precision recall curve was 0.996.

Full size image

Non-survivors and discharged patients differed significantly in the signature derived from the model (P < 0.0001). The AUC was 0.997 [95% CI 0.99, 1.00]. The sensitivity and specificity in predicting outcome of SARS-CoV-2 pneumonia were 98% [93%, 100%] and 91% [84%, 99%] respectively. The area under precision recall curve (AUPRC) was 0.996. Hosmer–Lemeshow test showed good calibration (P = 0.95) for the model.

The seven features included in the prediction model were as follows: PTA, urea, white blood cell (WBC), interleukin-2 receptor (IL-2r), indirect bilirubin (IB), myoglobin, and fibrinogen degradation products (FgDP). All features had coefficients of positive number except PTA. PTA and FgDP are from coagulation profile. Urea and IB are from renal and liver function respectively. WBC is from blood routine test. Myoglobin is a marker of myocardial injury. IL-2r is related to immune response. The signatures derived from the model could be positive or negative numbers.

Non-survivors and discharged patients did not differ in age or gender (median age 67 vs. 66, P = 0.75; percentage of males, 66% vs. 64%, P = 0.66). The comparisons of laboratory findings between non-survivors and discharged patients are shown in Table 2.

Table 2 Medians [inter-quartile range] of laboratory findings of patients with SARS-CoV-2 pneumonia were provided in the table.
Full size table

Blood routine test

WBC and neutrophils were significantly higher in non-survivor group versus discharge group. Lymphocyte, platelets and red blood cells were significantly lower in non-survivors. AUC for them were 0.646–0.910.

Electrolyte

Potassium, chlorine and sodium were significantly higher in non-survivor group versus discharge group. Calcium was significantly lower in non-survivors. AUC for them were 0.634–0.652.

Serum biochemical test

Glucose and globulin were significantly higher in non-survivor group versus discharge group. Albumin and total protein were significantly lower in non-survivors. AUC for them were 0.649–0.736.

Renal function

Urea and creatinine were significantly higher in non-survivor group versus discharge group. The eGFR was significantly lower in non-survivors. AUC for them were 0.672–0.907.

Liver function

Total bilirubin, direct bilirubin, IB and glutamic oxaloacetic transaminase were significantly higher in non-survivor group versus discharge group. AUC for them were 0.647–0.806.

Coagulation profile

Prothrombin time, activated partial thromboplastin time, D-dimer, international normalized ratio (INR), fibrinogen and FgDP were significantly higher in non-survivor group versus discharge group. PTA was significantly lower in non-survivors. AUC for them were 0.847–0.886.

Cytokine

IL-2r and IL-6 were significantly higher in non-survivor group versus discharge group. AUC for them were 0.689–0.909.

Infection-related markers and myocardial injury markers

Procalcitonin, high sensitive C-reactive protein, ferritin and N-terminal pro-brain natriuretic peptide (NT-proBNP) were significantly higher in non-survivor group versus discharge group. Myoglobin, MB isoenzyme of creatine kinase and high sensitive cardiac troponin I were significantly higher in non-survivors. AUC for them were 0.843–0.915.

Discussion

Non-survivors and discharged patients with SARS-CoV-2 pneumonia differed significantly in thirty-eight laboratory findings. By using machine learning method, we established a prediction model involving seven laboratory features. The model was found highly accurate in distinguishing non-survivors from discharged patients. The seven features selected by artificial intelligence also indicated that dysfunction of multiple organs or systems correlated with the prognosis of SARS-CoV-2 pneumonia.

The SARS-CoV-2 triggers a series of immune responses and induces cytokine storm, resulting in changes in immune components5,18. When immune response is dysregulated, it will result in an excessive inflammation, even cause death7,19. Excessive neutrophils may contribute to acute lung damage, and are associated with fatality20. Higher serum level of IL-2r was found in non-survivors, indicating excessive immune response. In addition, high leukocyte count in SARS-CoV-2 patients may be also due to secondary bacterial infection21,22.

