Artificial Intelligence Predictive Model for Severe COVID-19

第一作者：康剑洪（33岁 中山大学附属第一医院胸外科博士研究生294441422@qq.com 15208247890），广州市越秀区中山二路58号，中山大学附属第一医院胸外科

通讯作者：罗红鹤（60岁 中山大学附属第一医院胸外主任 博士生导师luohhzm@163.com 13600082298），广州市越秀区中山二路58号，中山大学附属第一医院胸外科

Abstract

The present study aimed to develop an artificial intelligence predictive model for severe COVID-19 in people infected with severe acute respiratory syndrome coronavirus 2. Clinical data were collected and scikit-learn was used to conduct data preprocessing, normalization, filtering, and feature selection. TensorFlow was subsequently applied to develop a neural network model. The data were separated into two groups to test the model. A training set was used to train the prediction model, and a validation set was used to assess the performance of the models individually. The model achieved good prediction performance, with an area under the curve value of 0.966 (0.825–0.980).

Keywords: severe COVID-19, artificial intelligence, predictive model

  1. 1. Introduction

In 2019, an outbreak of very contagious pneumonia began in Wuhan, China. The disease and the virus causing the disease were named coronavirus disease 2019 (COVID-19) and severe acute respiratory syndrome coronavirus two (SARS-COV-2), respectively. Mild COVID-19 has a self-limiting course with a low mortality rate, and patients with mild symptoms are reported to recover after one week. On the other hand, severe cases are reported to experience progressive respiratory failure due to alveolar damage from the virus, which may lead to death. Proinflammatory responses play a role in the pathogenesis of severe COVID-19. In vitro experiments have shown that delayed release of cytokines and chemokines occurs in respiratory epithelial cells, dendritic cells, and macrophages during the early stages of SARS-CoV2 infection. Later, the cells secrete low levels of antiviral factors, such as interferons, and high levels of proinflammatory cytokines, such as interleukin (IL)-1β, IL-6, and tumor necrosis factor, and chemokines, such as C–C motif chemokine ligand (CCL)-2, CCL-3, and CCL-51-3. The rapid increase in cytokines and chemokines attracts inflammatory cells, such as neutrophils and monocytes, resulting in excessive infiltration of inflammatory cells into the lung tissue, leading to lung injury. Serum cytokine and chemokine levels are significantly higher in patients with severe COVID-19 compared with those with mild and moderate COVID-194. Elevated serum cytokine and chemokine levels in COVID-19 patients are associated with a high number of neutrophils and monocytes in the lung tissues and peripheral blood, suggesting that these cells may play a role in lung pathology4-6. Onset COVID-19 is usually concealed and it is difficult to predict severe COVID-19, despite knowing its causes. While studies have reported ways of predicting COVID-19, these have mostly explored the linear relationship between each feature and COVID-19 severity to identify independent risk factors. However, some of the factors related to the severity of COVID-19 are nonlinear. Classical linear prediction methods do not take nonlinear phenomena into account. However, an artificial neural network can integrate the linear and nonlinear relationship of each feature to obtain prediction results, adding to the credibility of the forecast results. Our study aimed to introduce a neural network predictive model to predict severity of COVID-19 using the results from routine examinations. A neural network is a simplified model of the way in which the human brain processes information. There are typically three parts in a neural network: an input layer, with units representing the input fields; one or more hidden layers; and an output layer, with a unit or units representing the target field(s). The units are connected with varying connection strengths (or weights). Input data are presented to the first layer, and values are propagated from each neuron to each neuron in the next layer. Eventually, a result is delivered from the output layer. Finally, we built the predictive model and the model achieved good prediction performance, with area under the curve (AUC) values of 0.966 (0.825–0.980).

  • Materials and methods

2.1 Patient and public involvement

Data were collected from the Tumor Center of Union Hospital affiliated with Tongji Medical College of Huazhong University of Science and Technology, Hubei, China. All participants gave verbal consent to take part in the study.

