Two-step risk stratification model
In order to improve prognosis of patients without clinical symptoms on
admission, we designed a two-step risk-stratification model. (1)
Initially, we constructed a random forest classifier based on AdaBoost
with 39 asymptomatic and 34 presymptomatic patients. The diagnostic
model could accurately classify asymptomatic and presymptomatic patients
in 5-fold cross-validation. (2) Next, to predict the progression of
presymptomatic COVID-19 patients, we constructed a random forest
classifier based on AdaBoost with the training set of 1,751 symptomatic
patients. The model was then validated in 9 non- severe and 25 severe
presymptomatic patients.
The r-adabag package was used to perform the ensemble learning.
Candidate features used to construct the classifier were 63 laboratory
indicators indicated in Table S2, Table S3. Patients whose laboratory
tests with more than 80% missing value were excluded. To apply the
model for early application and avoid the interference of therapy, we
only selected the laboratory test data within one day after admission.
If the patient had multiple tests, the mean value was used for analysis.
Missing values were supplied by the method of Conditional Mean Completer
(CMC).