Liver injury has been reported to occur during the course of the disease23,24, and is associated with the severity of diseases. Increased serum bilirubin level was observed in fatal cases. Acute kidney injury could have been related to direct effects of the virus, hypoxia, or shock25,26. Blood urea level continued to increase in some cases. Non-survivors had higher blood urea compared to survivors. Myocardial injury was seen in non-survivors, which was suggested by elevated level of myoglobin. The mechanism of multiple organ dysfunction or failure may be associated with the death of patients with SARS-CoV-2 pneumonia. Some patients with SARS-CoV-2 infection progressed rapidly with sepsis shock, which is well established as one of the most common causes of disseminated intravascular coagulation (DIC)27. The non-survivors in our cohort revealed significantly lower PTA compared to survivors. At the late stages of SARS-CoV-2 infection, level of fibrin-related markers (FgDP) markedly elevated in most cases, suggesting a secondary hyperfibrinolysis condition.

A number of laboratory features were compared between non-survivors and discharged patients with SARS-CoV-2 pneumonia. The two groups differed significantly in as many as thirty-eight lab features. However, none of the futures provided adequate accuracy in predicting the outcome of SARS-CoV-2 pneumonia. Thus, a novel prediction model involving multiple features was established in the study. With machine learning methods previously used in radiomics, a prediction model combining seven out of thirty-eight laboratory features was built for predicting the outcome of SARS-CoV-2 pneumonia.

The mRMR algorithm was used for assessing significant features to avoid redundancy between features. The mRMR score of a feature is defined as the mutual information between the status of the patients and this feature minus the average mutual information of previously selected features and this feature17,28,29. The top fifteen features with high mRMR scores were selected for the next step of modeling. The least absolute shrinkage and selection operator logistic regression model was used to processing the features selected by mRMR algorithm. LASSO is actually a regression analysis method that improves the model prediction accuracy and interpretability30. The signature calculated with the model can be positive or negative number, corresponding with poor and good prognosis respectively. Our results showed that the AUC of the signature was 10–40% higher than that of a single feature.

The modeling process is a black box; however, the choice of variables seems reasonable. PTA can more accurately reflect the coagulation function compared to prothrombin time, and can also reflect the degree of liver injury. Urea is a good index to reflect the degree of renal function damage. WBC can not only reflect immune status, but also secondary infection. IL-2r is an indicator of inflammation and immune response20. IB is related to liver function and possible hemolysis. Myoglobin reflects the degree of myocardial injury. The increase of FgDP is related to coagulation disorders including DIC. Thus the current model involves multiple important systems related to prognosis. In consideration of the high accuracy of the model, it can be concluded that liver, kidney, myocardial damage, coagulation disorder and excess immune response all contribute to the outcome of SARS-CoV-2 pneumonia.

It is suitable to start to use this model after three repeated laboratory tests (about 2 weeks after admission), because doctors may have enough data at that time. Lots of laboratory findings are generated in hospitalization. Which are most important for predicting outcome? Our study at least answered such a problem. Seven laboratory features could be used to construct a new signature with the model. The new signature seems more useful than any single feature. We encourage such a simple-to-use model widely used in clinical practice.

Most of clinical factors are not continuous variables (such as underlying disease). We used a machine learning method similar to radiomics, which mainly deals with continuous features. Our study focused on continuous laboratory variables. We had to exclude non-continuous clinical factor with the current machine learning method. By using other methods, a model that involves both continuous variables and category variables can be established. Thus clinical factors raised as significant predictive factors (such as respiratory status or radiological features) could be included in the models. However, there are more than forty laboratory findings in our study, making establishment of model difficult. We felt it necessary to simplify laboratory features. Thus we establish a sub-model based on lab findings. A new lab signature is thus created, and is proved highly valuable. In future study, the signature may be combined with clinical factors to establish a more complex model.

Our study has some limitations. First, this is a single-center retrospective study. Multi-center large-sample studies are required to validate our prediction model. Second, our model may not be directly used in other centers. However, they could easily establish a prediction model using their own data with machine learning method. Third, some patients who did not have all the lab findings were excluded. Selection bias must be present due to patients exclusion. Other studies with more strict design were thus required to reveal the bias. Fourth, statistical approach conducted in this study is not perfect. As LASSO was used for 15 variables, 150 or more patients were needed. More patients should be collected in future study.

In conclusion, it is feasible to establish a accurate prediction model of outcome of SARS-CoV-2 pneumonia based on laboratory findings. Injury of liver, kidney and myocardium, coagulation disorder and excess immune response all correlate with the outcome of SARS-CoV-2 pneumonia.