2.2 Data collected

Data from consecutive patients with COVID-19 were collected between January 26, 2020, and March 20, 2020. Data were obtained at admission (Table 1). Inclusion criteria were patients with a confirmed diagnosis of SARS-COV-2 infection (according to Diagnosis and Treatment Protocol for Novel Coronavirus Pneumonia Version 77). Exclusion criteria were patients with existing severe or critical COVID-19 at the time of admission. A detailed description of the criteria is described in Table 2.

Table 1. Data collected.
Data typeParameter
QuantitativeAge, RBCs, Hb, WBCs, TP, ALB, GLB, CREA, BUN, mycoplasma IgM, mycoplasma IgG, chlamydial IgM 
CategoricalPatient condition (mild cases, moderate cases, severe cases, and critical cases), sex, diabetes, diabetes with complications, acquired immune deficiency syndrome, cancer, history of lung disease, solitary patchy foci, multiple patchy foci, solitary ground-glass opacity, multiple ground-glass opacity, diffuse interstitial change, solitary interstitial change, solitary pulmonary consolidation, multiple pulmonary consolidations, solitary infiltrate, multiple infiltrates, chronic kidney diseases (>3 months) 
OrdinalHypertension classification†, cardiac functional grading (according to New York Heart Association functional classification)

ALB, albumin; BUN, blood urea nitrogen; GLB, globulin; CREA, creatinine; Hb, hemoglobin; IgG, immunoglobulin G; IgM, immunoglobulin M; RBCs, red blood cells; TP, total protein; WBCs, white blood cells

† Measured according to the 2017 edition of American College of Cardiology/American Heart Association guidelines for hypertension.

Table 2. Criteria for assessing COVID-19 severity
SeverityCriteria
MildMinimal symptoms without pulmonary involvement in chest imaging studies
ModerateFever and/or respiratory symptoms; multiple limited patchy shadows and interstitial changes in chest imaging
SevereDyspnea with a respiratory rate >30 breathsper minute; resting oxygen saturation <95% or arterial blood oxygen partialpressure/oxygen concentration ≤300 mmHg(1 mmHg = 0.133 kPa); multilobular diseaseor lesion progression >50% within 48 h;SOFA ≥2 points; pneumothorax and/or other
Critically illrespiratory failure requiring mechanicalventilation; septic shock; additional organfailure

SOFA, sequential organ failure assessment

  • Software used

2.3.1 Scikit-learn

Scikit-learn is a software machine learning library for the Python programming language. It has a powerful data preprocessing function (https://scikit-learn.org/stable/).

TensorFlow is a framework for data stream-oriented programming, which is widely used in machine learning (https://github.com/tensorflow/tensorflow).

2.4 Study design

The study consisted of the following phases:

  1. Data preprocessing
    Cases with missing and invalid values were deleted. Next, data were normalized using the median normalization and the qualitative variable was coded as dummy variables to eliminate their effect on the model (using scikit-learn package in Python software).
  2. 2. Feature selection
    Pearson correlation coefficient was used to analyze correlations of quantitative data, and Kendall correlation coefficients were used to analyze correlations of qualitative data. Statistically significant (P < 0.05) features were extracted as the input for the neural network model (using R software).
  3. Model building

Data were divided into two subsets: a training set and a validation set. The training set was used for training and tuning the parameters, and the validation set was used to evaluate the performance (using the TensorFlow package in Python software).

  • Prevention of overfitting

The stopping criterion of the training was determined using a 10-fold internal cross-validation to prevent overfitting the data. In the 10-fold cross-validation, the original sample was randomly partitioned into 10 equal-sized subsamples. Of the 10 subsamples, a single subsample was retained as the validation data for testing the model, and the remaining nine subsamples were used as training data. The cross-validation process was then repeated 10 times, using each of the 10 subsamples once as the validation data. An average of the 10 results was then taken to produce a single estimation8. The 10-fold cross-validation tested the model’s ability to predict new data that was not used in the estimation to flag problems, such as overfitting. This method was effective at preventing overfitting.

