Models Maastricht University
Description of the tool for severity risk assessment and triage for COVID-19 patients at hospital admission: This tool indicates the risk that a COVID-19 patient will develop severe illness during hospitalization. Severe illness is defined as meeting at least one of the following criteria during hospitalization:
- respiratory failure requiring mechanical ventilation
- shock
- ICU admission
- organ failure
- death
- Model MU1 contained clinical features of patients without symptoms
Model MU1 predicts the risk of asymptomatic patients to develop severe COVID19 disease. It used clinical factors only: It has the ambition to identify the high risk patients that should follow a strict lockdown. The model contains only clinical features of patients without symptoms. - Model MU2 contained clinical features of symptomatic patients
Model MU2 predicts the risk of asymptomatic patients to develop severe COVID19 disease based on clinical factors only: It has the ambition to identify the high risk patients that have a high probability to need a respirator. - Model MU3 contained semantic CT features and age
Model MU3 predicts the risk of asymptomatic patients to develop severe COVID19 disease. It uses age and features of CT thorax: It has the ambition to identify the high risk patients that have a high probability to need a respirator. - Model MU4 employed a combination of clinical variables and lab values. This models has the highest accuracy
Model MU4 predicts the risk of asymptomatic patients to develop severe COVID19 disease. It uses age and laboratory values: It has the ambition to identify the high risk patients that have a high probability to need a respirator. This model has the highest accuracy.
To which patients does this tool apply?
The tool is developed for risk stratification and triage of patients who are diagnosed with COVID-19 and are being admitted to the hospital.
What information will you need?
Age (years)
C-reactive protein (mg/dL)
Lactate dehydrogenase (U/L)
Creatine kinase (U/L)
Urea (mg/dL)
Calcium (mg/dL)
Lymphocyte (%)
C-reactive protein (mg/dL)
Lactate dehydrogenase (U/L)
Creatine kinase (U/L)
Urea (mg/dL)
Calcium (mg/dL)
Lymphocyte (%)
How was this tool developed?
The prognostic model is based on retrospective data of 299 patients, treated at one hospital of Wuhan, China, and admitted to the hospital between December 2019 and February 2020.
In total 364 COVID-19 patients were admitted to the hospital.
Applying the exclusion criteria resulted in removal of 65 patients
14 patients had already severe illness at hospital admission
13 patients had a time interval between admission and examinations > 2 days
38 patients were excluded because of missing data.
The predicted outcome is the probability that a patient will develop severe illness during hospitalization (as defined above)
Imputation of missing data was not applied (complete case analysis)
The data was split into a training set of 239 patients and a testing set of 60 patients. Five additional datasets were available for external validation. Test 1: a Wuhan hospital from Hubei province (n=36), Test 2: a non-Wuhan hospital in Hubei province (n=90), Test 3: three hospitals from a non-Hubei province (n=32), Test 4: a hospital from Roma, Italy (n=48), and Test 5: another hospital from Genoa, Italy (n=42).
The data collection consisted of 29 clinical variables (including comorbidities), 42 laboratory measurements and 5 semantic CT characteristics (assessed by 2 radiologists).
Features showing high pairwise Spearman correlation (r > 0.8) and the highest mean correlation with all remaining features were removed, and features with high importance and robustness were selected using the Boruta method. Recursive feature elimination based on bagged tree models with a cross-validation technique (10 folds, 10 times) was performed to select the best combination of features. The variable selection process was used to select the optimal combination of clinical, laboratory, and CT semantic variables.
The selected variables were used to build a logistic regression model. The performance in terms of AUC was optimized using the testing set of 60 patients (split sample set). The maximum value of the AUC is 1.0, indicating a perfect prediction model. A value of 0.5 indicates that patients are correctly classified in 50% of the cases, i.e. as good as chance. The performance of the model on the external validation sets, ranged from 0.83 – 0.93.
Further details about the modeling approach can be found in the published article.
In total 364 COVID-19 patients were admitted to the hospital.
Applying the exclusion criteria resulted in removal of 65 patients
14 patients had already severe illness at hospital admission
13 patients had a time interval between admission and examinations > 2 days
38 patients were excluded because of missing data.
The predicted outcome is the probability that a patient will develop severe illness during hospitalization (as defined above)
Imputation of missing data was not applied (complete case analysis)
The data was split into a training set of 239 patients and a testing set of 60 patients. Five additional datasets were available for external validation. Test 1: a Wuhan hospital from Hubei province (n=36), Test 2: a non-Wuhan hospital in Hubei province (n=90), Test 3: three hospitals from a non-Hubei province (n=32), Test 4: a hospital from Roma, Italy (n=48), and Test 5: another hospital from Genoa, Italy (n=42).
The data collection consisted of 29 clinical variables (including comorbidities), 42 laboratory measurements and 5 semantic CT characteristics (assessed by 2 radiologists).
Features showing high pairwise Spearman correlation (r > 0.8) and the highest mean correlation with all remaining features were removed, and features with high importance and robustness were selected using the Boruta method. Recursive feature elimination based on bagged tree models with a cross-validation technique (10 folds, 10 times) was performed to select the best combination of features. The variable selection process was used to select the optimal combination of clinical, laboratory, and CT semantic variables.
The selected variables were used to build a logistic regression model. The performance in terms of AUC was optimized using the testing set of 60 patients (split sample set). The maximum value of the AUC is 1.0, indicating a perfect prediction model. A value of 0.5 indicates that patients are correctly classified in 50% of the cases, i.e. as good as chance. The performance of the model on the external validation sets, ranged from 0.83 – 0.93.
Further details about the modeling approach can be found in the published article.
Other Models
- Model E5 Early Warning Score for COVID-19
This is a diagnostic tool calculating the probability that a patient has COVID-19 based on clinical variables. - Model E6 predict risk of developing severe COVID-19 pneumonia
Nomogram to predict risk of developing severe COVID-19 pneumonia: a multicenter study from Wuhan and Guangdong, China - Model E7
The C-19 Index is an open source, AI-based predictive model that identifies people who are likely to have a heightened vulnerability to severe complications from COVID-19” - Model E8
This model is an open source COVID-19 prognostic tool that estimate mortality rates in patients with COVID-19. It is adapted from CDC materials. - Model E9
Diagnostic nomogram from China to predict COVID-19 pneumonia using semantic features.
- Model E5 Early Warning Score for COVID-19