Prognostication is tough.  The challenge is often not that we have too little data, but rather that we have too much data.  There are now about two dozen risk factors for poor outcomes with COVID-19.  How are we supposed to integrate all these bits of information at the bedside?  Which pieces of information are redundant with one another?  Which are more important than others?  

It would be useful to have a single index of COVID severity (CSI, the COVID Severity Index).  The CSI would be a measurement of the underlying physiologic disarray caused by COVID (i.e., the amount of cytokine activation and disseminated intravascular coagulation).  As such, the CSI would be expected to predict subsequent disease course (unlike most scores, such as SOFA or APACHE, which measure current physiologic derangement; figure below).  The best overall way to conceptualize the CSI is similar to the MELD score (a prognostic index used to describe cirrhosis severity and prognosis).  The CSI could serve three general uses…

Use #1:  Risk stratification and disposition

COVID is a humbling disease.  Perhaps no aspect of the disease is more humbling than sorting out which patients require critical care versus less intensive care.  Disposition is difficult for two main reasons:

1. Patients commonly have a “silent hypoxemia” phenotype, which causes them to look far better than they actually are. 
2. COVID is a multi-phasic illness (including a viremic phase and a subsequent adaptive immunity phase).  Patients are often clinically OK for several days during the viremic phase, before abruptly deteriorating. 

The CSI could be useful for disease prognostication as shown below:

The presence of a high CSI upon admission could predict worsening over time.  By combining an assessment of the patient’s current physiologic severity plus the CSI, the clinician could estimate where the patient is along the disease trajectory. 

Use #2:  Candidacy for more aggressive medication therapy

The use of any medication is predicated upon potential for benefit versus the potential for risk.  The potential for risk (drug toxicity) is generally somewhat stable.  Meanwhile, the potential for benefit will depend strongly on how severe the patient’s disease course is likely to be.  For example,

• Patients with mild disease stand to receive no benefit from therapeutics (e.g. steroid, hydroxychloroquine, remdesivir, tocilizumab, etc.) – because they will do fine regardless of therapy.
• Patients with increasing disease severity may have more potential to benefit from therapies (assuming that such therapies are effective, of course). 

Hopefully we will soon have evidence-based therapies for COVID (a boy can dream).  Unfortunately, any such therapy will be in short supply.  Therefore, we will be tasked with providing our limited supplies of medication to the patients who would benefit the most.  The CSI could potentially guide which patients are good candidates for early treatment:

• Low CSI patients are likely to do fine anyway – they wouldn’t require therapy.
• High CSI patients are at risk for deterioration – these may be the best candidates.
• Extremely high CSI patients are at very high risk for deterioration – these patients could be intermediate candidates for intervention (maybe not ideal candidates, because they could be at higher risk of treatment failure, even despite therapy).

Use #3:  Allocation of mechanical ventilators

We may reach a point where the number of ventilators and ICU beds is exhausted.  This would require us to allocate ventilators to the patients most likely to benefit from them. 

Many current allocation schemes are based on the SOFA score, which is a global description of the patient’s current degree of physiological abnormality.  The SOFA score has numerous limitations when used for this purpose:

• SOFA is a blunt tool which is generic (it ignores the underlying biology of COVID).
• The SOFA score isn’t entirely objective (it’s built upon the Glasgow Coma Scale, which is not entirely objective).
• SOFA requires an arterial blood gas (which be difficult to obtain, and can fluctuate rapidly over time).

A COVID Severity Index which is specifically designed for COVID would likely out-perform a generic SOFA score.  For example:

• The SOFA score includes bilirubin – a variable which hasn’t generally been shown to predict mortality in COVID.
• The SOFA score ignores C-reactive protein and D-dimer – variables which have been proven to be strong predictors of mortality in COVID.

Let’s imagine that we are faced with two patients with COVID and respiratory failure, who are competing for a single ventilator.  The patient with the lower CSI will be a better candidate for ventilation, for two reasons:

1. The lower CSI patient will have a higher likelihood of survival overall.
2. The lower CSI patient will be more likely to get off the ventilator rapidly (thereby freeing up the ventilator for another patient).  Alternatively, the patient with a high CSI is likely to get worse before they get better – potentially tying up the ventilator for a longer period of time.