  • Evaluation indexes of the model

The performance of the model was evaluated using receiver operating characteristics (ROC) curve analysis9 and AUC.

  • Results

3.1 Data preprocessing

A total of 166 cases were included in the study, although 15 cases were excluded due to missing data. The remaining 151 COVID-19 patients comprised 59 males and 92 females with a mean age of 62.4±16.12 (range 18–96) years. There were 58 mild and moderate cases, 88 severe cases, five critical cases (age 84, 84, 69, 65, and 34 years), one case of chronic kidney disease, 21 cases of diabetes (10 with complications), none were human immunodeficiency virus-infected, 11 patients had a history of cancer, and 20 had a history of lung disease (Fig. 1).

3.2 Feature selection

The feature set of the present study consisted of 33 features. After feature selection, six eligible features were used for modeling: history of lung disease, age, hemoglobin (Hb), albumin (ALB), globulin (GLB), and blood urea nitrogen (BUN) (Table 3). The data distribution of these features is illustrated in Fig. 2. For the correlation analysis, a correlation coefficient P ≥ 0.8 was considered a very strong correlation, P = 0.60–0.79 was strong, P = 0.4–0.59 was moderate, P = 0.20–0.39 a weak correlation, and P ≤ 0.19 was negligible. In the present study, ALB showed a strong negative correlation (r = 0.790, P < 0.001) whereas GLB (r = 0.678, P < 0.001) and BUN (r = 0.715, P < 0.001) showed a strong positive correlation with severity of COVID-19. Furthermore, age showed a weak positive correlation (r = 0.294, P < 0.001), Hb showed a weak negative correlation (r = −0.237, P = 0.003), and history of lung disease shows a negligible positive correlation (r = 0.188, P = 0.021) with severity of COVID-19.

Table 3. Correlation analysis
 FeaturesCorrelation coefficientP-valueSignificance level
Kendall correlation coefficientsex−0.0090.908 
hypertension classification0.0350.649 
chronic kidney diseases0.1030.205 
cardiac functional grading0.1510.054 
diabetes0.0120.878 
AIDS–– 
cancer0.1370.089 
history of lung disease0.1880.021<0.05
solitary patchy foci−0.1200.140 
multiple patchy foci−0.0300.707 
solitary ground-glass opacity0.0030.974 
multiple ground-glass opacity0.0910.246 
solitary interstitial change0.1090.180 
diffuse interstitial change0.0940.246 
solitary pulmonary consolidation0.0120.885 
multiple pulmonary consolidations0.1190.142 
solitary infiltrate−0.1170.151 
multiple infiltrates−0.0560.489 
Pearson correlation coefficientage0.2940.000<0.05
WBC0.0810.325 
RBC−0.1590.051 
Hb−0.2370.003<0.05
TP−0.0840.307 
ALB−0.7900.000<0.05
GLB0.6780.000<0.05
CREA0.0640.438 
BUN0.7150.000<0.05
Mycoplasma immunoglobulin M−0.1250.127 
Mycoplasma immunoglobulin G−0.1260.124 
Chlamydial immunoglobulin M−0.1260.122 
Chlamydial immunoglobulin G−0.1440.079 

AIDS, acquired immunodeficiency syndrome’ ALB, albumin; BUN, blood urea nitrogen; GLB, globulin; CREA, creatinine; Hb, hemoglobin; RBCs, red blood cells; TP, total protein; WBCs, white blood cells

3.3 Model building

The total data were divided into a training set (111 cases) and a test set (40 cases). The training dataset was used for model training and internal verification, while the test dataset was used for external validation of the model.

