For the identified cohort, we have considered demographics, body mass index (BMI), comorbidities [15 (link)], treatment with pressors, the quarter of COVID-19 diagnosis, patient severity at the time of diagnosis, and prescribed treatments as input features for model development.
To measure patient severity, we used an Ordinal Scale (OS) developed for use with EHR data [16 (link)]. Specifically, this was a 6-point ordinal scale assigned with odd integers from 1 to 11, devised explicitly for patients diagnosed with COVID-19 based on discrete EHR data elements. In this context, a level of 1 represents an outpatient or patient discharged from the hospital, level 3 indicates hospitalization, while being hospitalized on Oxygen or Mechanical Ventilator is an indicator of levels 5 and 7, respectively, with level 9 representing patients hospitalized on ECMO and level 11 representing death.
Fig 2 shows the lookback period used for determining the patient’s comorbidities in green with a minimum of 2 years, while highlighted in blue are the considered treatments’ duration within up to 28 days after the diagnosis, followed by the recorded patient’s outcome as of the last day of treatment.
Prescribed therapeutics on each day after the diagnosis were categorized and considered in eight distinct groups, defined as anticoagulants (Coag), steroid preparations (Ster), unproven antiviral therapies (ViralUnp), targeted antivirals (ViralTrgt), spike protein monoclonal antibodies (MonoSP), monoclonal antibody Immunomodulators (MonoI), macrolide and quinolone antibiotics (BiotMQ), and a miscellaneous treatments (Misc) category that included other treatments presumed to be administered for treatment of COVID-19. Medications in each category are shown inTable 1 .
The model considered the proportion of days on treatment combinations, any direct correlations between the treatment values and duration of treatment are removed, preventing the ML algorithm from leveraging this information directly for prediction. By using the proportion of days on treatment combinations, the modeling algorithm is forced to find the effect of different treatment distributions rather than attributing days on treatments to the outcome of interest.
To measure patient severity, we used an Ordinal Scale (OS) developed for use with EHR data [16 (link)]. Specifically, this was a 6-point ordinal scale assigned with odd integers from 1 to 11, devised explicitly for patients diagnosed with COVID-19 based on discrete EHR data elements. In this context, a level of 1 represents an outpatient or patient discharged from the hospital, level 3 indicates hospitalization, while being hospitalized on Oxygen or Mechanical Ventilator is an indicator of levels 5 and 7, respectively, with level 9 representing patients hospitalized on ECMO and level 11 representing death.
Prescribed therapeutics on each day after the diagnosis were categorized and considered in eight distinct groups, defined as anticoagulants (Coag), steroid preparations (Ster), unproven antiviral therapies (ViralUnp), targeted antivirals (ViralTrgt), spike protein monoclonal antibodies (MonoSP), monoclonal antibody Immunomodulators (MonoI), macrolide and quinolone antibiotics (BiotMQ), and a miscellaneous treatments (Misc) category that included other treatments presumed to be administered for treatment of COVID-19. Medications in each category are shown in
The model considered the proportion of days on treatment combinations, any direct correlations between the treatment values and duration of treatment are removed, preventing the ML algorithm from leveraging this information directly for prediction. By using the proportion of days on treatment combinations, the modeling algorithm is forced to find the effect of different treatment distributions rather than attributing days on treatments to the outcome of interest.
Free full text: Click here