Construction of the COVID severity index

This would not be difficult.  The methodology used to derive the MELD index in cirrhosis could be used (Malinchok M et al. 2000).  The MELD score was designed based on data from 231 patients using logistic regression analysis.  A database of ~300-500 COVID patients could probably be used to derive the CSI in a similar fashion.  Variables would need to be selected which are widely available and easily measured. 

The closest thing we currently have to a COVID severity index (“CSI version 1.0”)

Zhou et al. evaluated 393 hospitalized COVID patients for predictors of “severe pneumonia” (defined as hypoxemia, severe respiratory distress, or a respiratory rate >30 breaths/minute). Among about two dozen variables, multivariable logistic analysis identified four independent predictors of severe disease:  age, neutrophil/lymphocyte ratio (NLR), C-reactive protein (CRP), and D-dimer. 

These are ideal variables for several reasons:

• All are objective.
• They are continuous variables (avoiding arbitrary breakpoints).
• They are all easily measured upon admission to the hospital.
• They seem to reflect the biological underpinnings of the disease (e.g. C-reactive protein is a metric of cytokine storm, D-dimer is a metric of disseminated intravascular coagulation).  As such, these variables are expected to predict subsequent deterioration.
• They have universally consistent reference ranges (unlike, for example, troponin).
• These variables have been shown to have prognostic validity in other studies.

They simplified this logistic regression to obtain the following equation (the 10,000 correction factor is required for conversion to commonly used units in the United States).

For detection of patients with “severe pneumonia,” the optimal cutoff value for this formula was determined to be 5.3. Using a subset of patients to validate the equation, this cutoff had a sensitivity of 90% and specificity of 71% for detecting sicker patients. The combined CSI had better performance than any of its component variables taken alone.

Casually applying this formula to a handful of patients I’ve encountered recently suggests the following ranges (but please not that this is extremely crude).

• Outpatients:  0 to ~5 ??? (Note that the definition of “severe pneumonia” in this study might correlate with any degree of clinically significant pneumonia)
• Ward patients:  ~5 to ~20ish ???
• ICU patients:  Above ~20ish ???

caveats!

This post is intended primarily as a theoretical construct, not something which should immediately guide clinical practice. For example:

• CSI version 1.0 has yet to be formally published or validated at another center. Appropriate cutoff values remain unclear.
• A more precise model that incorporates logistic coefficients (similar to MELD) should be more accurate. 
• CSI is only intended as one piece of information which must also be integrated together with the overall clinical picture.
• As with any continuous variable, values close to cutoff points yield little useful information (whereas more extreme values are more illuminating).

• A single score is needed which reflects the underlying biological severity of COVID (the COVID Severity Index, CSI).
• CSI could have roles in patient triage, allocation of limited medications, and allocation of mechanical ventilators.
• CSI should be based on objective continuous variables which are immediately available upon admission.  One study suggests the best variables are neutrophil/lymphocyte ratio, C-reactive protein, and D-dimer.
• CSI would have numerous advantages compared to SOFA score (e.g., it would be objective, disease-specific, very easily calculated, and validated among patients with COVID).  
• Construction and validation of CSI should be fairly easy to accomplish (using as a model the development of the MELD score in cirrhosis).
• If you have a dataset which could be used to construct or validate the CSI, please consider sharing it with me. 

going further

• IBCC: Prognostication in COVID
• PulmCrit: The neutrophil/lymphocyte ratio

• Author
• Recent Posts

Social Me

Josh Farkas

Josh is the creator of PulmCrit.org. He is an associate professor of Pulmonary and Critical Care Medicine at the University of Vermont.

Social Me

Latest posts by Josh Farkas (see all)

• PulmCrit Wee – A better classification of heart failure (HFxEF-RVxEF) - August 26, 2024
• PulmCrit Wee: Rational selection of infusion rate based on loading dose - June 25, 2024
• PulmCrit: PPIs are safe and effective for GI prophylaxis… the end. - June 18, 2024