A neural network model has four parameters: number of units, number of hidden layers, number of iterations, and learning rate. As the number of iterations, units, and hidden layers increase, the model tends to become stable, and the amplitude of the error rate decreases. However, overfitting of the model may occur if the number of iterations, units, and hidden layers are higher than the optimized value. A learning rate that is higher than the optimized value may affect the model accuracy to a certain extent, while a learning rate lower than the optimized value may lead to overfitting of neural network models and lengthen the training time. To avoid overfitting, we determined the optimum value of the above parameters using the 10-fold cross-validation technique.

After adjusting for parameters, the final predictive model made up an input layer (six units), two hidden layers (64 units), and an output layer (one unit: severe COVID-19 or non-severe COVID-19). Furthermore, the learning rate was set at 0.001 and the stochastic gradient descent iterations were set at 200. Hidden layer nodes use the ReLU (rectified linear unit) activation function (Eq. a), the output node uses Sigmoid activation function (Eq. b), mean squared error (Eq. c) was used for the loss function, and the cost function was minimized using the adaptive moment estimation method (Fig. 3).

Eq. (a)

Eq. (b)

Eq. (c)

ADAM, adaptive moment; ReLU, rectified linear unit

3.4 Evaluation of the model

The AUC of the model was 0.966 (0.825–0.980) (Fig. 4).

  • Discussion

The present study included a total of 151 cases. At the data preprocessing stage, 33 features among all cases were subjected to relatedness analyses, and six features were extracted for modeling (P < 0.05). Our results show that the AUC of the model was 0.966 (0.825–0.980), and the model had a perfect fit and accuracy. Our study aimed to predict the severity of COVID-19 using the results from routine examinations. Routine laboratory variables are extremely meaningful as they are readily accessible at the initial diagnosis, which helps early prediction.

Correlation analysis revealed that age showed a weak positive correlation (r = 0.294, P < 0.001), ALB showed a strong negative correlation (r = −0.790, P < 0.001), and GLB (r = 0.678, P < 0.001) and BUN (r = 0.715, P < 0.001) showed a strong positive correlation with severity of COVID-19, whereas Hb showed a weak negative correlation (r = −0.237, P = 0.003) and history of lung disease showed a negligible positive correlation (r = 0.188, P = 0.021).

Previous studies have reported that age is a risk factor for severe COVID-1910-12. Fei Zhou et al. analyzed 191 patients with severe COVID-19 requiring hospitalization. In this cohort, patients had a median age of 56.0 years13. ALB is an acidic, hydrophilic, and highly stable globular protein that is synthesized specifically in the liver. Our study showed that decreased ALB levels was a risk factor. Several studies have shown a negative correlation between ALB and severity of COVID-1914,15, which parallel the findings of the present study. A meta-analysis of 90 cohort studies that evaluated hypoalbuminemia as a prognostic biomarker in acutely ill patients showed that each 10-g/L decrease in serum albumin concentration was associated with a 137% increase in odds of death, 89% increase in morbidity, and 71% increase in length of hospital stay16. Thus, there is a clear association between albumin level and severity of the insult17. Based on the existing literature, we speculated three reasons for this phenomenon. First, low albumin levels can influence the pharmacokinetics. Albumin transports multiple endogenous and exogenous substances18; therefore, changes in albumin concentration during critical illness can have potentially marked effects on drug delivery and efficacy19. In a systematic review, Ulldemolins et al. reported that protein binding of antibacterials was frequently decreased in critically ill patients with hypoalbuminemia, notably with increased volume of distribution and drug clearance20. These changes could result in suboptimal treatment. Second, low albumin levels can change the acid–base balance in the human body. The balance of acidic to basic residues on albumin makes it a weak acid in physiological concentrations19,21, and a decrease in albumin concentration increases the anion gap. This passively increases the bicarbonate concentration. Third, low albumin levels could affect endothelial cell function since albumin is a crucial part of the endothelial surface layer. However, experiments in isolated organs have shown that the endothelial surface layer appears to function well until the albumin concentration falls to values as low 10 g/L22. Hence, the major insult when the vascular barrier fails to function due to severe acute illness is likely not the low albumin level, but the breakdown of the molecular structure of the endothelial glycocalyx due to hypervolemia or ischemia/reperfusion injury and other forms of systemic inflammation23. However, these are only speculations, and definitive conclusions cannot be drawn. Therefore, it remains unclear whether the effect of hypoalbuminemia on the outcome is a cause and effect relationship or whether hypoalbuminemia is a marker of serious disease.

