Method for predicting risk of unfavorable outcomes, e.g., in COVID-19 hospitalization, from clinical characteristics and basic laboratory findings

1. A method for predicting an unfavorable outcome for a hospitalized COVID-19 patient, comprising the steps of:

a) obtaining attributes from a health record for the patient comprising at least findings obtained at admission, basic patient characteristics and laboratory data,

b) iteratively training each of a set of initial binary classifiers from a development set and child classifiers from development subsets comprising data obtained from records for a set of hospitalized COVID-19 patients including at least findings obtained at admission, basic patient characteristics, and laboratory data for each of such patient;

c) supplying the attributes obtained in step a) to a classifier using the trained set of initial binary and child classifiers implemented in a programmed computer trained to predict a risk of the unfavorable outcome, wherein:

1) the classifier is arranged as a hierarchical combination of (a) the initial binary classifier stratifying the patient into either a high risk group or a low risk group, and (b) the child classifiers further classifying the patient in a lowest risk group or a highest risk group depending how the initial binary classifier stratified the patient as either a member of the high risk or low risk group,

2) the initial binary classifier is configured as a combination of a trained classification decision tree and a logistical combination of atomic classifiers with drop-out regularization; and

3) the child classifiers are configured as a logistical combination of atomic classifiers with drop-out regularization; and

generating an output specifying a predicted risk of the unfavorable outcome for the patient based on a risk group in which the patient is classified.

2. The method of claim 1 , wherein the unfavorable outcome comprises at least one of ICU admission, acute respiratory distress syndrome, any complication, and intubation.

3. The method of claim 1 , further comprising the step of predicting one or more missing attributes in the health record for the patient, and wherein step c) comprises supplying the attributes obtained in step a) and the predicted one or more missing attributes.

4. The method of claim 1 , wherein the initial binary and child classifiers are developed from attributes which are inclusive of the attributes presented in the health record for the patient.

5. The method of claim 1 , wherein the findings obtained at admission in step a) are in a binary format and wherein the initial binary and child classifiers are developed from attributes in the form of findings obtained at admission in the development set which are converted into the binary format.

6. The method of claim 1 , wherein the attributes used for training the classification decision tree of step c) 2) comprise age, race, weight, gender, and the output classification label produced by the logistical combination of atomic classifiers with drop-out regularization.

7. A method for developing a classifier for predicting risk of an unfavorable outcome of a patient admitted to a hospital, comprising the steps of:

a) obtaining a development set comprising data from records for a multitude of patients admitted to one or more hospitals, wherein the data includes laboratory data, findings obtained at admission, basic patient characteristics, and clinical data including data indicating whether or not the unfavorable outcome occurred during hospitalization of the patients;

b) iteratively training an initial binary classifier from the development set to stratify members of the development set into high and low risk groups;

c) iteratively training child classifiers to further stratify high and low risk groups into at least highest and lowest risk groups;

d) supplying the development set to the classifier using the trained set of initial binary and child classifiers;

e) configuring a test for predicting risk of unfavorable outcome as a hierarchical combination of the initial binary classifier and the child classifiers using characteristics from a portion of data in the development set, and

wherein: 1) the initial binary classifier is configured as a combination of a trained classification decision tree and a logistical combination of atomic classifiers with drop-out regularization, and 2) the child classifiers are configured as a logistical combination of atomic classifiers with drop-out regularization; and

f) generating an output of the test specifying a predicted risk of the unfavorable outcome on a risk group in which the characteristics from the portion of data in the development set is classified.

8. The method of claim 7 , wherein the development set comprises data for a multitude of hospitalized influenza patients.

9. The method of claim 7 , wherein, for the initial binary classifier, the attributes used for training the tree comprise age, race, weight, gender, and the output classification label produced by the logistical combination of atomic classifiers with drop-out regularization.

10. The method of claim 7 , further comprising the step of predicting one or more missing attributes in the health record for the patients in the development set, and wherein step b) comprises supplying the attributes obtained in step a) and the predicted one or more missing attributes.