ALB (r = −0,79) and GLB (r = 0.71) showed similar relevant intensity in our study, and an increase in GLB may be a risk factor for COVID-19 severity. GLB and ALB are commonly used as markers for assessing patient hepatic function, and low levels of ALB and high levels of GLB could indicate impaired liver function24. Thus, deterioration of COVID-19 patients maybe be correlated with damage to liver function. Furthermore, it is also important to consider the impact of BUN on the severity of COVID-19. Elevated BUN is a key indicator of kidney malfunction. This seems to indicate that renal impairment may result in deterioration of COVID-19. This speculation is consistent with our statements concerning ALB. The antioxidant and anti-inflammatory effects of ALB properties can mediate renoprotective effects25. However, it is worth noting that the present study only included one case with chronic kidney disease. This means that BUN had likely already resulted in a harmful impact on patients with COVID-19 before BUN reached the level of renal damage. Studies have indicated basic disease as a risk factor for COVID-19 exacerbation26-31. On the other hand, our results showed that the correlation between basic diseases and severity of COVID-19 was not significant. The may be due to a low number of cases with the basic disease in our cohort.

An independent analysis of various imaging manifestations of COVID-19 was performed in the present study. Our results showed a very weak (r < 0.2) correlation between imaging manifestations and severe COVID-19. This may be because predicting severe COVID-19 needs to observe dynamic changes of imaging manifestations (Fig. 5) rather than a single imaging manifestation32,33.

Artificial neural network technology, which is widely implemented in various fields of science, was used in the establishment of our model34-37. This model has good accuracy as long as there is suitable parameter adjustment. Thus, the neural network model is extremely useful for complex diseases. Furthermore, the operational process of the artificial neural network model is complicated, like the neural activity of the human brain. Therefore, there is currently no quantitative indicator that can express the relevance between predictors and forecast results in the artificial neural network model. This is a limitation of this study, as well as difficulties with artificial intelligence38.

Although our model showed good results in the test set, there are several other limitations to our study. First, although we examined our model in an internal population, we did not validate it in an external population, and its generalizability needs to be confirmed. Second, the data of our study was not comprehensive enough. Yong Gao et al. reported that IL-6 and D-dimer were closely connected with occurrence of severe COVID-19 in adult patients39. Hongyi Zhang et al. reported that patients with severe COVID-19 had a significant reduction in granulocytes compared with patients with mild COVID-1940. We did not collect information on these factors. Third, verification of our model using prospective testing with a larger sample sizes is warranted. Our sample size was relatively small and we are currently collecting recent data from a larger sample size to validate further and improve the current models.

Author contributions

Jianhong Kang designed and performed the study and critical editing. Ting Chen drafted the manuscript and provided subsequent analysis with statistics. Guipeng Du assisted and provided support in sample collection and performed the study and critical editing.

Declaration of Competing Interest

The authors declared that they have no conflicts of interest in this

work.

Acknowledgments

We appreciate the support of the Health Commission of Sichuan Province for the study.

References

1. Law HK, Cheung CY, Ng HY, et al. Chemokine up-regulation in SARS-coronavirus-infected, monocyte-derived human dendritic cells. Blood 2005; 106(7): 2366-74.

2. Cheung CY, Poon LL, Ng IH, et al. Cytokine responses in severe acute respiratory syndrome coronavirus-infected macrophages in vitro: possible relevance to pathogenesis. J Virol 2005; 79(12): 7819-26.

3. Lau SKP, Lau CCY, Chan KH, et al. Delayed induction of proinflammatory cytokines and suppression of innate antiviral response by the novel Middle East respiratory syndrome coronavirus: implications for pathogenesis and treatment. J Gen Virol 2013; 94(Pt 12): 2679-90.

4. Yang X, Yu Y, Xu J, et al. Clinical course and outcomes of critically ill patients with SARS-CoV-2 pneumonia in Wuhan, China: a single-centered, retrospective, observational study. Lancet Respir Med 2020; 8(5): 475-81.

5. Wang Y, Zhang D, Du G, et al. Remdesivir in adults with severe COVID-19: a randomised, double-blind, placebo-controlled, multicentre trial. Lancet 2020; 395(10236): 1569-78.

6. Weiss P, Murdoch DR. Clinical course and mortality risk of severe COVID-19. Lancet 2020; 395(10229): 1014-5.

7. Diagnosis and Treatment Protocol for Novel Coronavirus Pneumonia (Trial Version 7). Chin Med J (Engl) 2020; 133(9): 1087-95.

8. Victor WC, Raymond KW, Chi-Hung C. Over-Fitting and Error Detection for Online Role Mining. International Journal of Web Services Research 2012; 9(4): 1-23.

9. Feinstein AR. Principles of Medical Statistics: CRC Press; 2001.

10. Chen N, Zhou M, Dong X, et al. Epidemiological and clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in Wuhan, China: a descriptive study. Lancet 2020; 395(10223): 507-13.

11. Huang C, Wang Y, Li X, et al. Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China. Lancet 2020; 395(10223): 497-506.

12. Wang D, Hu B, Hu C, et al. Clinical Characteristics of 138 Hospitalized Patients With 2019 Novel Coronavirus-Infected Pneumonia in Wuhan, China. JAMA 2020.

13. Zhou F, Yu T, Du R, et al. Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet 2020.

14. Liu M, He P, Liu HG, et al. [Clinical characteristics of 30 medical workers infected with new coronavirus pneumonia]. Zhonghua Jie He He Hu Xi Za Zhi 2020; 43(3): 209-14.

15. Liu C, Jiang ZC, Shao CX, et al. [Preliminary study of the relationship between novel coronavirus pneumonia and liver function damage: a multicenter study]. Zhonghua Gan Zang Bing Za Zhi 2020; 28(2): 148-52.

16. Vincent JL, Dubois MJ, Navickis RJ, Wilkes MM. Hypoalbuminemia in acute illness: is there a rationale for intervention? A meta-analysis of cohort studies and controlled trials. Ann Surg 2003; 237(3): 319-34.

17. Gibbs J, Cull W, Henderson W, Daley J, Hur K, Khuri SF. Preoperative serum albumin level as a predictor of operative mortality and morbidity: results from the National VA Surgical Risk Study. Arch Surg 1999; 134(1): 36-42.

18. Nicholson JP, Wolmarans MR, Park GR. The role of albumin in critical illness. Br J Anaesth 2000; 85(4): 599-610.

19. Fanali G, di Masi A, Trezza V, Marino M, Fasano M, Ascenzi P. Human serum albumin: from bench to bedside. Mol Aspects Med 2012; 33(3): 209-90.

20. Ulldemolins M, Roberts JA, Rello J, Paterson DL, Lipman J. The effects of hypoalbuminaemia on optimizing antibacterial dosing in critically ill patients. Clin Pharmacokinet 2011; 50(2): 99-110.

21. Figge J, Rossing TH, Fencl V. The role of serum proteins in acid-base equilibria. J Lab Clin Med 1991; 117(6): 453-67.

22. Zausig YA, Chappell D, Becker BF, et al. The impact of crystalloidal and colloidal infusion preparations on coronary vascular integrity, interstitial oedema and cardiac performance in isolated hearts. Crit Care 2013; 17(5): R203.

23. Chappell D, Westphal M, Jacob M. The impact of the glycocalyx on microcirculatory oxygen distribution in critical illness. Curr Opin Anaesthesiol 2009; 22(2): 155-62.

24. Jawahar A, Gonzalez B, Balasubramanian N, Adams W, Goldberg A. Comparison of computed tomography hepatic steatosis criteria for identification of abnormal liver function and clinical risk factors, in incidentally noted fatty liver. Eur J Gastroenterol Hepatol 2020; 32(2): 216-21.

25. Iglesias J, Abernethy VE, Wang Z, Lieberthal W, Koh JS, Levine JS. Albumin is a major serum survival factor for renal tubular cells and macrophages through scavenging of ROS. Am J Physiol 1999; 277(5): F711-22.

26. Dantzer R, Heuser I, Lupien S. Covid-19: An Urgent Need For A Psychoneuroendocrine Perspective. Psychoneuroendocrinology 2020: 104703.

27. Montoya-Barthelemy AG, Lee CD, Cundiff DR, Smith EB. COVID-19 and the Correctional Environment: The American Prison as a Focal Point for Public Health. Am J Prev Med 2020.

28. Nowak-Wegrzyn A, Cianferoni A, Bird JA, Fiocchi A, Caubet JC, Medical Advisory Board of the International FA. Managing FPIES during the COVID-19 pandemic-expert recommendations. Ann Allergy Asthma Immunol 2020.

29. Fill L, Hadney L, Graven K, Persaud R, Hostoffer R. The Clinical Observation of a CVID Patient Infected with COVID-19. Ann Allergy Asthma Immunol 2020.

30. Hwee J, Chiew J, Sechachalam S. The Impact of Coronavirus Disease 2019 (COVID-19) on the Practice of Hand Surgery in Singapore. J Hand Surg Am 2020.

31. Schultz K, Wolf JM. Digital Ischemia in COVID-19 Patients: Case Report. J Hand Surg Am 2020.

32. Guan W, Liu J, Yu C. CT Findings of Coronavirus Disease (COVID-19) Severe Pneumonia. AJR Am J Roentgenol 2020: W1-W2.

33. Lee EYP, Ng MY, Khong PL. COVID-19 pneumonia: what has CT taught us? Lancet Infect Dis 2020.

34. Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting Splicing from Primary Sequence with Deep Learning. Cell 2019; 176(3): 535-48 e24.

35. Pan C, Schoppe O, Parra-Damas A, et al. Deep Learning Reveals Cancer Metastasis and Therapeutic Antibody Targeting in the Entire Body. Cell 2019; 179(7): 1661-76 e19.

36. Stokes JM, Yang K, Swanson K, et al. A Deep Learning Approach to Antibiotic Discovery. Cell 2020; 180(4): 688-702 e13.

37. Zeiser FA, da Costa CA, Zonta T, et al. Segmentation of Masses on Mammograms Using Data Augmentation and Deep Learning. J Digit Imaging 2020.

38. Nishant R, Kennedy M, Corbett J. Artificial intelligence for sustainability: Challenges, opportunities, and a research agenda. International Journal of Information Management 2020; 53: N.PAG-N.PAG.

39. Gao Y, Li T, Han M, et al. Diagnostic Utility of Clinical Laboratory Data Determinations for Patients with the Severe COVID-19. J Med Virol 2020.

40. Zheng HY, Zhang M, Yang CX, et al. Elevated exhaustion levels and reduced functional diversity of T cells in peripheral blood may predict severe progression in COVID-19 patients. Cell Mol Immunol 2020.

Figure legends:

Figure 1. Data preprocessing

Figure 2. Data distribution

Figure 3. The final model

Figure 4. ROC curve of our model

Figure 5. Imaging manifestations of severe COVID-19 (a) before treatment, (b) during treatment, and (c) after treatment.